机器学习在发作间期颅内脑电图上的应用可预测耐药性癫痫的手术结果-动脉网

Machine learning on interictal intracranial EEG predicts surgical outcome in drug resistant epilepsy

Nature 等信源发布 2025-03-05 10:17



可切换为仅中文







Abstract

摘要

Surgical success for patients with focal drug resistant epilepsy (DRE) relies on accurate localization of the epileptogenic zone (EZ). Currently, no exam delineates this zone unambiguously. Instead, the EZ is approximated by the area where seizures begin, which is identified manually through a tedious process that is prone to errors and biases.

对于局灶性耐药性癫痫（DRE）患者，手术成功依赖于对致痫区（EZ）的准确定位。目前，没有任何检查能够明确划定该区域。相反，致痫区通常通过癫痫发作起始的区域来近似确定，而这一区域是通过手动识别的，过程繁琐且容易出现错误和偏差。

More importantly, resection of this area does not always predict good surgical outcome. Here, we propose an artificially intelligent, patient-specific framework that automatically identifies the EZ requiring little to no input from clinicians, without having to wait for a seizure to occur. The framework transforms interictal intracranial electroencephalography data into spatiotemporal representations of brain activity discriminating the interictal epileptogenic network from background activity.

更重要的是，切除该区域并不总是预示着良好的手术效果。在此，我们提出了一种人工智能化的、针对患者特定情况的框架，该框架能够自动识别致痫区（EZ），几乎不需要临床医生的输入，且无需等待癫痫发作。该框架将发作间期颅内脑电图数据转化为大脑活动的时空表征，从而区分出发作间期的致痫网络与背景活动。

The epileptogenic network delineates the EZ with high precision and predicts surgical outcome. Our framework eliminates the need for manual data inspection, reduces prolonged monitoring, and enhances surgical planning for DRE patients..

癫痫发作网络以高精度划定癫痫区域（EZ）并预测手术结果。我们的框架消除了手动数据检查的需要，减少了长时间的监测，并增强了对DRE患者的手术规划。

Introduction

介绍

For patients with focal drug resistant epilepsy (DRE), neurosurgery is the best available treatment and achieves seizure freedom in up to 60% of cases

对于局灶性药物难治性癫痫 (DRE) 患者，神经外科手术是最佳的治疗方法，在多达 60% 的病例中可实现无癫痫发作。

. The neurosurgical outcome depends upon precise and accurate delineation of the epileptogenic zone (EZ), the brain area that is indispensable for the generation of seizures

神经外科手术的结果取决于对癫痫发作区（EZ）的精确和准确的划定，这个脑区是产生癫痫发作不可或缺的部分。

，

. Yet, the EZ is a theoretical concept since its elucidation before resection is uncertain

然而，EZ是一个理论概念，因为在切除前对其的阐明并不确定。

. Even after a successful resection, the EZ cannot be precisely defined since the resected area can also be larger than its actual extent

即使在成功切除后，也无法精确界定癫痫灶（EZ），因为切除区域可能比其实际范围更大。

. Therefore, there is currently no method of delineating the EZ directly; instead, this area is defined indirectly through several diagnostic tests whose results are often inconclusive or nonconcordant

因此，目前还没有直接划定EZ的方法；相反，这个区域是通过几种诊断测试间接定义的，而这些测试的结果往往不确定或不一致。

. The seizure onset zone (SOZ), the brain area where seizures initiate, currently serves as the best approximator of the EZ

发作起始区（SOZ），即癫痫发作开始的大脑区域，目前是致痫区（EZ）的最佳近似指标。

. This area is defined by analyzing ictal intervals of intracranial electroencephalography (iEEG) data

该区域是通过分析颅内脑电图 (iEEG) 数据的发作间期来定义的。

. Yet, removal of the SOZ does not always predict surgical outcome

然而，移除SOZ并不总是能预测手术结果。

，

. Moreover, seizures may take several days (or even weeks) to occur, extending the duration of invasive recording at the expense of considerable human and financial resources

此外，癫痫发作可能需要数天（甚至数周）才能发生，这延长了侵入性记录的持续时间，耗费大量人力和财力资源。

. An alternate interictal biomarker of the EZ is therefore of paramount importance

因此，EZ的另一种发作间期生物标志物至关重要。

。

Several electrophysiological abnormalities are observed during interictal intervals, where background brain activity is altered by transient interictal epileptiform discharges (IEDs), such as spikes and sharp waves, and ripples

在发作间期观察到几种电生理异常，其中背景脑活动因短暂的发作间期癫痫样放电（IEDs）如尖波、锐波和涟漪而发生改变。

，

. These abnormalities currently serve as established interictal biomarkers of epilepsy

这些异常目前可作为癫痫发作间期的已确定生物标志物。

. However, they lack specificity in delineating the EZ since they can include areas larger than the EZ, which should be preserved during surgery

然而，它们在描绘EZ时缺乏特异性，因为它们可能包括比EZ更大的区域，而这些区域在手术过程中应该被保留。

. Moreover, IEDs and ripples are identified through visual inspection of iEEG data or with the aid of automated software

此外，IEDs 和 ripples 是通过视觉检查 iEEG 数据或借助自动化软件来识别的。

. Yet, both approaches are prone to errors and biases

然而，这两种方法都容易出错并带有偏见。

. Previous studies have shown that both IEDs and ripples do not occur only as isolated events in iEEG; in contrast, they often exhibit coherent or co-activated patterns

以前的研究表明，IEDs 和涟漪在 iEEG 中并非仅作为孤立事件出现；相反，它们常常表现出相干或共激活的模式。

，

, or propagate across contacts

，或通过接触传播

，

forming brain networks (i.e., interconnected regions characterized by covariations and correlations among iEEG time-series)

形成大脑网络（即，以iEEG时间序列之间的协变和相关为特征的相互连接区域）

，

. Moreover, they occur transiently, alternating with background activity. Thus, the characterization of the brain network through the analysis of iEEG data segments, which contain both epileptiform as well as background activity, may be problematic for the accurate delineation of the EZ

此外，它们是短暂发生的，与背景活动交替出现。因此，通过包含癫痫样活动和背景活动的颅内脑电图（iEEG）数据段来表征脑网络，可能会对致痫区（EZ）的准确定位带来问题。

. Furthermore, IEDs are absent in ~10% of patients with DRE

此外，在约10%的DRE患者中不存在IED。

. Hence, the epileptologists cannot visually pinpoint all activity related to the disease in all patients. Thus, there is an ultimate need for a method that can capture both prominent and subtle epileptiform signatures and automatically distinguish the interictal epileptogenic network (IEN) from background activity.

因此，癫痫学家无法通过视觉观察来精确定位所有患者中与疾病相关的所有活动。因此，亟需一种能够捕捉显著和细微癫痫特征的方法，并自动区分发作间期癫痫网络（IEN）与背景活动。

。

Numerous artificial intelligence (AI) tools have recently been proposed to facilitate the presurgical evaluation of DRE patients

最近提出了许多人工智能（AI）工具，以促进对DRE患者的术前评估。

，

. Supervised machine learning (ML) models have been used to extract epileptiform patterns from multimodal neuroimaging data and automatically delineate the EZ from interictal iEEG data using morphological, spectral, and temporal waveform features

监督机器学习（ML）模型已被用于从多模态神经影像数据中提取癫痫样模式，并使用形态学、频谱和时间波形特征从发作间期的颅内脑电图（iEEG）数据中自动描绘癫痫区（EZ）。

，

. Yet, automated delineation of the EZ through supervised ML models is challenging since it requires a priori information regarding the exact EZ location to train the model. Unsupervised ML techniques are attractive alternatives aiming to identify dominant networks in the iEEG data

然而，通过监督机器学习模型自动描绘EZ具有挑战性，因为这需要关于EZ确切位置的先验信息来训练模型。无监督机器学习技术是很有吸引力的替代方案，旨在识别iEEG数据中的主要网络。

，

. These methods can identify relevant epileptiform patterns and inherent “structures” in the data without explicitly training the model

这些方法可以在不显式训练模型的情况下识别数据中相关的癫痫样模式和内在的“结构”。

. Thus far, these methods have only been applied to small cohorts and their potential use for outcome prediction remains unexplored.

到目前为止，这些方法仅应用于小队列，其在结果预测中的潜在用途仍有待探索。

Here, we introduce a novel AI-based framework that identifies the EZ with high precision and predicts surgical outcome in DRE patients. The framework is patient-specific and requires little to no input from clinicians; thus, it offers a significant advancement in the personalized surgical planning for DRE patients undergoing resective neurosurgery.

在此，我们介绍了一种基于人工智能的新型框架，该框架能够以高精度识别致痫区 (EZ)，并预测药物难治性癫痫 (DRE) 患者的手术效果。该框架针对患者个体化设计，几乎不需要临床医生的输入；因此，它为接受切除性神经外科手术的 DRE 患者的个性化手术规划提供了重要的进步。

The framework transforms interictal iEEG data into brain networks and their corresponding temporal maps in different frequency bands through dynamic signal decomposition (DSD) and unsupervised ML. The extracted networks characterize highly correlated and/or quasi-synchronized coherent areas that are repeatedly active across time in continuous interictal iEEG data.

该框架通过动态信号分解 (DSD) 和无监督机器学习，将发作间期的颅内脑电图 (iEEG) 数据转化为不同频段的大脑网络及其相应的时间映射。提取出的网络特征表现为在连续的发作间期 iEEG 数据中，跨时间反复活跃的高度相关和/或准同步的相干区域。

The temporal maps indicate when these networks are active in time. Our method classifies the derived networks into epileptogenic and background depending on the frequency of their activity in the temporal maps. We hypothesize that the identified IEN delineates the EZ with high precision; resection (or ablation) of this brain area yields good surgical outcome.

时间图显示了这些网络在何时活跃。我们的方法根据衍生网络在时间图中的活动频率，将其分类为致痫性网络和背景网络。我们假设所识别的IEN能够高精度地描绘EZ；切除（或消融）该脑区可获得良好的手术效果。

To test our hypothesis, we compared the spatial distribution of the IEN with the resection and clinically defined SOZ in a relatively large cohort of children and young adults with DRE. We evaluated the ability of these networks to identify the resection and SOZ and to predict outcome. More specifically, we explored the concordance of each network’s active time-windows in temporal maps with IEDs and ripples, compared each network’s power inside vs.

为了检验我们的假设，我们比较了在较大儿童和青年难治性癫痫（DRE）队列中，IEN的空间分布与切除区域及临床定义的发作起始区（SOZ）的关系。我们评估了这些网络识别切除区域和SOZ的能力，并预测其结果。更具体地说，我们探索了每个网络在时间图中的活跃时间窗与IEDs和涟漪的一致性，并比较了每个网络内部相对于其他区域的功率。

outside the resection and SOZ, and used support vector machine to predict outcome based on the IEN’s focality and proximity to resection. We further assessed the applicability of our framework for data with frequent and sparse IEDs as well as for different implantati.

在切除区域和癫痫发作区（SOZ）之外，并使用支持向量机基于IEN的病灶性和与切除区域的接近程度来预测结果。我们进一步评估了我们的框架对频繁和稀疏IED数据以及不同植入情况的适用性。

Results

结果

Patient cohort

患者队列

We retrospectively analyzed interictal iEEG data from 43 children and young adults (39.5% female) with DRE having good (27 patients, Engel I) and poor (16 patients, Engel

我们回顾性分析了43名患有耐药性癫痫（DRE）的儿童和年轻人（39.5%为女性）的发作间期颅内脑电图（iEEG）数据，其中预后良好（27名患者，Engel I级）和预后较差（16名患者，Engel分级差）。

$>$

I) outcome following epilepsy surgery [median follow-up: 4 years (2–6)]. Patient demographics are summarized in Table

I) 癫痫手术后的结果[中位随访时间：4年（2-6年）]。患者人口统计资料总结在表中。

. The patients had a median age at surgery of 13 years (range: 9.3–16.8) with median seizure onset age of four years (range: 1–7.5). The median time between diagnosis and surgery was eight years (range: 3.6–11.8). Our dataset included patients with three implantation types: (i) subdural electrodes [electrocorticography (ECoG)] (15 patients; 10 with good outcome); (ii) stereotactic EEG (sEEG) implantation (10 patients; 6 with good outcome); and (iii) both subdural and depth electrodes (18 patients; 11 with good outcome).

患者的手术中位年龄为13岁（范围：9.3–16.8），癫痫发作的中位起始年龄为4岁（范围：1–7.5）。从诊断到手术的中位时间为8年（范围：3.6–11.8）。我们的数据集包括三种植入类型的患者：(i) 脑表面电极 [皮层脑电图 (ECoG)]（15名患者；10名效果良好）；(ii) 立体定向脑电图 (sEEG) 植入（10名患者；6名效果良好）；以及 (iii) 同时使用脑表面和深部电极（18名患者；11名效果良好）。

The median number of implanted electrodes was 99 (range: 88–119). The median percentage of resected electrodes was 18% (range: 8.8–31.1) and the percentage of electrodes determined to be part of the SOZ was 9% (range: 4.6–23.7). The median resected volume determined by coregistering the preoperative and postoperative MRIs was 23.1 cm.

植入电极的中位数为99个（范围：88-119）。切除电极的中位百分比为18%（范围：8.8-31.1），而被确定为癫痫发作起始区（SOZ）部分的电极百分比为9%（范围：4.6-23.7）。通过将术前和术后MRI进行核心对齐确定的切除体积中位数为23.1立方厘米。

(range: 11.5–36.9) in good- and 20.3 cm

（范围：11.5–36.9）在良好条件下，20.3厘米

(range: 8.3–43.9) in poor-outcome patients. Eleven patients (26%) had non-lesional MRI, 26 patients (60%) had a malformation of cortical development, and six patients (14%) had acquired brain injury. Eight patients (19%) had temporal epilepsy. The median IEDs rate was 40 IEDs per minute (range: 13–58.5).

（范围：8.3–43.9）在预后不良的患者中。11名患者（26%）的MRI无病灶，26名患者（60%）存在皮质发育畸形，6名患者（14%）有获得性脑损伤。8名患者（19%）患有颞叶癫痫。IEDs的中位数为每分钟40次（范围：13–58.5）。

No significant differences in these characteristics were found between good- and poor-outcome patients..

这些特征在预后良好和预后不良的患者之间没有发现显著差异。

Table 1 Patient demographic information categorized by post-surgical outcome

表1 按术后结果分类的患者人口统计信息

Full size table

全尺寸表格

Construction of brain networks and temporal maps

脑网络和时间地图的构建

The proposed framework transforms multimodal brain imaging data (preoperative and postoperative MRIs, CT-scans, and interictal iEEG) into interpretable spatiotemporal maps comprised of active brain networks and their corresponding activity across time in the form of temporal maps using DSD and unsupervised ML.

该框架将多模态脑成像数据（术前和术后的MRI、CT扫描以及发作间期的颅内脑电图）转化为可解释的时空图，这些图由活跃的脑网络及其在时间上的对应活动以时间图的形式组成，利用了DSD和无监督机器学习方法。

By categorizing the networks as epileptogenic and background, we perform a comparative analysis to validate the ability of the IEN to delineate the EZ and predict outcome (Fig. .

通过将网络分类为致痫性和背景性，我们进行对比分析，以验证IEN描绘EZ和预测结果的能力（图。

）。

Fig. 1: Overall processing pipeline.

图1：整体处理流程。

Coregistration of MRI and computed tomography with intracranial implantations identified intracranial EEG (iEEG) electrode coordinates. Pre- and post- operative MRIs defined the resected volume. The resection and clinically defined seizure onset zone (SOZ) were used as gold standards for the epileptogenic zone (EZ) prediction.

通过颅内植入物的磁共振成像 (MRI) 和计算机断层扫描 (CT) 配准，确定了颅内脑电图 (iEEG) 电极坐标。术前和术后的 MRI 定义了切除体积。切除区域和临床定义的癫痫发作起始区 (SOZ) 被用作致痫区 (EZ) 预测的金标准。

。

Five minutes of

五分钟的

-channel iEEG were filtered in two frequency bands: spike band (

-通道 iEEG 在两个频段中进行了滤波：尖峰频段 (

某人

) [1-80 Hz] and ripple band (

) [1-80 Hz] 和涟漪频段 (

) [80–250 Hz] and dissected into

) [80–250 Hz] 并分解为

ms time-windows with 95% overlap. Dynamic mode decomposition (DMD) with

毫秒级时间窗口，重叠率为95%。动态模式分解（DMD）与

$h$

time delay embeddings extracted

提取的时间延迟嵌入

${r}_{1}$

(for

（对于

某人

) and

) 和

${r}_{2}$

(for

（对于

) oscillatory components and spectra per channel across time-windows.

) 每个通道在时间窗口上的振荡成分和频谱。

Feature matrices were constructed by averaging the DMD spectra across seven physiologically relevant frequency bands [delta (

通过在七个生理相关频段 [delta（δ）] 上对 DMD 光谱进行平均，构建了特征矩阵。

$\delta$

= 1–4 Hz), theta (

= 1–4 Hz)，theta (

$\theta$

= 4–8 Hz), alpha (

= 4–8 Hz)，alpha (

$\alpha$

= 8–12 Hz), beta (

= 8–12 Hz)，β（

$\beta$

= 12–3 Hz), gamma (

= 12–3 Hz)，伽马(

$\gamma$

= 30–80 Hz), spike band (

= 30–80 Hz)，尖峰频段 (

${sb}$

${某人}$

= 1–80 Hz), and ripple band (

= 1–80 Hz），以及涟漪频段（

${rb}$

= 80–250 Hz)], and scaled between 0 and 1. The entire network was computed by averaging the feature matrices across time-windows in each frequency band. Non-negative matrix factorization (NNMF) then extracted in each frequency band two brain networks and a temporal map.

= 80–250 Hz）]，并将其缩放至0到1之间。整个网络通过在每个频段内对时间窗口的特征矩阵进行平均计算得出。非负矩阵分解（NNMF）随后在每个频段提取出两个脑网络和一个时间映射。

Networks were categorized as epileptogenic (IEN) (red-colored) and background (blue-colored) whose focality, overlap, and distance from r

网络被分为致痫性（IEN）（红色）和背景（蓝色），其焦点性、重叠性和与 r 的距离。

section were then computed.

部分随后被计算。

Temporal maps showed active IEN (red-colored) segments and background network (blue-colored) segments across time which were concordant with interictal epileptiform discharge (IED) (pink-colored) and ripple annotations (purple-colored).

时间地图显示了活跃的IEN（红色）片段和背景网络（蓝色）片段，这些片段与发作间期癫痫样放电（IED）（粉红色）和涟漪注释（紫色）一致。

Full size image

全尺寸图像

We initially processed the preoperative MRI, postoperative MRI, and electrode implantation CT to define the exact location of iEEG electrodes and resection in relation to brain anatomy (Fig.

我们最初处理了术前MRI、术后MRI和电极植入CT，以确定iEEG电极和切除位置与大脑解剖结构的确切关系（图。

). Next, we filtered the multichannel interictal 5-minute iEEG segment in spike (

). 接下来，我们对多通道间歇性5分钟iEEG片段进行了滤波处理 (

某人

) (1–80 Hz) and ripple (

) (1–80 Hz) 和涟漪 (

) (80–250 Hz) bands (Fig.

) (80–250 Hz) 频段 (图。

). These two bands contain most interictal activity and correspond to the frequency components of IEDs and ripples

）。这两个频带包含大多数发作间活动，并对应于IEDs和涟漪的频率成分。

，

. We then dissected the data in each band into time-windows and processed them using dynamic mode decomposition (DMD) (Fig.

然后，我们将每个频段的数据分解为时间窗口，并使用动态模式分解（DMD）进行处理（图。

). The DMD has been previously used to extract coherent brain regions by decomposing neural recordings in both space and time

). DMD 以前曾被用来通过在空间和时间上分解神经记录来提取连贯的脑区。

. Since IEDs and ripples occur coherently in multiple regions

由于IEDs和涟漪在多个区域中连贯地发生

，

, DMD is a suitable approach to characterize their dynamics. We used DMD to extract power features from interictal iEEG that characterize coherent channels oscillating at specific frequency components in the form of DMD power spectra (Fig.

，DMD 是一种适合用于表征其动态特性的方法。我们使用 DMD 从发作间期 iEEG 中提取功率特征，这些特征以 DMD 功率谱的形式表征在特定频率成分上振荡的相干通道（图。

). These components refer to common frequencies present across multichannel signals in a time-window. We further processed the DMD power spectra to construct feature data matrices containing average DMD powers in physiologically relevant frequency bands (Fig.

）。这些成分指的是在多通道信号中跨时间窗口存在的共同频率。我们进一步处理了DMD功率谱，以构建包含生理相关频带内平均DMD功率的特征数据矩阵（图。

) [delta (

$\delta$

= 1–4 Hz), theta (

= 1–4 Hz)，theta (

$\theta$

= 4–8 Hz), alpha (

= 4–8 Hz)，alpha (

$\alpha$

= 8–12 Hz), beta (

= 8–12 Hz)，beta (

$\beta$

= 12–3 Hz), gamma (

= 12–3 Hz)，伽马 (

$\gamma$

= 30–80 Hz), spike band (

= 30–80 Hz)，尖峰频带 (

${sb}$

{sb}

= 1–80 Hz), and ripple band (

= 1–80 Hz)，以及涟漪波段 (

${rb}$

= 80–250 Hz)] (see “

= 80–250 Hz)]（见“

Methods

方法

”).

”`).

To characterize the dynamics of interictal iEEG data and extract recurrent active networks across time-windows, we processed the feature data matrices in each band separately. We first derived the entire network by averaging DMD spectral power of channels across time-windows in each band. We then extracted brain networks and temporal maps (Fig.

为了描述发作间期iEEG数据的动态特性并提取跨时间窗口的反复活跃网络，我们分别处理了每个频段的特征数据矩阵。首先，我们通过平均每个频段中跨时间窗口的通道DMD频谱功率，构建了整个网络。然后，我们提取了脑网络和时间映射（图。

。

). Brain networks represent coherent iEEG electrodes which are repeatedly active across time

）。大脑网络代表了在时间上反复活跃的相干iEEG电极。

. The active electrodes in the network are displayed in Fig.

网络中的活动电极如图所示。

1天

as spheres whose radii are proportional to the activation strength. Temporal maps characterize when these networks are active across time; they are displayed in Fig.

其半径与激活强度成正比的球体。时间图谱描述了这些网络在不同时间的活跃情况；它们显示在图中。

as color-coded maps indicating which network is active during a specific time-window. We define ‘active time-windows’ as time periods in the temporal map where a network is active, and ‘inactive time-windows’ as periods where a network is not active. To extract the networks and their corresponding temporal map, we used an unsupervised ML approach, namely non-negative matrix factorization (NNMF).

通过颜色编码地图指示在特定时间窗口内哪些网络处于活动状态。我们将“活跃时间窗口”定义为时间地图中网络处于活动状态的时间段，而“非活跃时间窗口”定义为网络未处于活动状态的时间段。为了提取网络及其对应的时间地图，我们使用了一种无监督的机器学习方法，即非负矩阵分解（NNMF）。

. Since IEDs and ripples occur transiently, alternating with background activity

由于IEDs和涟漪是短暂发生的，与背景活动交替出现

，

, we presumed that there are two networks which are active. We then categorized the less frequently active network as IEN and the more active as background (see “

我们假设有两个网络是活跃的。然后，我们将较不活跃的网络归类为IEN，而更活跃的则归为背景（见“

Methods

方法

”). Finally, we evaluated the ability of the networks (i.e., entire, interictal epileptogenic, and background) to delineate the EZ and predict outcome. The details of the framework are provided in “

”).最后，我们评估了网络（即整个网络、发作间期致痫网络和背景网络）划定癫痫区（EZ）和预测结果的能力。该框架的详细信息见“

Methods

方法

”.

”

We first investigated whether time-windows corresponding to the IEN were concordant with the activity of IEDs and ripples (annotated by EEG experts) across various frequency bands (see “

我们首先调查了与IEN对应的时间窗口是否与IEDs和涟漪的活动（由EEG专家标注）在各个频段上一致（见“

Methods

方法

”). Our analysis revealed that time-windows corresponding to the IEN on the temporal map had superior performance in detecting IEDs and ripples compared to the background network (

”).我们的分析显示，与背景网络相比，时间图上对应于IEN的时间窗口在检测IED和涟漪方面表现更优（

values < 0.0001,

值 < 0.0001，

Wilcoxon signed-rank

威尔科克森符号秩

test) (Supplementary Note

测试) (补充说明

). Figure

). 图

depicts the temporal concordance of the active segments of the IEN with the annotated IEDs and ripples for different frequency bands in a portion of iEEG data from patient #1. Notably, time-windows of iEEG data with frank epileptiform activity (i.e., IEDs and ripples) were temporally concordant with those corresponding to the IEN..

描绘了患者#1的iEEG数据片段中，IEN的活跃段与注释的IED和ripple在不同频带上的时间一致性。值得注意的是，具有明显癫痫样活动（即IED和ripple）的iEEG数据时间窗口与对应于IEN的时间窗口是一致的。

To verify spatial stationarity of the identified networks over time, we assessed their consistency when networks were derived from different iEEG segments using the Dice score

为了验证所识别的网络在时间上的空间平稳性，我们使用Dice分数评估了从不同iEEG片段导出网络时它们的一致性。

, a measure that quantifies the similarity and overlap between networks. A Dice score

，一种量化网络之间相似性和重叠的度量。Dice 分数

$\ge$

0.8 is considered to have almost perfect agreement

0.8 被认为具有几乎完美的一致性

. We extracted the IEN and background networks from different 1-minute-long segments and computed their Dice scores with the networks identified from the remaining 4-minute-long segments (see “

我们从不同的1分钟长的片段中提取了IEN和背景网络，并计算了它们与从剩余4分钟长的片段中识别出的网络的Dice得分（见“

Methods

方法

”). Both the IEN and background network were consistent in terms of spatial distribution in all bands when different segments were analyzed with a median Dice score

“）。在使用中位Dice分数分析不同段时，IEN和背景网络在所有频段的空间分布上都是一致的。”

$\ge$

0.8 (Supplementary Figure

0.8（补充图

）。

We finally performed a randomization test to evaluate whether the IENs were meaningful and not simply due to random fluctuations (or noise) in the data. Therefore, we randomized the feature data matrices, extracted random networks using our framework, and compared them with the original IEN and background network using the Dice score (see “.

我们最终进行了随机化测试，以评估IEN是否有意义，而不仅仅是由于数据中的随机波动（或噪声）造成的。因此，我们对特征数据矩阵进行了随机化处理，使用我们的框架提取了随机网络，并通过Dice分数将它们与原始的IEN和背景网络进行了比较（见“”。

Methods

方法

”). The resulting Dice scores were consistently low (

“）。最终的Dice分数始终较低（

$\le$

≤

0.3), indicating minimal overlap between the random networks and the IEN and background network (Supplementary Figure

0.3），表明随机网络与IEN和背景网络之间的重叠最小（补充图

). Therefore, our framework generates robust IENs that are unlikely due to random data fluctuations.

). 因此，我们的框架生成的IEN具有鲁棒性，不太可能由于随机数据波动引起。

Higher power of IEN inside resection and SOZ

在切除和SOZ内IEN的更高功率

To assess the relationship between the network power distribution and EZ, we estimated and compared the power of the entire, interictal epileptogenic, and background networks inside vs. outside the resection and SOZ in good- and poor-outcome patients, separately. In good-outcome patients, we observed increased power with moderate effect size for the entire network inside (compared to outside) resection for .

为了评估网络功率分布与癫痫区（EZ）之间的关系，我们分别估算了良好预后和较差预后患者中，整个网络、发作间期致痫网络以及背景网络在切除区内与区外、以及在癫痫发作区（SOZ）内与外的功率，并进行了比较。在良好预后患者中，我们观察到整个网络在切除区内（相较于区外）的功率增加，具有中等效应量。

$\delta$

(

= 0.0004,

= 0.0004，

= 0.42) and

= 0.42）和

$\theta$

(

= 0.0005,

= 0.34) bands (Fig.

= 0.34）波段（图。

). For these patients, we also observed higher power for the IEN inside resection for

）。对于这些患者，我们还观察到在切除范围内的IEN具有更高的功率。

$\delta$

(

= 0.0011,

= 0.0011，

= 0.34),

$\theta$

(

< 0.0001,

= 0.55),

$\alpha$

(

= 0.0012,

= 0.0012，

= 0.33),

$\beta$

(

= 0.0001,

= 0.47), and

= 0.47），并且

${sb}$

(

= 0.0066,

= 0.40) (Fig.

= 0.40)（图。

). We found no differences between inside and outside resection for the background network in good-outcome patients for any band (Fig.

). 我们在良好预后患者的任何频段的背景网络中，未发现内部与外部切除之间存在差异（图。

). Poor-outcome patients did not show differences in power for the entire, interictal epileptogenic, and background networks across all bands, except for the entire network in

). 预后不良的患者在整个网络、发作间期致痫网络和背景网络的所有频段上均未显示出功率差异，整个网络除外。

$\delta$

band (Fig.

乐队（图。

）。

Fig. 2: Median network power in good- and poor-outcome patients.

图 2：良好和较差预后患者的中位网络功率。

Median spectral power inside vs. outside resection for the entire, interictal epileptogenic (IEN), and background networks in different frequency bands in good- (green-colored) and poor- (orange-colored) outcome patients.

在良好（绿色）和较差（橙色）预后患者中，整个网络、发作间期癫痫源（IEN）网络和背景网络在不同频段内切除区域内与区域外的中位频谱功率比较。

Median dynamic mode decomposition spectral power inside vs. outside the SOZ for the entire, interictal epileptogenic (IEN), and background networks in different frequency bands in good- (green-colored) and poor- (orange-colored) outcome patients. The notch of the box plots represents the median values, lower and upper edges represent the 25.

在良好（绿色）和较差（橙色）预后患者中，整个网络、发作间期致痫网络（IEN）和背景网络在不同频段内与癫痫发作区（SOZ）内外的中位动态模态分解谱功率。箱线图的凹槽表示中位值，上下边缘分别表示第25百分位数。

and 75

和75

percentiles. Whiskers extend to the minimum and maximum values after omitting the outliers. The colored circles represent the median power of the networks of different patients and the black lines connect paired values. Significant differences are indicated by asterisks: *

百分位数。剔除异常值后，须线延伸至最小值和最大值。彩色圆圈代表不同患者的网络中位功率，黑线连接成对的值。显著性差异用星号表示：*

< 0.05, **

< 0.01, ***

< 0.001, ****

< 0.0001 (

Wilcoxon signed-rank

威尔科克森符号秩

test, significance levels were corrected for multiple comparisons using Bonferroni correction). The number displayed below the asterisks represents the effect size. Frequency bands: delta (

测试，使用Bonferroni校正对多重比较的显著性水平进行了校正）。星号下方显示的数字代表效应量。频段：delta (

$\delta$

= 1–4 Hz), theta (

= 1–4 Hz)，theta (

$\theta$

= 4–8 Hz), alpha (

= 4–8 Hz)，alpha (

$\alpha$

= 8–12 Hz), beta (

= 8–12 Hz)，β（

$\beta$

= 12–3 Hz), gamma (

= 12–3 Hz)，伽马 (

$\gamma$

= 30–80 Hz), spike band (

= 30–80 Hz)，尖峰频段 (

${sb}$

{某人}

= 1–80 Hz), and ripple band (

= 1–80 Hz），以及涟漪波段（

${rb}$

= 80–250 Hz).

= 80–250 Hz）。

Full size image

全尺寸图像

Good-outcome patients showed higher power for the entire network inside (vs. outside) the SOZ in

结果良好的患者在整个网络内部（相对于外部）的SOZ中显示出更高的功率

$\delta$

(

= 0.0002,

= 0.42),

$\theta$

(

= 0.0006,

= 0.39), and

= 0.39），并且

$\alpha$

(

= 0.0036,

= 0.32) bands (Fig.

= 0.32）波段（图。

). Similarly, the IEN showed higher power inside the SOZ for

). 同样，IEN 在 SOZ 内部显示出更高的功率。

$\delta$

(

= 0.0025,

= 0.30),

$\theta$

(

< 0.0001,

= 0.55),

$\alpha$

(

= 0.0002,

= 0.45),

$\beta$

(

< 0.0001,

= 0.53),

$\gamma$

(

= 0.0042,

= 0.0042，

= 0.37), and

= 0.37），并且

${sb}$

{sb}

(

= 0.0008,

= 0.53) (Fig.

= 0.53)（图。

). We found no differences between inside and outside the SOZ for the background network in good-outcome patients for any band (Fig.

). 在良好预后患者的任何频段中，我们发现SOZ内外的背景网络没有差异（图。

). Notably, poor-outcome patients did not show differences in power for the entire, interictal epileptogenic, and background networks across all bands (Fig.

）。值得注意的是，预后不良的患者在整个频段、发作间期致痫网络和背景网络的功率上均未显示出差异（图。

）。

Network properties

网络属性

For each network (i.e., entire, interictal epileptogenic, and background) of good-outcome patients, we estimated and compared three properties: focality measuring the proximity of the identified network electrodes to each other (

对于每个网络（即，整个网络、发作间期致痫网络和背景网络）的良好预后患者，我们估计并比较了三个属性：焦点性测量识别的网络电极彼此之间的接近程度（

net

网

), overlap with resection (

），与切除重叠（

res

结果

), and distance from resection (

），以及与切除部位的距离（

res

结果

) (see “

)（见“

Methods

方法

”). We presume that the IEN is more focal, has larger overlap, and is closer to the resection compared to the background network. The IEN showed higher

“）。我们推测，与背景网络相比，IEN 更为集中，重叠更大，并且更接近切除区域。IEN 显示出更高的”

net

网

(in

（在

$\theta$

，

$\beta$

，

$\gamma$

, and

，以及

某人

;

；

< 0.05,

≥ 0.38), higher

≥ 0.38)，更高

res

结果

(in all bands;

（在所有波段；

< 0.01,

≥ 0.40), and lower

≥ 0.40），并且更低

res

结果

(in all bands;

（在所有波段；

< 0.05,

≥ 0.33) compared to the background network (Fig.

≥ 0.33）相比于背景网络（图。

). Compared to the entire network, the IEN had higher

与整个网络相比，IEN更高

net

网

(in all bands except

（在所有波段中，除了

$\delta$

;

；

< 0.05,

≥ 0.38), higher

≥ 0.38)，更高

res

结果

(in all bands except

（在所有波段中，除了

$\delta$

and

和

$\alpha$

;

；

< 0.05,

≥ 0.26), and lower

≥ 0.26），并且更低

res

资源

(in all bands except

（在所有频段中，除了

$\delta$

and

和

$\alpha$

;

；

< 0.01,

≥ 0.32) (Fig.

≥ 0.32)（图。

). Finally, the entire network had higher

）。最后，整个网络的水平更高了

res

结果

(in

（在

$\delta$

，

$\theta$

, and

，以及

$\alpha$

;

；

< 0.01,

≥ 0.27) and lower

≥ 0.27）且更低

res

结果

(in

（在

$\theta$

and

和

$\alpha$

;

；

< 0.01,

≥ 0.06) compared to the background (Fig.

≥ 0.06）相比背景（图。

). There were no differences in

）。不存在差异

net

网

between the entire and background networks in all bands (Fig.

整个网络和背景网络在所有频段之间（图。

）。

Fig. 3: Network properties in good-outcome patients.

图3：良好预后患者的网络属性。

Focality (

焦点性 (

${F}_{{net}}$

${F}_{{净}}$

, normalized values), percentage of overlap with resection (

，标准化值），与切除重叠的百分比 (

${O}_{{res}}$

, %), and distance from resection (

，%)，以及与切除部位的距离 (

${D}_{{res}}$

, mm) of the entire, interictal epileptogenic (IEN), and background networks of good-outcome patients in different frequency bands. The notch of the box plots represents the median values, lower and upper edges represent the 25

，毫米）的整个网络、发作间期癫痫发生网络（IEN）和预后良好患者在不同频段的背景网络。箱线图的凹口代表中位数值，下边缘和上边缘分别代表第25百分位数。

and 75

和75

percentiles. Whiskers extend to the minimum and maximum values after omitting the outliers. Significant differences are indicated by asterisks: *

百分位数。剔除异常值后，须线延伸至最小值和最大值。显著性差异用星号表示：*

< 0.05, **

< 0.01, ***

< 0.001, ****

< 0.0001 (

Wilcoxon signed-rank

威尔科克森符号秩

test, significance levels were corrected for multiple comparisons using Bonferroni correction). The number displayed below or to the right of the asterisks represents the effect size. Frequency bands: delta (

测试，使用Bonferroni校正对多重比较的显著性水平进行了校正）。星号下方或右侧的数字表示效应量。频段：delta (

$\delta$

= 1–4 Hz), theta (

= 1–4 Hz)，theta (

$\theta$

= 4–8 Hz), alpha (

= 4–8 Hz)，alpha (

$\alpha$

= 8–12 Hz), beta (

= 8–12 Hz)，β（

$\beta$

= 12–3 Hz), gamma (

= 12–3 Hz)，伽马 (

$\gamma$

= 30–80 Hz), spike band (

= 30–80 Hz)，尖峰频段 (

${sb}$

{sb}

= 1–80 Hz), and ripple band (

= 1–80 Hz），以及ripple频段（

${rb}$

= 80–250 Hz).

= 80–250 Hz）。

Full size image

全尺寸图像

The IEN predicts the EZ

IEN预测EZ

We then investigated whether the entire, interictal epileptogenic, and background networks could predict the resection and SOZ in good-outcome patients by comparing the network performance metrics. We first used the power of each network’s electrodes as predictor and the resection as target. In terms of EZ prediction, the IEN outperformed the entire and background networks and was able to delineate the resection with 65% accuracy (range: 54–72%) and 85.7% precision (range: 50–100%) in .

我们随后研究了整个网络、发作间期致痫网络和背景网络是否能够通过比较网络性能指标来预测良好预后患者的切除区域和癫痫发作起始区（SOZ）。我们首先使用每个网络电极的功率作为预测因子，切除区域作为目标。在癫痫发作起始区（EZ）预测方面，发作间期致痫网络的表现优于整个网络和背景网络，并能够以65%的准确率（范围：54-72%）和85.7%的精确率（范围：50-100%）描绘出切除区域。

$\theta$

band (Fig.

乐队（图。

). Since resection often contains areas larger than the actual EZ and not all electrodes inside resection exhibit epileptogenic activity, the observed low accuracy (but high precision) of the IEN in the

）。由于切除区域通常比实际的癫痫发作区（EZ）更大，并且切除区域内的电极并非都表现出癫痫发生活动，因此观察到的IEN的低准确率（但高精确率）在

$\theta$

band is justified. We performed a similar analysis using the SOZ as a target. Although higher accuracy was achieved compared to when resection was used as a target [92% (range: 83–96%) in

频带是合理的。我们使用SOZ作为目标进行了类似的分析。尽管与使用切除作为目标相比，准确率更高 [92%（范围：83-96%）在

$\theta$

], the networks poorly performed in terms of precision [49% (range: 25–66%) in

]，网络在精度方面的表现不佳 [49%（范围：25-66%）在

$\theta$

] (Supplementary Figure

]（补充图

）。

Fig. 4: Network performance metrics for predicting resection.

图4：预测切除的网络性能指标。

Sensitivity, specificity, positive (PPV) and negative predictive values (NPV), and accuracy of the entire (purple-colored), interictal epileptogenic (IEN) (red-colored), and background (blue-colored) networks to predict resection. The notch of the box plots represents the median values, lower and upper edges represent the 25.

整个（紫色）、发作间期致痫（IEN）（红色）和背景（蓝色）网络预测切除的敏感性、特异性、阳性预测值（PPV）、阴性预测值（NPV）和准确性。箱线图的凹槽代表中位数值，下边缘和上边缘分别代表第25百分位数。

第

and 75

和75

percentiles. Whiskers extend to the minimum and maximum values after omitting the outliers. Significant differences are indicated by asterisks: *

百分位数。剔除异常值后，须线延伸至最小值和最大值。显著性差异用星号表示：*

< 0.05, **

< 0.01, ***

< 0.001, ****

< 0.0001 (

Wilcoxon signed-rank

威尔科克森符号秩

test, significance levels were corrected for multiple comparisons using Bonferroni correction). Frequency bands: delta (

测试，使用邦弗朗尼校正对多重比较的显著性水平进行了校正）。频段：德尔塔（

$\delta$

= 1–4 Hz), theta (

= 1–4 Hz)，theta (

$\theta$

= 4–8 Hz), alpha (

= 4–8 Hz)，alpha (

$\alpha$

= 8–12 Hz), beta (

= 8-12 Hz)，beta (

$\beta$

= 12–3 Hz), gamma (

= 12–3 Hz)，伽马 (

$\gamma$

= 30–80 Hz), spike band (

= 30–80 Hz)，尖峰频段 (

${sb}$

{sb}

= 1–80 Hz), and ripple band (

= 1–80 Hz)，以及ripple频段 (

${rb}$

= 80–250 Hz).

= 80–250 Hz）。

Full size image

全尺寸图像

In terms of EZ prediction, the derived IEN was able to delineate the resection with an average receiver operating characteristic (ROC) area under the curve (AUC) of 0.7440 (

在EZ预测方面，所衍生的IEN能够以平均0.7440的接收者操作特征（ROC）曲线下面积（AUC）来划定切除范围（

$\sigma$

= ±0.11) in

= ±0.11) 在

$\theta$

band (Fig.

乐队（图。

). The average AUC with resection in all other bands was <0.7 (Fig.

）。在所有其他频段中，切除后的平均AUC小于0.7（图。

). ROC curves of all good-outcome patients were above the no-discrimination line (i.e., line equivalent to chance) only in

). 所有预后良好患者的ROC曲线仅在无判别线（即相当于随机概率的线）之上

$\theta$

band (Fig.

乐队（图。

). Similarly, the IEN was able to delineate the SOZ with an average AUC of 0.7474 (

）。同样，IEN能够以平均AUC为0.7474描绘SOZ（

$\sigma$

= ±0.18) in

= ±0.18) 在

$\theta$

band (Fig.

乐队（图。

). The average AUC with SOZ in all other bands was <0.7 (Fig.

）。其他所有频段中包含SOZ的平均AUC均小于0.7（图。

）。

Fig. 5: Prediction of resection and seizure onset zone (SOZ).

图 5：切除预测和癫痫发作起始区（SOZ）。

Receiver operating characteristic (ROC) curves of interictal epileptogenic networks (IENs) as predictors of resection and SOZ for all good-outcome patients across different frequency bands. The ROC curve of each patient is shown in different color. The mean ROC curves are depicted in bold black. The diagonal line represents the line of no-discrimination or chance [area under the curve (AUC) equals 0.5].

不同频段下，发作间期致痫网络（IENs）作为切除和癫痫发作区（SOZ）预测指标的所有良好预后患者的受试者操作特征（ROC）曲线。每位患者的ROC曲线以不同颜色显示。平均ROC曲线用粗黑线表示。对角线代表无辨别力或随机情况的分界线[曲线下面积（AUC）等于0.5]。

The average AUC values are shown at the top of each panel for each band. .

每个波段的平均AUC值显示在每个面板的顶部。

AUC of the entire, interictal epileptogenic (IEN), and background networks with resection and SOZ in different bands. The notch of the box plots represents the median values, lower and upper edges represent the 25

整个网络、发作间期致痫网络（IEN）和背景网络在不同频段下，与切除区域和癫痫发作区（SOZ）的AUC。箱线图的凹口表示中位数值，下边缘和上边缘分别表示第25百分位数。

and 75

和75

percentiles. Whiskers extend to the minimum and maximum values after omitting the outliers. Significant differences are indicated by asterisks: *

百分位数。剔除异常值后，须线延伸至最小值和最大值。显著性差异用星号表示：*

< 0.05, **

< 0.01, ***

< 0.001, ****

< 0.0001 (

Wilcoxon signed-rank

威尔科克森符号秩

test, significance levels were corrected for multiple comparisons using Bonferroni correction). Frequency bands: delta (

测试，使用Bonferroni校正对多重比较的显著性水平进行了校正）。频段：delta (

$\delta$

= 1–4 Hz), theta (

= 1–4 Hz)，theta (

$\theta$

= 4–8 Hz), alpha (

= 4-8 Hz)，alpha (

$\alpha$

= 8–12 Hz), beta (

= 8-12 Hz)，beta (

$\beta$

= 12–3 Hz), gamma (

= 12–3 Hz)，伽马 (

$\gamma$

= 30–80 Hz), spike band (

= 30–80 Hz)，尖峰频带 (

${sb}$

= 1–80 Hz), and ripple band (

= 1–80 Hz)，以及涟漪波段 (

${rb}$

= 80–250 Hz).

= 80–250 Hz）。

Full size image

全尺寸图像

In predicting resection, the IEN showed superior performance with higher AUC compared to background (in

在预测切除时，IEN 表现出比背景更高的 AUC，性能更优。

$\theta$

，

$\beta$

，

$\gamma$

, and

，以及

某人

;

；

< 0.01) and entire (in

<0.01）和整个（在

$\theta$

and

和

$\beta$

;

；

< 0.05) networks (Fig.

<0.05）网络（图。

). The entire network had higher AUC compared to background (in

). 整个网络的AUC高于背景（在

$\theta$

and

和

$\beta$

;

；

< 0.01) (Fig.

<0.01）（图。

). In predicting the SOZ, the IEN showed higher AUC when compared to background (in

). 在预测SOZ时，与背景相比，IEN显示出更高的AUC（

$\theta$

，

$\alpha$

，

$\beta$

，

$\gamma$

, and

，以及

某人

;

；

< 0.05) and entire (only in

<0.05）和整个（仅在

$\theta$

;

；

< 0.01) networks (Fig.

<0.01）网络（图。

). Additionally, the entire network had higher AUC compared to background when predicting the SOZ (in

）。此外，整个网络在预测 SOZ 时相比背景具有更高的 AUC（在

$\theta$

，

$\alpha$

，

$\beta$

, and

，以及

某人

;

；

< 0.05) (Fig.

<0.05)（图。

）。

Robust IENs across variable IED rates and implantation types

在不同IED速率和植入类型下均具有鲁棒性的IEN

We investigated the robustness of the IEN across segments with frequent and sparse IEDs for the good-outcome patients. Only 16 out of the 27 good-outcome patients had segments of 1-minute duration with both frequent (57.2 IEDs per minute) and sparse IEDs (8.7 IEDs per minute) (Supplementary Table

我们研究了在频繁和稀疏IEDs的片段中，IEN在良好预后患者中的稳健性。在27名良好预后患者中，只有16名患者有持续1分钟的片段，同时包含频繁（每分钟57.2个IEDs）和稀疏（每分钟8.7个IEDs）的IEDs（补充表）。

). For these patients, we estimated the IENs using our framework and calculated the AUC of the IENs with the resection and SOZ. We then compared the AUC values derived from the segments with frequent IEDs with those derived from the segments with sparse IEDs (see “

). 对于这些患者，我们使用我们的框架估计了IEN，并计算了IEN与切除区域和SOZ的AUC。然后，我们将从频繁IED的片段得出的AUC值与从稀疏IED的片段得出的AUC值进行了比较（见“

Methods

方法

”). We found no differences between the AUC values from segments with frequent vs. sparse IEDs (

”).我们发现频繁与稀疏IEDs的片段之间的AUC值没有差异(

> 0.05) (Fig.

> 0.05) (图。

）。

Fig. 6: Consistency of interictal epileptogenic networks (IENs) across variable interictal epileptic discharge (IED) rates and implantation types.

图6：不同发作间期癫痫样放电（IED）频率和植入类型下，发作间期致痫网络（IENs）的一致性。

Area under the curve (AUC) of IENs with resection and SOZ in segments with frequent and sparse interictal IEDs.

包含切除和SOZ的IEN在频繁和稀疏发作间期IED段下的曲线面积（AUC）。

AUC of IENs with resection and SOZ in patients with different implantation types. The notch of the box plots represents median values; lower and upper edges represent the 25

不同植入类型患者IENs切除和SOZ的AUC。箱线图的凹口代表中位数值；下边缘和上边缘代表第25百分位数。

and 75

和75

percentiles. Whiskers extend to the minimum and maximum values after omitting the outliers. Significant differences are indicated by asterisks: *

百分位数。剔除异常值后，须线延伸至最小值和最大值。显著性差异用星号表示：*

< 0.05, **

< 0.01, ***

< 0.001, ****

< 0.0001 (

Wilcoxon signed-rank

威尔科克森符号秩

test for rate consistency analysis and

用于速率一致性分析的测试和

Wilcoxon rank-sum

威尔科克森秩和检验

test for implantation consistency analysis, significance levels were corrected for multiple comparisons using Bonferroni correction). Frequency bands: delta (

用于植入一致性分析的测试，使用邦弗朗尼校正对多重比较的显著性水平进行了校正）。频段：δ（

$\delta$

= 1–4 Hz), theta (

= 1–4 Hz)，theta (

$\theta$

= 4–8 Hz), alpha (

= 4–8 Hz)，alpha (

$\alpha$

= 8–12 Hz), beta (

= 8-12 Hz)，beta (

$\beta$

= 12–3 Hz), gamma (

= 12–3 Hz)，伽马 (

$\gamma$

= 30–80 Hz), spike band (

= 30–80 Hz)，尖峰频段 (

${sb}$

{sb}

= 1–80 Hz), and ripple band (

= 1–80 Hz），以及ripple频段（

${rb}$

= 80–250 Hz). ECoG = electrocorticography; sEEG = stereotactic electroencephalography.

= 80-250 Hz）。ECoG = 皮层脑电图；sEEG = 立体定向脑电图。

Full size image

全尺寸图像

We then examined whether our framework provides consistent findings among the three implantation types in our dataset: (i) subdural electrodes; (ii) sEEG implantations, and (iii) both subdural and depth electrodes. For good-outcome patients, we initially estimated the IENs and then calculated their AUC with the resection and SOZ.

我们随后检查了我们的框架是否在数据集中的三种植入类型中提供一致的结果：(i) 硬膜下电极；(ii) sEEG 植入，以及 (iii) 硬膜下和深部电极。对于预后良好的患者，我们首先估算了 IEN，然后计算了它们与切除区域和 SOZ 的 AUC。

We then compared the AUC of IENs with resection and SOZ among the three implantation types (see “.

我们随后比较了三种植入类型中IENs与切除和SOZ的AUC（见“。

Methods

方法

”). We found no differences between the AUCs of IENs with resection among the different implantation types in any frequency band (

“）。在任何频率范围内，我们发现不同植入类型之间，切除IENs的AUC没有差异（

> 0.05) except the

> 0.05) 除了

$\alpha$

band (

乐队 (

= 0.0023) (Fig.

= 0.0023)（图。

). We also found no differences between the AUCs for predicting the SOZ (p > 0.05) (Fig.

). 我们还发现预测 SOZ 的 AUC 之间没有差异 (p > 0.05) (图。

）。

Focality and proximity of IEN to resection

病变焦点与切除部位的邻近性

We examined the focality and proximity of the IEN to resection. We presumed that the IEN would have higher

我们检查了IEN的焦点性和与切除位置的接近程度。我们推测IEN会有更高的

net

网

, lower

，更低

res

结果

, and lower

，以及更低的

res

资源

to resection in good- compared to poor-outcome patients. At the population level, the IEN had higher

与预后较差的患者相比，预后良好的患者更适合进行切除手术。在群体水平上，IEN更高。

net

网

in good-outcome patients (compared to poor) in

良好预后患者（相较于预后较差的患者）在

$\theta$

(

= 0.0002,

= 0.69),

$\alpha$

(

= 0.0051,

= 0.52), and

= 0.52），并且

$\beta$

(

= 0.0040,

= 0.53) bands (Fig.

= 0.53）波段（图。

). The IEN also had higher

）。IEN还具有更高的

res

结果

in good-outcome patients in

在结果良好的患者中

$\theta$

(

= 0.0003,

= 0.0003，

= 0.54) and

= 0.54）和

$\beta$

(

= 0.0013,

= 0.34) bands (Fig.

= 0.34) 波段 (图。

). Finally, the

）。最后，

res

结果

of the IEN was lower in good-outcome in

IEN在良好结果中的值较低

$\theta$

(

= 0.0004,

= 0.0004，

= 0.54),

$\beta$

(

= 0.0022,

= 0.42), and

= 0.42），并且

$\gamma$

(

= 0.0037,

= 0.0037，

= 0.54) bands compared to poor-outcome patients (Fig.

= 0.54）条带，相较于预后较差的患者（图。

）。

Fig. 7: Interictal Epileptogenic network (IEN) properties.

图7：发作间期致痫网络（IEN）特性。

Network focality (

网络焦点性 (

${F}_{{net}}$

${F}_{{净}}$

, normalized values), percentage of overlap with resection (

，标准化值），与切除重叠的百分比 (

${O}_{{res}}$

, %), and distance from resection (

， %），以及与切除部位的距离（

${D}_{{res}}$

, mm) of the IENs in different frequency bands in good- (green-colored) and poor- (orange-colored) outcome patients. The notch of the box plots represents the median values, lower and upper edges represent the 25

不同频率波段的IEN（绿色表示良好结果患者，橙色表示较差结果患者）的盒图中，缺口代表中位数值，下边缘和上边缘分别代表第25百分位数。

and 75

和75

percentiles. Whiskers extend to the minimum and maximum values after omitting the outliers. Significant differences are indicated by asterisks: *

百分位数。剔除异常值后，须线延伸至最小值和最大值。显著性差异用星号表示：*

< 0.05, **

< 0.01, ***

< 0.001, ****

< 0.0001 (

Wilcoxon rank-sum

Wilcoxon 秩和检验

test, significance levels are corrected for multiple comparisons using Bonferroni correction). The number displayed below the asterisks represents the effect size.

测试，使用Bonferroni校正对多重比较的显著性水平进行校正）。星号下方显示的数字代表效应量。

Properties of the IEN across patients and outcomes in different frequency bands. For visualization purposes, the reciprocal of the distance from resection (

不同频率波段中跨患者的IEN属性及结果。为了便于可视化，切除距离的倒数（

$\dagger$

) is shown. All values were normalized between 0 and 1. Circle radii are proportional to the normalized values with larger circles indicating higher focality, higher overlap, and proximity to resection. The bar graphs on the sides of the panels for each property represent the number of cases that were above the corresponding thresholds .

）已显示。所有值均在0到1之间进行了归一化处理。圆的半径与归一化值成正比，较大的圆表示更高的聚焦性、更高的重叠度以及更接近切除区域。每个属性面板侧面的条形图代表超过相应阈值的案例数量。

${th}$

(where

（其中

${th}$

represents

表示

$t{h}_{F}$

= 0.48 for

= 0.48 对于

${F}_{{net}}$

${F}_{{净}}$

，

$t{h}_{O}$

= 49% for

= 49% 支持

${O}_{{res}}$

，

$t{h}_{D}$

= 18.25 mm for

= 18.25 毫米对于

${D}_{{res}}$

,) for good- (

，）永远（

left

左

) and poor-outcome (

）和不良结果（

right

正确

) patients per frequency band. Frequency bands: delta (

每位患者每频段。频段：δ（

$\delta$

= 1–4 Hz), theta (

= 1–4 Hz)，theta (

$\theta$

= 4–8 Hz), alpha (

= 4–8 Hz)，alpha (

$\alpha$

= 8–12 Hz), beta (

= 8–12 Hz)，β（

$\beta$

= 12–3 Hz), gamma (

= 12–3 Hz)，伽马 (

$\gamma$

= 30–80 Hz), spike band (

= 30–80 Hz)，尖峰频段 (

${sb}$

{sb}

= 1–80 Hz), and ripple band (

= 1–80 Hz），以及涟漪波段（

${rb}$

= 80–250 Hz).

= 80–250 Hz）。

Full size image

全尺寸图像

At the patient level, we report the IEN properties for all patients aiming to evaluate differences between good- and poor-outcome patients. The

在患者层面，我们报告了所有患者的整体指数属性，旨在评估结局良好和结局不良患者之间的差异。

net

网

and

和

res

结果

were higher in most patients with good (compared to poor) outcome (Fig.

在大多数预后良好（相对于预后较差）的患者中更高（见图。

). Similarly,

）。同样，

res

结果

was lower in most patients with good outcome. We report

在大多数预后良好的患者中较低。我们报告

net

网

，

res

结果

, and

，以及

res

资源

for each patient (Fig.

对于每位患者（图。

) after estimating optimal thresholds (see “

）在估计最佳阈值后（见“

Methods

方法

”) of each property for predicting outcome (

”) 的每个属性用于预测结果 (

$t{h}_{F}$

= 0.48,

= 0.48，

$t{h}_{O}$

= 49%, and

= 49%，并且

$t{h}_{D}$

= 18.25 mm, respectively). The thresholds for each band are reported in Supplementary Table

= 18.25 毫米，分别）。每个波段的阈值报告在补充表中。

. We observed that 20 out of 27 (74.1%) good-outcome patients had

我们观察到27名（74.1%）预后良好患者中有20名

net

网

≥

$t{h}_{F}$

在

$\theta$

band (Fig.

乐队（图。

); contrarily, only 4 out of 16 (25%) poor-outcome patients had

）；相反，16名（25%）预后不良患者中只有4名具有

net

网

≥

$t{h}_{F}$

. We also found that 85.1% of good-outcome patients (23 out of 27) had

。我们还发现，85.1%的预后良好患者（27人中的23人）具有

res

结果

≥

$t{h}_{O}$

with resection in

切除后

$\theta$

band (Fig.

乐队（图。

); contrarily, only 31.2% (5 out of 16) of poor-outcome patients had

）；相反，只有31.2％（16人中的5人）的预后不良患者有

res

结果

≥

$t{h}_{O}$

with resection. Finally, we found that 92.6% of good-outcome patients (25 out of 27) had

进行切除。最后，我们发现 92.6% 的预后良好患者（27 人中的 25 人）有

res

结果

≤

$t{h}_{D}$

在

$\theta$

band; contrarily, only 37.5% of poor-outcome patients (6 out of 16) had

乐队；相反，只有37.5%的预后不良患者（16人中的6人）有

res

结果

≤

$t{h}_{D}$

(Fig.

（图。

）。

IEN predicts surgical outcome

IEN预测手术结果

To evaluate the IEN as interictal biomarker of the EZ, we examined its ability to predict outcome using network properties as predictors. Particularly, we trained a linear support vector machine classifier (SVM)

为了评估IEN作为EZ的发作间期生物标志物，我们检验了其利用网络属性作为预测因子来预测结果的能力。特别是，我们训练了一个线性支持向量机分类器（SVM）。

to estimate the outcome using the three IEN properties as features and outcome as target in each of the seven bands. We also trained another linear SVM classifier (denoted by SVM-all) using 21 properties (seven bands × three properties per band) simultaneously as features that incorporated information from all bands combined.

使用三个IEN属性作为特征，以结果为目标，在七个频段中的每一个上估计结果。我们还训练了另一个线性SVM分类器（记为SVM-all），同时使用21个属性（七个频段×每个频段三个属性）作为特征，综合了所有频段的信息。

Each SVM model was validated using five-fold cross-validation. SVM outputs the probability of being classified as good and poor (see “.

每个SVM模型都使用了五折交叉验证进行验证。SVM输出被分类为良好和较差的概率（见“。

Methods

方法

”). Similarly, for each band, we trained SVMs using entire network properties (as well as SVM-all with 21 properties) as outcome predictors.

”).同样，对于每个波段，我们使用整个网络属性（以及具有21个属性的SVM-all）作为结果预测因子来训练SVM。

Predictive models using SVM revealed that the entire network predicted outcome in

使用SVM的预测模型显示，整个网络预测结果在

$\theta$

and

和

某人

(

< 0.05) (Fig.

<0.05)（图。

). Moreover, the IEN demonstrated predictive capabilities in multiple bands, including

). 此外，IEN 在多个频段中展示了预测能力，包括

$\theta$

，

$\beta$

，

$\gamma$

, and

，以及

某人

(

< 0.05) (Fig.

<0.05)（图。

). Overall, the IEN outperformed the entire network in predicting outcome. In particular, SVM trained using the IEN in

总体而言，IEN在预测结果方面优于整个网络。特别是使用IEN训练的SVM在

$\theta$

band performed best with 93% sensitivity, 63% specificity, 81% precision [positive predictive value (PPV)], 83% negative predictive value (NPV), 81% accuracy, and 0.85 AUC (Fig.

乐队表现最佳，敏感性为93%，特异性为63%，精确度为81% [阳性预测值 (PPV)]，阴性预测值 (NPV) 为83%，准确率为81%，AUC为0.85（图。

). On the other hand, SVM trained using the entire network in

). 另一方面，使用整个网络训练的SVM在

$\theta$

band had 85% sensitivity, 44% specificity, 72% PPV, 64% NPV, 70% accuracy, and 0.71 AUC (Fig.

带状物具有85%的敏感性、44%的特异性、72%的阳性预测值、64%的阴性预测值、70%的准确性和0.71的AUC（图。

）。

Fig. 8: Outcome prediction using machine learning.

图 8：使用机器学习进行结果预测。

Confusion matrices of the support vector machine (SVM) classifier using three properties (focality, overlap, and distance from resection) of the entire network in each band as predictors of outcome. “All” (checkered red-colored) is the result of training SVM-all using all properties of all bands as predictors.

使用整个网络在每个频段的三个属性（焦点性、重叠性和与切除的距离）作为结果预测因子的支持向量机（SVM）分类器的混淆矩阵。“全部”（方格红颜色）是使用所有频段的所有属性训练 SVM-all 的结果。

Numbers in parentheses represent .

括号中的数字代表。

values (

值（

Fisher’s exact

费舍尔精确检验

test). Significant differences are indicated in bold.

测试）。显著差异以粗体显示。

Confusion matrices of SVM using three network properties of the interictal epileptogenic network (IEN) in each band as predictors. “All” (red-colored) is the result of training SVM-all classifier. Numbers in parentheses represent

使用每个频段的发作间期致痫网络 (IEN) 的三个网络属性作为预测因子的 SVM 混淆矩阵。“All”（红色）是 SVM-all 分类器的训练结果。括号中的数字表示

values (

值（

Fisher’s exact

费舍尔精确

test). Significant differences are indicated in bold.

测试）。显著差异以粗体显示。

Bar graphs depict the performance metrics in percentage [sensitivity, specificity, positive and negative predictive values (PPV and NPV, respectively), accuracy, and five-fold cross-validation area under the curve (AUC)] of the confusion matrices of (

条形图以百分比形式显示了混淆矩阵的性能指标[敏感性、特异性、阳性与阴性预测值（分别为PPV和NPV）、准确率，以及五折交叉验证曲线下面积（AUC）]。

) and (

) 和 (

). In each frequency band, the first bar represents the entire network (checkered pattern), while the second bar represents the IEN (fully-colored).

）。在每个频段中，第一个条形代表整个网络（方格图案），而第二个条形代表IEN（全彩色）。

Feature importance analysis using ANOVA where

使用ANOVA进行特征重要性分析，其中

$F$

，

$O$

, and

，以及

$D$

represent focality, overlap, and distance from resection, respectively. Significant features are indicated by asterisks: *

分别表示病灶、重叠和与切除的距离。显著性特征用星号表示：*

< 0.05, **

< 0.01, ***

< 0.001, ****

< 0.0001 (corrected for multi

< 0.0001（已校正多重比较）

le comparisons using Bonferroni correction). Frequency bands: delta (

使用邦弗朗尼校正的比较）。频段：德尔塔（

$\delta$

= 1–4 Hz), theta (

= 1–4 Hz)，theta (

$\theta$

= 4–8 Hz), alpha (

= 4–8 Hz)，alpha (

$\alpha$

= 8–12 Hz), beta (

= 8–12 Hz)，β（

$\beta$

= 12–3 Hz), gamma (

= 12–3 Hz)，伽马 (

$\gamma$

= 30–80 Hz), spike band (

= 30–80 Hz)，尖峰频带 (

${sb}$

{sb}

= 1–80 Hz), and ripple band (

= 1–80 Hz)，以及涟漪波段 (

${rb}$

= 80–250 Hz).

= 80–250 Hz）。

Full size image

全尺寸图像

SVM-all trained using all 21 IEN properties predicted outcome (

使用全部21个IEN属性训练的SVM-all预测结果（

= 0.0007) (Fig.

= 0.0007) （图。

) with 89% sensitivity, 63% specificity, 81% PPV, 77% NPV, 79% accuracy, and 0.86 AUC (Fig.

）具有89%的敏感性、63%的特异性、81%的阳性预测值、77%的阴性预测值、79%的准确性和0.86的AUC（图。

). SVM-all trained using the entire network properties predicted outcome (

). SVM-all 使用整个网络属性训练预测结果 (

= 0.0104) (Fig.

= 0.0104)（图。

) with 78% sensitivity, 63% specificity, 78% PPV, 63% NPV, 72% accuracy, and 0.75 AUC (Fig.

）具有78%的敏感性、63%的特异性、78%的阳性预测值、63%的阴性预测值、72%的准确性和0.75的AUC（图。

）。

Finally, we performed feature importance analysis using analysis of variance (ANOVA)

最后，我们使用方差分析 (ANOVA) 进行了特征重要性分析。

四十三

to rank the IEN properties in the seven bands in decreasing order of importance and identified the IEN properties and bands that best discriminated between good and poor outcome. We found that

根据重要性对七个频段的IEN属性进行排序，并确定了最能区分良好和较差结果的IEN属性和频段。我们发现

net

网

，

res

结果

, and

，以及

res

结果

在

$\theta$

band had the highest importance for predicting outcome followed by

乐队对预测结果的重要性最高，其次是

net

网

and

和

res

结果

在

$\beta$

, and

，以及

res

结果

在

$\gamma$

band (Fig.

乐队（图。

8天

）。

Representative cases

代表性案例

To highlight the ability of the IEN to predict the EZ and outcome at the patient level, we report findings from two representative cases: a patient with good and one with poor outcome. For these patients, we depict a snapshot of annotated iEEG data, the corresponding temporal maps (Fig.

为了突出IEN在患者层面上预测EZ和结果的能力，我们报告了两个代表性病例的发现：一个预后良好，一个预后较差。对于这些患者，我们展示了一个注释的iEEG数据快照以及相应的时间图（图。

), and the IEN and the background network in

），以及IEN和背景网络在

$\theta$

band (Fig.

乐队（图。

九b

). We also report network properties, their predictive values for resection and SOZ, and the concordance of temporal maps with IEDs (Fig.

）。我们还报告了网络属性、它们对切除和SOZ的预测值，以及时间图与IEDs的一致性（图。

）。

Fig. 9: Case studies of a good- and a poor-outcome patient.

图 9：良好结局和不良结局患者的案例研究。

Five-second 10-channel sample intracranial EEG data with annotated spikes (pink-colored) and corresponding temporal maps showing activity of the background network (blue-colored) and interictal epileptogenic network (IEN) (red-colored) for a good- (patient #1, 11-year-old male) and a poor- (patient #31,16-year-old male) outcome patient.

五秒钟的10通道样本颅内脑电图数据，标注了尖峰（粉红色）以及显示背景网络活动（蓝色）和发作间期致痫网络（IEN）（红色）的相应时间图，分别对应预后良好（患者#1，11岁男性）和预后较差（患者#31，16岁男性）的患者。

。

The background (blue-colored) and IEN (red-colored) networks in

背景（蓝色）和IEN（红色）网络在

$\theta$

band projected on MRI with the resection (green-colored) and zoomed around the resection for both good- and poor-outcome cases. The spheres were centered at the network’s electrodes and their radii were proportional to the dynamic mode decomposition (DMD) spectral power (only electrodes greater than the threshold are shown).

在MRI上以绿色标示出切除区域，并围绕切除区域进行放大，展示良好和较差预后病例。球体以网络电极为中心，其半径与动态模式分解（DMD）频谱功率成正比（仅显示超过阈值的电极）。

。

For both patients, IEN and background network properties [i.e., focality (

对于两位患者，IEN 和背景网络属性 [即，病灶性 (

${F}_{{net}}$

${F}_{{净}}$

), overlap (

），重叠（

${O}_{{res}}$

, %) and distance from resection (

，%) 以及与切除部位的距离 (

${D}_{{res}}$

, mm)] are displayed for the

，毫米）] 显示了

$\theta$

band. Performance metrics [i.e., area under the curve (AUC) with resection (AUC-RES) and seizure onset zone (AUC-SOZ), and concordance with interictal epileptiform discharges (IEDs) (AUC-IED)] for both the background network (blue-colored) and IEN (red-colored) are displayed for the

乐队。性能指标 [即，曲线下面积 (AUC) 与切除 (AUC-RES) 和癫痫发作起始区 (AUC-SOZ)，以及与发作间期癫痫样放电 (IED) 的一致性 (AUC-IED)] 显示了背景网络（蓝色）和 IEN（红色）的

$\theta$

band. Frequency bands: delta (

频带。频率波段：δ（

$\delta$

= 1–4 Hz), theta (

= 1–4 Hz)，theta (

$\theta$

= 4–8 Hz), alpha (

= 4–8 Hz)，alpha (

$\alpha$

= 8–12 Hz), beta (

= 8-12 Hz)，β（

$\beta$

= 12–3 Hz), gamma (

= 12–3 Hz)，伽马 (

$\gamma$

= 30–80 Hz), spike band (

= 30–80 Hz)，尖峰频段 (

${sb}$

= 1–80 Hz), and ripple band (

= 1–80 Hz），以及涟漪波段（

${rb}$

= 80–250 Hz).

= 80–250 Hz）。

Full size image

全尺寸图像

For the good-outcome patient, we observed higher

对于结果良好的患者，我们观察到更高的

net

网络

(0.74 >

$t{h}_{F}$

), higher

），更高

res

结果

(100% >

$t{h}_{O}$

), and lower

），并且更低

res

结果

(6 mm <

(6毫米 <

$t{h}_{D}$

) for the IEN (compared to background;

) 对于IEN（相对于背景；

net

网

$=$

0.51,

0.51，

res

结果

$=$

18%, and

18%，并且

res

结果

31 mm) in

31毫米）在

$\theta$

band (Fig.

乐队（图。

). The IEN had higher AUC with resection (0.92) and SOZ (0.93) compared to background network (0.62 and 0.43, respectively) (Fig.

）。与背景网络（分别为0.62和0.43）相比，IEN在切除（0.92）和SOZ（0.93）方面具有更高的AUC（图。

). In terms of network’s temporal map concordance with IED annotations (Fig.

). 在网络的时间图与IED注释的一致性方面（图。

), the IEN had higher AUC (0.75) compared to background (0.24) (Fig.

)，IEN的AUC（0.75）高于背景（0.24）（图。

）。

For the poor-outcome patient, both networks in

对于预后不良的患者，两个网络均

$\theta$

band had

乐队有

net

网

$t{h}_{F}$

，

res

结果

$t{h}_{O}$

, and

，以及

res

结果

$t{h}_{D}$

(Fig.

（图。

). The IEN was able to predict resection with an AUC = 0.76. Both networks had low AUC when predicting the SOZ (0.53 and 0.63, respectively) (Fig.

). IEN能够以AUC=0.76预测切除。两个网络在预测SOZ时的AUC均较低（分别为0.53和0.63）（图。

). In terms of network’s temporal map concordance with IED annotations (Fig.

). 关于网络时间地图与IED注释的一致性（图。

), both networks had low AUC (0.43 and 0.58, respectively). Ultimately, the IEN identified candidate resection areas. Notably, while these areas overlapped with resection in the good-outcome patient, they covered wider areas and were far away from resection in the poor-outcome patient.

），两个网络的AUC均较低（分别为0.43和0.58）。最终，IEN确定了候选切除区域。值得注意的是，这些区域在预后良好的患者中与切除区域重叠，但在预后较差的患者中覆盖了更广泛的区域且远离切除区域。

Discussion

讨论

We propose a novel patient-specific framework that extracts the IEN from interictal iEEG data using unsupervised ML, delineates the EZ, and predicts outcome in patients with DRE. The proposed framework automatically transforms the interictal iEEG data into temporal maps and brain networks and categorizes them as background and epileptogenic.

我们提出了一种新颖的患者特定框架，该框架使用无监督机器学习从发作间期的颅内脑电图数据中提取IEN，划定EZ，并预测DRE患者的治疗结果。所提出的框架自动将发作间期的颅内脑电图数据转换为时间图和脑网络，并将其分类为背景性和癫痫性。

The IEN presents high-power coherent activity which corresponds to IEDs and ripples in temporal maps; surgical resection of this network predicts good outcome. Its performance is better than the one of the entire network. Our framework works on both data with frequent and sparse IEDs as well as different implantation types reducing the burden of manually identifying epileptogenic iEEG time-windows and localizing the corresponding underlying epileptogenic generator..

IEN呈现出与IEDs和时间图上的涟漪相对应的高功率相干活动；对该网络进行手术切除预示着良好的结果。其性能优于整个网络。我们的框架适用于频繁和稀疏IEDs的数据以及不同类型的植入，减轻了手动识别致痫性iEEG时间窗口和定位相应潜在致痫性发生器的负担。

Interictal biomarkers in the form of elevated power, coherence, source-sink, or functional connectivity have been associated with the EZ

以功率升高、相干性、源-汇或功能连接性为形式的发作间期生物标志物与癫痫发作区（EZ）有关。

，

. However, these biomarkers rely on manual selection of IEDs by experts, which is time-consuming and prone to errors

然而，这些生物标志物依赖于专家手动选择IED，这既耗时又容易出错。

. AI-based tools that delineate the EZ have been recently proposed;

基于人工智能的工具已经被提出用于描绘EZ；

，

yet, they are not widely used in clinical practice possibly because they have been trained and validated only in small potentially biased datasets. Moreover, they rely on information about the EZ provided by experts as a priori, which is prone to errors

然而，它们在临床实践中并没有被广泛使用，可能是因为它们仅在小型潜在偏差数据集上进行了训练和验证。此外，它们依赖专家提供的关于 EZ 的先验信息，这种信息容易出错。

. Furthermore, in many cases, these methods are validated without properly separating testing and training sets, which can lead to overfitting; thus, they suffer from poor generalization and may not perform well on new unseen data

此外，在许多情况下，这些方法在验证时没有正确区分测试集和训练集，这可能导致过拟合；因此，它们的泛化能力较差，在新的未见数据上可能表现不佳。

. Our framework overcomes these hurdles by identifying coherent IENs without the need to detect IEDs or ripples. It adopts an unsupervised ML approach

我们的框架通过识别连贯的IEN来克服这些障碍，而无需检测IED或涟漪。它采用了一种无监督的机器学习方法。

，

which does not rely on a model that is trained using prior information. Moreover, it separates epileptiform from background activity, enabling the identification of epileptogenic regions without the need for manual data processing; thus, it enhances the localization of the EZ in comparison to previously described methods.

这种方法不依赖于使用先验信息训练的模型。此外，它将癫痫样活动与背景活动分离，能够在无需手动数据处理的情况下识别致痫区域，从而相比先前描述的方法提高了对癫痫发作区（EZ）的定位能力。

，

, which rely on aggregated computations such as averaging. This notion is validated by our findings showing the superiority of the IEN (compared to the entire) to identify critical epileptogenic areas. The entire and IEN had higher power inside resection in several bands for patients with good outcome (Fig.

，这些计算依赖于诸如平均值之类的聚合计算。我们的研究结果证实了这一概念，表明 IEN（相较于整体）在识别关键的致痫区域方面的优越性。对于预后良好的患者，整体和 IEN 在多个频段内切除区域内的功率较高（图。

。

). Moreover, the IEN showed higher focality, greater overlap with resection, and shorter distance to resection compared to the entire network (Fig.

). 此外，与整个网络相比，IEN显示出更高的聚焦性、与切除区域更大的重叠以及更短的切除距离（图。

). This suggests that the IEN is more focal and spatially overlapped with the EZ compared to the entire network. The IEN is superior to the entire network to delineate the EZ achieving higher precision (Fig.

这表明，与整个网络相比，IEN 更为集中且在空间上与 EZ 有更多重叠。IEN 在描绘 EZ 方面优于整个网络，达到了更高的精确度（图。

) and AUC (Fig.

）和AUC（图。

). Therefore, although the entire network predicts the EZ, the IEN outperforms it.

因此，虽然整个网络预测了EZ，但IEN的表现优于它。

Previous studies showed the association of elevated interictal power and functional connectivity with epileptogenicity and surgical outcome

以前的研究表明，发作间期功率和功能连接的增加与癫痫发生性和手术结果有关。

，

. These studies suggest that seizure-generating regions in focal epilepsy patients are isolated (from surrounding regions) and that the likelihood of good outcome is increased by resecting these regions. Previously, the overlap of the identified regions with resection was used as outcome predictor

这些研究表明，局灶性癫痫患者的癫痫发作区域是孤立的（与周围区域隔离），并且切除这些区域可以增加良好预后的可能性。以前，确定的区域与切除区域的重叠被用作结果预测指标。

，

. Our findings support this notion since we observed a decrease in percentage overlap of the IEN with resection in poor-outcome patients (Fig.

我们的研究结果支持了这一观点，因为在预后不良的患者中，我们观察到IEN与切除范围的百分比重叠减少（图。

). Furthermore, we found that the distance of the IEN from resection was higher in poor-outcome patients (Fig.

此外，我们发现预后不良患者的IEN距离切除边缘更远（图。

). We also observed a decrease in the IEN focality in poor-outcome patients (Fig.

). 我们还观察到预后不良患者的IEN病灶局限性降低（图。

). Our findings suggest that the more focal the epileptogenic activity is the higher the chances are for the patient to become seizure-free if these regions are resected. Therefore, in addition to overlap, we employed the IEN focality and its distance from resection as outcome predictors. We also compared the outcome prediction of the IENs and entire networks.

）。我们的研究结果表明，癫痫病灶活动越集中，如果这些区域被切除，患者无癫痫发作的可能性就越高。因此，除了重叠区域外，我们还采用了IEN病灶集中性及其与切除区域的距离作为结果预测指标。我们还比较了IEN和整个网络的结果预测。

Our findings suggest that, although the entire network could predict outcome, the IEN is superior in terms of sensitivity (93% vs. 88%), PPV (81% vs. 78%), NPV (83% vs. 64%), accuracy (81% vs. 72%), and AUC (86% vs. 75%) (Fig. .

我们的研究结果表明，尽管整个网络可以预测结果，但 IEN 在敏感性（93% 对 88%）、阳性预测值（81% 对 78%）、阴性预测值（83% 对 64%）、准确性（81% 对 72%）和 AUC（86% 对 75%）方面表现更优（图。

). In 25 (out of the 27) good-outcome patients, there was an agreement between the EZ (identified by the IEN) and the one delineated by the clinicians (Fig.

在 27 名预后良好患者中的 25 名患者中，IEN 识别的 EZ 与临床医生划定的 EZ 之间达成一致（图。

). In poor-outcome patients, the IEN predicted outcome correctly in 10 out of 16 cases (Fig.

). 在预后不良的患者中，IEN在16例中有10例正确预测了结果（图。

). These findings may indicate other epileptogenic regions, distant from resection, that have not been resected due to overlap of the IEN with eloquent areas [e.g., patient #31 (Fig.

这些发现可能表明其他远离切除区域的致痫区，由于与重要功能区重叠而未被切除（例如，患者#31（图）。

)].

)]。

IEDs are often characterized by different morphologies that occur at varying durations. Spikes often last between 20 and 70 ms, while sharp spikes have a duration of 70–200 ms

碰撞电离事件通常以不同形态出现，并且持续时间各异。尖峰通常持续20到70毫秒，而锐利尖峰的持续时间为70到200毫秒。

，

. In the spectral (frequency) domain, these durations are equivalent to components oscillating in

在频谱（频率）域中，这些持续时间相当于振荡的成分

$\theta$

and

和

$\beta$

bands. Here, we found that the IEN in

带。在这里，我们发现 IEN 在

$\theta$

(and

（并且

$\beta$

) bands outperformed other bands when identifying the EZ and predicting outcome. The IEN in the

）波段在识别EZ和预测结果方面优于其他波段。IEN在

$\theta$

band identified resection with a precision of 85.7%. This is probably due to the concordance of the IENs temporal activity with IEDs whose morphological characteristics were predominantly expressed in the form of

频带识别切除的精度为85.7%。这可能是由于IENs的时间活动与IEDs的一致性，其形态学特征主要以以下形式表现：

$\theta$

waves. Additionally, ANOVA during outcome prediction showed that the IEN in

波浪。此外，结果预测期间的方差分析显示，IEN 在

$\theta$

band had the best discriminative power between good- and poor-outcome patients when all networks in all bands were used as outcome predictors. Our findings are in line with previous studies showing increased functional connectivity in

当所有频段的所有网络都被用作结果预测因子时，频段对良好和不良结果患者具有最佳的区分能力。我们的发现与之前显示功能连接性增加的研究一致。

$\theta$

band in patients with epilepsy when compared to healthy controls

癫痫患者与健康对照组相比的频带

，

. Increased interictal synchronization in

. 发作间期同步性增强

$\theta$

band may not necessarily be a unique characteristic of epilepsy. Instead, it could be a consequence of compensatory mechanisms or disinhibition resulting from the disease. As the brain attempts to compensate for the damage caused by the disease, it may increase the synchronized activity in

频带不一定是癫痫的独特特征。相反，它可能是由于疾病引起的代偿机制或去抑制作用的结果。当大脑试图补偿疾病造成的损伤时，可能会增加同步活动。

$\theta$

band as a compensatory mechanism

作为代偿机制的带

. Several studies have related

。几项研究已经关联了

$\theta$

activity to abnormal plasticity in epilepsy patients with lesions; in these patients, abnormal neural connections may lead to excessive synchronization in

癫痫病灶患者异常可塑性活动；在这些患者中，异常的神经连接可能导致过度同步化

$\theta$

band

乐队

，

。

Automated IED detectors

自动IED探测器

，

, including AI-based tools

，包括基于人工智能的工具

，

, have been designed to detect IEDs in iEEG recordings by exhibiting varying performances (specificity: 68–94%, sensitivity: 47–99%, and accuracy: 68–100%)

，通过展示不同的性能（特异性：68-94%，敏感性：47-99%，准确率：68-100%）已被设计用于在iEEG记录中检测IED。

，

. These tools do not have the ability to assess the spatial extent of the EZ without further processing. Tools that allow the spatiotemporal mapping of interictal data have the potential to identify the spatial and temporal extent of epileptiform activity

这些工具无法在没有进一步处理的情况下评估 EZ 的空间范围。允许对发作间期数据进行时空映射的工具有潜力识别癫痫样活动的空间和时间范围。

，

. In line with this notion, our framework identified the IEN whose temporal activity was concordant with IED and ripple annotations with 77% accuracy (in

与此概念一致，我们的框架识别出了与IED和涟漪注释在时间活动上相一致的IEN，准确率达到77%（在...

某人

) (Supplementary Figure

) （补充图

). This was achieved by analyzing interictal iEEG recordings of short duration; this is in line with previous studies showing that short time-windows are sufficient to map the spatiotemporal dynamics of iEEG data with high consistency

）。这是通过分析短暂的间歇期iEEG记录实现的；这与之前的研究一致，表明短时间窗口足以高一致性地映射iEEG数据的时空动态。

，

. Additionally, our methodology robustly estimated the IENs regardless of the IEDs frequency (Fig.

此外，我们的方法能够稳健地估计IENs，而不论IEDs的频率如何（图。

）。

Previous studies have shown that the interpretation of epileptiform activity can be influenced by the implantation type used in each patient

以往的研究表明，癫痫样活动的解释可能受到每位患者所使用的植入类型的影响。

，

. In particular, sEEG employs a network-based implantation strategy that differs fundamentally from subdural and depth electrodes

特别是，sEEG采用了一种基于网络的植入策略，这与硬膜下电极和深部电极有根本区别。

，

. sEEG has the ability to capture propagating epileptiform activity beyond the EZ leading to more distributed activity across multiple brain regions. On the other hand, subdural electrodes record activity exclusively at the cortical level

.sEEG能够捕捉超出致痫区（EZ）传播的癫痫样活动，导致多个脑区之间更广泛的活动。另一方面，硬膜下电极仅记录皮层水平的活动。

. Both modalities have limited spatial resolution since they record epileptiform activity in the direct vicinity of contacts and thus are blind to other brain regions. Therefore, the actual focus of the EZ may be missed leading to surgical failure. Here, we examined whether the implantation type can affect the robustness of our methodology.

两种方式的空间分辨率都有限，因为它们记录的是接触点附近的癫痫样活动，因此对其他脑区是“盲区”。这可能导致实际的致痫区（EZ）焦点被遗漏，从而导致手术失败。在此，我们研究了植入类型是否会影响我们方法的稳健性。

Our analysis showed no differences between the AUCs of IENs with resection among the different implantation types in any frequency band except the .

我们的分析显示，在任何频率范围内，不同植入类型之间具有切除的IEN的AUC没有差异，除了。

$\alpha$

band (Fig.

乐队（图。

6b) and no

6b) 以及没有

differences for the SOZ. The differences between the AUCs of the IENs of subdural and sEEG implantations in the

SOZ的差异。硬膜下和sEEG植入的IEN的AUC之间的差异

$\alpha$

band may be explained by the fact that subdural electrodes are unable to capture dysregulations of thalamocortical interactions observed in patients with epilepsy in

带宽可能由于皮质下电极无法捕捉癫痫患者中观察到的丘脑皮层相互作用失调来解释。

$\alpha$

band

乐队

, which may be captured with sEEG. In summary, our data demonstrate that our framework provides robust findings independent of the implantation type.

，这可能通过sEEG捕捉到。总之，我们的数据表明，我们的框架提供了独立于植入类型的强大发现。

Our retrospective study analyzed iEEG data from a single-center cohort of patients with DRE having different pathologies. Future prospective studies from larger multicenter cohorts would allow us to examine its applicability in clinical settings and in homogeneous groups of patients. Our analysis was performed on short data segments; analysis of longer data segments may enhance the accuracy of the IEN to delineate the EZ and predict outcome.

我们的回顾性研究分析了来自单一中心队列的DRE患者的不同病理学的iEEG数据。未来更大规模多中心队列的前瞻性研究将使我们能够检查其在临床环境和同质患者群体中的适用性。我们的分析是在短数据段上进行的；分析更长的数据段可能会提高IEN划定EZ和预测结果的准确性。

Due to the lack of interictal data segments with absent IEDs in our database, further studies should examine the applicability of our methodology for patients where IEDs are absent in intracranial EEG recordings. Moreover, iEEG does not cover the entire brain but is rather implanted in locations agreed upon during pre-implantation analysis which can be subjective.

由于我们的数据库中缺乏无发作间期癫痫样放电（IEDs）的数据段，因此需要进一步研究我们的方法对颅内脑电图（iEEG）记录中无IEDs的患者的适用性。此外，颅内脑电图并不能覆盖整个大脑，而是植入在术前分析中确定的位置，这一过程可能存在主观性。

While most of the implantation correctly covered the EZ in good-outcome patients, epileptogenic regions may have been missed in poor-outcome patients. Furthermore, the resection and SOZ were subjectively identified and may either miss critical epileptogenic regions or describe areas larger than the actual EZ.

虽然在预后良好的患者中，大部分植入物正确覆盖了癫痫区（EZ），但在预后较差的患者中可能遗漏了致痫区域。此外，切除范围和发作起始区（SOZ）是主观确定的，可能会遗漏关键的致痫区域或描述比实际癫痫区更大的范围。

Advances in whole brain iEEG implantations may overcome these drawbacks. Our framework may not discriminate the EZ from IEDs propagation. Future studies may incorporate the phase information of DMD in order to distinguish between the primary epileptogenic activity and propagating IEDs. Additionally, the phase information may provide directionality of IED propagations providing better understanding of the epileptogenic network dynamics.

全脑 iEEG 植入技术的进步可能会克服这些缺点。我们的框架可能无法区分致痫区 (EZ) 和 IEDs 传播。未来的研究可以结合 DMD 的相位信息，以区分原发性致痫活动和传播中的 IEDs。此外，相位信息可能提供 IED 传播的方向性，从而更好地理解致痫网络的动态。

Future studies could consider applying the proposed framework to a large range of virtual channels, estimated through electrical source imaging performed on iEEG recordings. This approach allows the characterization of neural activity i.

未来的研究可以考虑将所提出的框架应用于通过颅内脑电图（iEEG）记录进行电源成像估计出的广泛虚拟通道。这种方法允许对神经活动进行表征。

，

六十三

. Thus, it may allow capturing the entire phenomena of IEDs propagation potentially discriminating it from the primary EZ. Finally, our framework presumes that only two networks were active. Future studies should examine the exact number of networks that were active during interictal iEEG and duration of data analysis under different scenarios.

因此，它可能允许捕捉IEDs传播的整个现象，从而潜在地区分出原发EZ。最后，我们的框架假设只有两个网络处于活跃状态。未来的研究应考察在发作间期iEEG期间确切的活跃网络数量，以及在不同情景下的数据分析时长。

Future directions may include clinically applicable 3D surgical navigation systems that integrate the IEN enabling the precise identification of iEEG contacts to resect or ablate..

未来的发展方向可能包括临床适用的3D手术导航系统，该系统集成IEN，能够精确定位需要切除或消融的iEEG触点。

To conclude, our proposed framework identifies an IEN which delineates the EZ and predicts surgical outcome from interictal iEEG data of patients with DRE. The framework’s automated nature reduces post-acquisition preprocessing cost and time, reducing risks associated with the presurgical evaluation.

总之，我们提出的框架识别出一个IEN，该网络描绘了EZ，并从耐药性癫痫患者的发作间期iEEG数据中预测手术结果。该框架的自动化特性降低了采集后的预处理成本和时间，减少了与术前评估相关的风险。

Such a framework would be particularly useful to epilepsy centers that lack the multidisciplinary expertise to delineate accurately and precisely the EZ in complex cases of patients with DRE. If validated prospectively, our framework may potentially replace ictal recordings to localize the EZ..

这样的框架对于缺乏多学科专业知识来准确和精确划定耐药性癫痫（DRE）患者复杂病例中致痫区（EZ）的癫痫中心尤其有用。如果前瞻性验证，我们的框架可能潜在替代发作期记录以定位致痫区。

Methods

方法

Ethics statement

伦理声明

The protocol was approved by North Texas Regional IRB (2019-166; PI: C. Papadelis) that waived the need for informed consent considering the study’s retrospective nature. All methods and analyses were performed in accordance with relevant guidelines and regulations.

该协议经北德克萨斯地区IRB（2019-166；PI：C. Papadelis）批准，鉴于研究的回顾性性质，免除了知情同意的要求。所有方法和分析均按照相关指南和规定进行。

Study cohort

研究队列

We retrospectively analyzed iEEG data recorded from 43 children and young adults with DRE who had resective neurosurgery at Boston Children’s Hospital between June 2011 and June 2018. We selected patients based on the following criteria: (i) availability of at least 5-minute interictal iEEG data; (ii) availability of post-implantation computerized tomography (CT); (iii) availability of preoperative and postoperative MRIs; (iv) information about the resection volume and the clinically defined SOZ; and (v) availability of post-surgical outcome at least one year after surgery.

我们回顾性分析了2011年6月至2018年6月期间在波士顿儿童医院接受切除性神经外科手术的43名患有药物难治性癫痫（DRE）的儿童和年轻成人的颅内脑电图（iEEG）数据。我们根据以下标准选择了患者：(i) 至少5分钟的发作间期iEEG数据可用；(ii) 植入后的计算机断层扫描（CT）可用；(iii) 术前和术后的磁共振成像（MRI）可用；(iv) 关于切除体积和临床定义的发作起始区（SOZ）的信息；以及 (v) 手术后至少一年的术后结果可用。

The outcome was determined by a pediatric epileptologist (J.B.) after follow-up visits. We use the Engel score to classify patients as good (Engel.

结果由儿科癫痫专家（J.B.）在随访后确定。我们使用Engel评分将患者分类为良好（Engel）。

$=$

I, seizure-free) or poor outcome (Engel

我，无癫痫发作）或不良结果（Engel

$>$

I, non-seizure-free). The patient demographic and clinical information are provided in Supplementary Table

我，非无癫痫发作）。患者的人口统计学和临床信息见补充表格。

。

Interictal iEEG recordings

发作间期颅内脑电图记录

The multidisciplinary clinical team decided the locations of implanted iEEG electrodes independently of this study. Long-term iEEG data were acquired with subdural electrodes, sEEG implantations, and subdural and depth electrodes using XLTEK NeuroWorks (Natus Inc., USA). The number and type of implanted electrodes are presented in Supplementary Table .

多学科临床团队独立于本研究决定植入iEEG电极的位置。长期iEEG数据通过硬膜下电极、sEEG植入、以及使用XLTEK NeuroWorks（Natus Inc.，美国）的硬膜下和深部电极获取。植入电极的数量和类型见补充表。

. Subdural electrodes were 2–3 mm in diameter with a 10 mm inter-contact distance, whereas depth electrodes were composed of 6 to 16 linearly arranged contacts ~1.5–2.5 mm apart. Each contact had a diameter of 0.8 mm and a length of 2 mm. The average acquisition time using iEEG implantations was 5.6 days, 12.6 hours, and 25.3 minutes.

硬膜下电极直径为2-3毫米，接触点间距为10毫米，而深部电极由6至16个线性排列的接触点组成，接触点间距约为1.5-2.5毫米。每个接触点的直径为0.8毫米，长度为2毫米。使用颅内脑电图植入的平均采集时间为5天6小时12分25.3秒。

The sampling frequency ranged between 1000 and 2048 Hz. We selected 5-minute interictal segments with frequent IEDs from non-rapid eye movement slow-wave sleep (whenever applicable) at least one hour before/after clinical seizures or half an hour before/after electrographic seizures.

采样频率范围在1000到2048赫兹之间。我们选择了5分钟的频繁发作间期癫痫样放电（IEDs）片段，这些片段来自非快速眼动慢波睡眠（如适用），并且距离临床发作至少前后一小时或脑电图发作前后半小时。

. This selection ensured the inclusion of segments with the highest IED rate and minimal motion artifacts

此选择确保了包含最高IED率和最小运动伪影的片段

，

。

Localization of iEEG electrodes

iEEG电极的定位

MRI scans with standard magnetization-prepared rapid acquisition gradient-echo sequences were performed before and after resection using a 3 T scanner (TIM TRIO, Siemens AG). CT scans were performed after iEEG implantation. We localized the iEEG electrode coordinates by coregistering the post-implantation CT scans with the preoperative MRI using .

使用3T扫描仪（TIM TRIO，西门子）在切除术前后进行了标准磁化准备快速梯度回波序列的MRI扫描。在颅内脑电图（iEEG）植入后进行了CT扫描。我们通过将植入后的CT扫描与术前MRI进行核心配准来定位iEEG电极坐标。

Brainstorm

头脑风暴

(Fig.

（图。

)

. We adjusted the electrode locations to compensate for possible brain shift occurring after electrocorticography implantation

我们调整了电极位置，以补偿皮层电图植入后可能发生的大脑移位。

。

Defining the resection and SOZ

定义切除区域和发作起始区（SOZ）

To define resection, we coregistered the preoperative and postoperative MRIs and manually drew the resection volume boundary on consecutive slices using

为了定义切除范围，我们对术前和术后的MRI进行了核心配准，并使用连续切片手动绘制了切除体积的边界。

Brainstorm

头脑风暴

(Fig.

（图。

)

. Pediatric epileptologists defined the SOZ through visual inspection of ictal data independently from this study. iEEG contacts whose dynamics changed with seizure onsets were marked as SOZ contacts. To define the EZ, we considered as gold standards the resected electrodes and the clinically defined SOZ electrodes (Fig.

小儿癫痫专家通过独立于本研究的视觉检查发作期数据来定义致痫区（SOZ）。动态变化与癫痫发作起始相关的颅内脑电图（iEEG）触点被标记为致痫区触点。为了定义癫痫灶（EZ），我们将切除的电极和临床定义的致痫区电极视为金标准（图。

。

）。

Preprocessing of iEEG data

iEEG数据的预处理

We initially preprocessed the data by applying a DC offset removal and common average referencing. We then notch-filtered the data to remove the 60 Hz power line noise and its harmonics, and bandpass filtered them in two frequency bands:

我们首先通过应用直流偏移去除和共同平均参考对数据进行了预处理。然后，我们对数据进行了陷波滤波以去除60 Hz电源线噪声及其谐波，并在两个频带中对它们进行了带通滤波：

${sb}$

{sb}

(1–80 Hz) and

（1–80 Hz）和

${rb}$

(80–250 Hz)

. Bad channels and channels with artifacts were excluded from this study.

坏通道和有伪迹的通道被排除在本研究之外。

Feature extraction using DMD

使用DMD进行特征提取

The first step of our framework extracts coherent spatial features from interictal iEEG data. Using a sliding window approach, each time-window is processed with DMD

我们框架的第一步是从发作间期的颅内脑电图（iEEG）数据中提取连贯的空间特征。通过使用滑动窗口方法，每个时间窗口都通过动态模式分解（DMD）进行处理。

. DMD decomposes high-dimensional time-series data into a set of coherent structures that exhibit similar linear dynamics in time, in the form of oscillations and exponential growth and decay

DMD将高维时间序列数据分解为一组在时间上表现出相似线性动力学的相干结构，形式包括振荡、指数增长和衰减。

. Since the spatial coherent patterns identified by DMD are relative to inherent temporal dynamics, these structures can be considered as functionally connected networks in terms of coherence

由于DMD识别的空间相干模式与内在的时间动力学相关，因此这些结构可以被视为在相干性方面具有功能连接的网络。

，

。

DMD approximates locally (in a specific time-window) a non-linear dynamical system linearly in the discrete domain as:

DMD 在离散域中通过线性方式局部近似（在特定时间窗口内）非线性动力系统，表达式如下：

$${x}_{t+1}=A{x}_{t}+{w}_{t},t=0,1,2,...,m-1$$

(1)

where

其中

${x}_{t}\in {{\mathbb{R}}}^{n}$

denotes the measurements from

表示来自以下的测量值

$n$

channels at instant

瞬间通道

$t$

，

$A$

is an

是

$n\times n$

matrix,

矩阵，

$m$

is the number of samples, and

是样本的数量，且

${w}_{t}$

represents the residual error. DMD finds in the least square sense, a low-rank eigen-decomposition of

表示残差误差。DMD 在最小二乘意义上找到了一个低秩的特征分解

$A$

by minimizing:

通过最小化：

$${{||}{x}_{t+1}-A{x}_{t}{||}}_{2}$$

(2)

The matrix

矩阵

$X\in {{\mathbb{R}}}^{n\times m}$

resulting from horizontally stacking

由水平堆叠产生的

$m$

measurements taken every

每次测量采取

$\Delta t$

can be represented as:

可以表示为：

$$X=\left(\begin{array}{ccc}| & | & |\\ {x}_{0} &\;\;\;\;\; {x}_{1}\ldots &\;\;\;\;\;{x}_{m-1}\\ | & | & |\end{array}\right)$$

$$X=\left(\begin{array}{ccc}| & | & | \\x_{0} & \;\;\;\;\; x_{1}\ldots & \;\;\;\;\; x_{m-1} \\| & | & |\end{array}\right)$$

(3)

$X$

can represent a time-window of iEEG with

可以表示 iEEG 的时间窗口

$n$

-channels and

-通道和

米

samples.

样本。

Constructing

构建中

$X{\prime} \in {{\mathbb{R}}}^{n\times m}$

which contains one-time shifted version of

其中包含一次性移位版本的

$X$

：

$$X{\prime} =\left(\begin{array}{ccc}| & | & |\\ {x}_{1} &\;\;\;\;\;\;{x}_{2}\ldots &\;\;\;{x}_{m}\\ | & | & |\end{array}\right)$$

$$X' = \begin{pmatrix}| & | & | \\x_1 & x_2 \ldots & x_m \\| & | & |\end{pmatrix}$$

(4)

Equation (

方程（

) can be written as:

) 可以写成：

$${X}^{\prime} ={AX}$$

$$X^{\prime} = AX$$

(5)

The solution of the least squares problem approximates A:

最小二乘问题的解近似于 A：

$$A={X}^{{\prime} }{X}^{\dagger }$$

(6)

where † is the Moore-Penrose pseudo-inverse. The eigenvectors and eigenvalues of

其中 † 表示摩尔-彭罗斯伪逆。特征向量和特征值的

$A$

correspond to DMD modes and eigenvalues. DMD computes a lower dimensional representation of the data to find an approximation

对应于DMD模式和特征值。DMD计算数据的低维表示以找到近似值。

$\widetilde{A}$

的

$A$

by keeping the leading

通过保持领先

$r$

eigenvalues, eigenfunctions, and modes. The output of DMD is the DMD tuple

特征值、特征函数和模式。DMD 的输出是 DMD 元组。

$(\lambda ,\psi ,\phi )$

defined by:

定义为：

The eigenvalues

特征值

${\lambda }_{j}$

from which the frequency of oscillations

由此得出振荡频率

${f}_{j}$

of the eigenfunctions can be computed where

特征函数的计算可以进行，其中

$\scriptstyle{f}_{j}=|\frac{{imag}({\omega}_{j})}{2\pi}|$

$\scriptstyle{f}_{j}=|\frac{{虚部}({\omega}_{j})}{2\pi}|$

，

${\omega }_{j}=\log ({\lambda }_{j})/\Delta t$

，

${imag}(.)$

is the imaginary part of a complex number where

是复数的虚部，其中

$j=\mathrm{1,2},\ldots ,r$

$j=1,2,\ldots ,r$

。

The eigenfunctions

特征函数

$\psi ={e}^{\varOmega t}$

oscillating (with or without growth or decay) at a frequency

以某一频率振荡（有或没有增长或衰减）

${f}_{j}$

。

The spatial modes

空间模式

$\phi$

representing the weights that reconstruct the data from the

表示从数据中重建的权重

$r$

eigenfunctions.

特征函数。

Therefore,

因此，

$$X\approx \phi {e}^{\varOmega t}b$$

(7)

where

其中

$b$

is computed from initial conditions

从初始条件计算得出

${x}_{0}$

such that

使得

${x}_{0}=\phi b$

。

DMD fails if

DMD 失败的情况是

$m\,\gg\, n$

since

自从

$n$

modes are insufficient to accurately approximate the time dynamics in

模式不足以准确近似时间动态

$m$

samples

样本

. Hankel time delay embeddings can overcome this deficiency by augmenting the data by

汉克尔时间延迟嵌入可以通过扩充数据来克服这一缺陷。

$h$

measurements with past history

具有历史记录的测量

. To guarantee an optimal representation, we used

为了保证最佳的表示，我们使用了

$h$

time delay embeddings satisfying

满足时间延迟嵌入

${hn}$

$2m$

. For more detail about DMD and Hankel time delay embeddings, refer to Supplementary Note

有关DMD和汉克尔时间延迟嵌入的更多细节，请参阅补充说明。

. DMD applied on a sample time-window is described in Supplementary Note

应用于样本时间窗口的DMD在补充说明中描述

。

In our study, we first filtered the 5-minute-long interictal iEEG segments in

在我们的研究中，我们首先过滤了 5 分钟长的发作间期 iEEG 片段

${sb}$

{sb}

and

和

${rb}$

(Fig.

（图。

). The segments were dissected into

）。节段被解剖为

$L$

(

${L}_{1}$

$L_1$

在

某人

and

和

${L}_{2}$

在

) time-windows with

) 时间窗口与

$m$

samples each using a sliding window with 95% overlap. We used 250 ms time-windows for

每个样本使用95%重叠的滑动窗口。我们使用了250毫秒的时间窗口。

${sb}$

{某人}

and 150 ms time-windows for

以及150毫秒的时间窗口用于

${rb}$

since they were sufficient to accurately capture the characteristic waveforms of IEDs and ripples

由于它们足以准确捕捉IED和涟漪的特征波形

，

. Since a 250 ms time-window spans only one cycle of

由于250毫秒的时间窗口仅跨越一个周期

$\delta$

activity, two cycles of

活动，两个周期

$\theta$

activity, and three cycles of

活动，以及三个周期的

$\alpha$

activity, estimating the signal power reliably with a brief time-window is challenging mostly due to the windowing edge effect. Yet, unlike the discrete Fourier and continuous wavelet transforms, which are constrained by the window size for capturing low frequency components, the DMD is not subject to these limitations.

活动时，由于窗口边缘效应的影响，在短时间内可靠地估计信号功率具有挑战性。然而，与离散傅里叶变换和连续小波变换不同，它们受窗口大小限制以捕捉低频成分，而DMD不受这些限制。

. Additional analysis to validate the ability of DMD to accurately estimate the frequency components from 250 ms time-windows are provided in Supplementary Note

补充说明中提供了额外的分析，以验证DMD从250毫秒时间窗口准确估计频率成分的能力。

。

Each time-window

每次时间窗口

${X}_{i}$

(

$i=\mathrm{1,2},\ldots ,L$

) was processed using DMD with

) 使用DMD处理了

${r}_{1}$

modes in

模式在

${sb}$

{某人}

and

和

${r}_{2}$

modes in

模式在

${rb}$

to extract spatial mode matrices (

提取空间模式矩阵 (

${\phi }_{i}^{{sb}}\in {{\mathbb{C}}}^{n\times {r}_{1}},{\phi }_{i}^{{rb}}\in {{\mathbb{C}}}^{n\times {r}_{2}}$

) and eigenfunctions (

）和特征函数（

${\psi }_{i}^{{sb}}\in {{\mathbb{C}}}^{{r}_{1}\times m},{\psi }_{i}^{{rb}}\in {{\mathbb{C}}}^{{r}_{2}\times m}$

) with their corresponding oscillation frequencies (

）及其对应的振荡频率（

${f}_{i}^{{sb}}{{\mathbb{\in }}{\mathbb{R}}}^{{r}_{1}},{f}_{i}^{{rb}}{{\mathbb{\in }}{\mathbb{R}}}^{{r}_{2}}$

${f}_{i}^{{sb}} \in {\mathbb{R}}^{{r}_{1}}, {f}_{i}^{{rb}} \in {\mathbb{R}}^{{r}_{2}}$

). We justified and used

). 我们证明了并使用了

${r}_{1}=50$

and

和

${r}_{2}=100$

(Supplementary Note

（补充说明

）。

The magnitude of the spatial mode matrices (

空间模式矩阵的大小 (

$\left|{\phi }_{i}^{{sb}}\right|,\left|{(\phi }_{i}^{{rb}}\right|$

) in each time-window

）在每个时间窗口中

$i$

at frequencies (

在频率（

${f}_{i}^{{sb}},{f}_{i}^{{rb}}$

) are computed to generate DMD spectra of each channel (Fig.

) 计算以生成每个通道的DMD频谱 (图。

). We define the following physiologically relevant frequency bands: delta (

）。我们定义了以下生理相关的频段：δ（

$\delta =$

1–4 Hz), theta (

1–4 Hz)，theta (

$\theta =\,$

[4–8] Hz), alpha (

[4-8] Hz)，alpha (

$\alpha =$

[8–12] Hz, beta (

[8-12] Hz，β（

$\beta =\,$

[12–30] Hz), gamma (

[12-30] Hz)，伽马(

$[$

30–80] Hz), and broadband (

30-80] Hz），以及宽带（

${sb}=$

[1–80] Hz). For

[1–80] Hz)。对于

${sb}$

, DMD spectra powers

，DMD光谱功率

$\left|{\phi }_{i}^{{sb}}\right|$

are distributed in the six defined bands by collecting the average power of the modes whose frequency of oscillation

通过收集振荡频率在六个已定义频带中的模式的平均功率来分布。

${f}_{i,j}^{{sb}}$

(

$j=\mathrm{1,2},\ldots ,{r}_{1}$

$j=1,2,\ldots ,r_{1}$

) falls within the band boundaries. For

）在带边界范围内。对于

${rb}$

, the average spectrum is computed using all

，使用所有数据计算平均频谱

${r}_{2}$

components. We denote by

组件。我们用

$Y$

the feature matrices that contain the DMD spectra of all the time-windows in both

包含所有时间窗口的DMD频谱的特征矩阵

${sb}$

and

和

${rb}$

(Fig.

（图。

）。

Extraction of the entire network

整个网络的提取

The entire network was computed by averaging the feature matrix

整个网络通过平均特征矩阵计算得出。

$Y$

across time-windows for each patient in each band. These networks represent average DMD power of channels across the entire iEEG segment serving as a baseline to compare with the proposed approach. The resulting vectors (

跨每个患者在每个频段的时间窗口。这些网络代表了整个 iEEG 段中通道的平均 DMD 功率，作为与所提出方法进行比较的基线。生成的向量 (

$\in {{\mathbb{R}}}^{n\times 1}$

) are thresholded using the mean plus one standard deviation as threshold.

) 使用均值加一个标准差作为阈值进行阈值处理。

Dominant brain network extraction using NNMF

使用NNMF提取主导脑网络

The next step of our processing pipeline involved the identification of the consistently active interictal networks that occur recurrently across time-windows. We used NNMF, a dimensionality reduction and an unsupervised ML method, to extract these dominant spatial configurations and their corresponding temporal activity.

我们处理流程的下一步涉及识别在时间窗口间反复出现的持续活跃的发作间网络。我们使用了非负矩阵分解（NNMF），一种降维且无监督的机器学习方法，来提取这些主要的空间构型及其对应的时间活动。

。

NNMF decomposes a non-negative matrix

NNMF分解一个非负矩阵

$Y$

(

$\in {{\mathbb{R}}}^{n\times L}$

) into two non-negative matrices

) 分解为两个非负矩阵

$W$

(

$\in {{\mathbb{R}}}^{n\times k}$

) and

) 和

$H$

(

$\in {{\mathbb{R}}}^{k\times L}$

) such that

) 使得

$Y\approx W\times H$

where

其中

$k$

is an integer

是一个整数

$< \min \{n,m\}$

$ < \min \{n,m\} $

that defines the number of spatial basis functions to extract. NNMF estimates

定义要提取的空间基函数的数量。NNMF 估计

$W$

and

和

$H$

by minimizing the reconstruction error and constraining both

通过最小化重建误差并约束两者

$W$

and

和

$H$

to be positive:

保持积极：

$$\underbrace{{minimize}}_{W,H}{{||}Y-{WH}{||}}_{2}{subject}\; {to}W\,>\,0,{H}\,>\,0$$

$$\underbrace{\text{minimize}}_{W,H}{{||}Y-{WH}{||}}_{2}\;\text{subject to}\; W>0,\,H>0$$

(8)

Typically,

通常，

$k$

is a user-defined parameter.

是一个用户定义的参数。

$W$

is termed the basis matrix and contains the

被称为基矩阵，包含

$k$

-dimensional representation of the data, while

数据的维表示，而

$H$

is the coefficient matrix which contains reconstruction coefficients that reconstructs the original data

是包含重构原始数据的重构系数的系数矩阵。

$Y$

from the

来自

$k$

basis functions in

基函数在

$W$

。

In our study, we used NNMF to process the feature data matrix

在我们的研究中，我们使用了NNMF来处理特征数据矩阵。

$Y$

in each band separately to extract

分别在每个波段中提取

$k$

dominant spatial basis functions (

主导的空间基函数 (

$\in {{\mathbb{R}}}^{n\times 1}$

) that represent the recurrent spatial configurations across the time-windows in the form of brain networks (columns of

）以脑网络（列的形式）表示跨时间窗口的反复空间配置。

$W$

) and their corresponding temporal activity strengths (corresponding rows in

）及其相应的时间活动强度（对应行在

$H$

). Supplementary Note

). 补充说明

describes NNMF on sample data. We assumed two distinct active networks (i.e., interictal epileptogenic and background) in interictal segments and set

描述了样本数据上的NNMF。我们假设在发作间期片段中有两个不同的活跃网络（即，发作间期致痫性和背景网络）并设置

$k=2$

. We justified our choice using the dispersion coefficient method

我们使用色散系数方法证明了我们的选择是合理的。

in Supplementary Note

在补充说明中

。

Determining activity of brain networks across time

确定大脑网络随时间的活动

By assuming that one network is active in each time-window, we quantized the coefficient matrix

通过假设每个时间窗口内有一个网络处于活跃状态，我们对系数矩阵进行了量化。

$H$

per column by finding the index

按列查找索引

$q$

with maximum coefficient. In other words, for each time-window, where

具有最大系数。换句话说，对于每个时间窗口，其中

$q\in \left\{\mathrm{1,2}\right\}$

$q\in \{1,2\}$

, we chose the index of the maximum per column:

，我们选择了每列最大值的索引：

$$V={\arg }\mathop{\max }\limits_{q}({H}_{j})$$

(9)

The vector

向量

$V$

forms the temporal map, showing the indices of the active networks across time-windows. The basis functions (columns of

形成时间映射，显示跨时间窗口的活跃网络索引。基函数（列

$W$

) are thresholded using the mean plus one standard deviation per column, forming the two networks (Fig.

) 使用每列的平均值加一个标准差进行阈值处理，形成两个网络（图。

1天

) (Supplementary Note

) （补充说明

）。

Network categorization

网络分类

After identifying the two networks and their corresponding temporal maps, we categorized them as epileptogenic and background. We presumed that the IEN contaminates the background activity intermittently across time. Therefore, we considered the network with the least frequent activity in the temporal map to be the epileptogenic (Fig.

在识别出两个网络及其相应的时间映射后，我们将它们分类为致痫性和背景性。我们假设致痫网络 (IEN) 会在时间上间歇性地污染背景活动。因此，我们认为时间映射中活动频率最低的网络是致痫性网络（图。

。

）。

Generating stable networks and temporal maps

生成稳定的网络和时间映射

Since NNMF factorizes the matrices starting from random initial values, for each patient and band we repeated the following three steps 30 times: NNMF, identification of networks and temporal maps, and categorization of networks as epileptogenic and background. We used k-means clustering

由于NNMF从随机初始值开始分解矩阵，对于每位患者和每个频段，我们重复以下三个步骤30次：NNMF、网络及时间图的识别、将网络分类为癫痫相关和背景。我们使用了k均值聚类。

to smooth the indices of the temporal map

平滑时间映射的索引

$V$

. Since k-means assigns the cluster numbers randomly, we matched the indices of

. 由于k-means随机分配聚类编号，我们匹配了

$V$

with the ones generated with k-means (Supplementary Note

与使用k-means生成的那些（补充说明

). We then averaged the resulting 30 IENs and background networks, as well as the temporal maps.

). 随后，我们对生成的30个IEN和背景网络以及时间图进行了平均处理。

Network properties

网络属性

We estimated and compared DMD power of the entire, interictal epileptogenic, and background networks inside and outside the resection and SOZ for good- and poor-outcome patients, separately (

我们分别估计并比较了良好预后和不良预后患者切除区域内外及癫痫发作间期致痫网络和背景网络的整体DMD功率（

Wilcoxon signed-rank

威尔科克森符号秩检验

test). We then computed three measures to characterize the networks in reference to the resection volume. For each of the three networks (i.e., entire, interictal epileptogenic, and background), we estimated the

测试）。然后，我们计算了三个指标来表征与切除体积相关的网络。对于三个网络中的每一个（即整个网络、发作间期致痫网络和背景网络），我们估算了

net

网

，

res

结果

, and

，以及

res

结果

(Fig.

（图。

1天

). We defined

）。我们定义了

net

网

as the normalized reciprocal of the average Euclidean distance between electrodes of the thresholded networks. We defined

作为阈值网络中电极之间平均欧几里得距离的归一化倒数。我们定义了

res

资源

as the percentage of electrodes of the thresholded network within 10 mm from resection. Finally, we defined

作为距离切除区域10毫米内阈值网络电极的百分比。最后，我们定义了

res

结果

as the average Euclidean distance of the thresholded electrodes in a network to resection. For each band, we compared

作为网络中阈值电极到切除区域的平均欧几里得距离。对于每个频段，我们进行了比较

net

网

，

res

资源

, and

，以及

res

资源

of the entire, interictal epileptogenic, and background networks (

整个、发作间期致痫和背景网络 (

Wilcoxon signed-rank

威尔科克森符号秩

test). The selection of the 10 mm cut-off was based on studies that showed that the gyral width is between 11 to 21 mm

测试）。选择10毫米的截止值是基于研究表明脑回宽度在11到21毫米之间的研究。

。

Test for consistency and robustness of extracted networks

测试提取网络的一致性和鲁棒性

To test the robustness of our methodology, we performed five-fold cross-validation. For each patient, we dissected the 5-minute-long iEEG segment into five 1-minute-long segments with no overlap forming five folds. For each fold, we estimated the IEN and the background networks from the 1-minute segment and the remaining 4-minute segment.

为了测试我们方法的鲁棒性，我们进行了五折交叉验证。对于每位患者，我们将5分钟长的iEEG片段分割为五个1分钟长的不重叠片段，形成五折。对于每一折，我们从1分钟片段和剩余的4分钟片段中分别估计了IEN和背景网络。

For each fold, we then computed the Dice.

对于每个折叠，我们随后计算了Dice系数。

score between the networks (IEN and background) identified from the 1-minute segment and the ones identified using the remaining 4-minute segments. We then computed the average Dice score across folds. Refer to Supplementary Figure

从1分钟段中识别出的网络（IEN和背景）与使用剩余4分钟段识别出的网络之间的得分。然后我们计算了跨折叠的平均Dice得分。请参阅补充图。

for an illustration of the methodology. The analysis evaluates the consistency of the identified networks when different segments of the same data are analyzed. In general, a Dice score

作为方法论的插图。该分析评估了在分析相同数据的不同片段时所识别网络的一致性。通常，Dice 分数

$>$

0.8 is considered to have almost perfect agreement

0.8 被认为具有几乎完美的一致性

. The average Dice score of the five-fold tests were computed for both the IEN and background network.

五折测试的平均Dice得分分别针对IEN和背景网络进行了计算。

As a form of surrogate testing

作为一种替代测试的形式

, for each band and patient, we first generated 100 amplitude adjusted phase shuffled surrogates of the feature matrix derived from the 5-minute iEEG. This procedure keeps the power information while randomizing the temporal relationships by shuffling the order of the values across time-windows. The randomized feature matrices are processed using NNMF to extract the IENs and the background networks, which are termed here as random networks.

对于每个频段和患者，我们首先基于5分钟的iEEG生成了100个振幅调整后的相位随机替代特征矩阵。该过程通过在时间窗口间打乱值的顺序来保留功率信息，同时随机化时间关系。使用NNMF处理这些随机化的特征矩阵以提取IENs和背景网络，这些网络在此被称为随机网络。

The methodology can help verify that the identified networks are not due to random fluctuations in the data, but rather due to patterns that repeat consistently across time. We then computed the average Dice scores of the random networks with the IEN and the background networks estimated from the 5-minute-long data segments..

该方法有助于验证所识别的网络不是由于数据中的随机波动，而是由于在时间上持续重复的模式。然后，我们计算了随机网络与从5分钟长的数据段中估计出的IEN和背景网络的平均Dice得分。

Predicting the EZ

预测EZ

To assess the ability of the entire, interictal epileptogenic, and background networks to predict the EZ, we used the power of the networks as predictors and the resection as target. We performed the following analysis in each band separately and only for good-outcome patients since their resected tissue is assumed to contain the actual EZ.

为了评估整个网络、发作间期致痫网络和背景网络预测癫痫区（EZ）的能力，我们使用网络的功率作为预测因子，以切除区域为目标。由于假设切除组织包含实际的癫痫区，我们仅对预后良好的患者分别在每个频段进行了以下分析。

For each network, we considered electrodes whose power was above the threshold as active and the rest as inactive. Electrodes ≤10 mm from resection were considered to be inside resection. We then defined as: (i) true positives (TP), the number of active electrodes that were inside resection; (ii) false positives (FP), the number of inactive electrodes that were inside resection; (iii) false negatives (FN), the number of active electrodes that were not inside resection; and (iv) true negatives (TN), the number of inactive electrodes that were not inside resection.

对于每个网络，我们将功率高于阈值的电极视为活跃，其余视为不活跃。距离切除区域 ≤10 毫米的电极被认为位于切除区域内。然后我们定义如下：(i) 真阳性 (TP)，位于切除区域内的活跃电极数量；(ii) 假阳性 (FP)，位于切除区域内的不活跃电极数量；(iii) 假阴性 (FN)，不在切除区域内的活跃电极数量；(iv) 真阴性 (TN)，不在切除区域内的不活跃电极数量。

We then calculated the following performance metrics per patient in each band: (i) sensitivity [TP/(TP + FN)]; (ii) specificity [TN/(TN + FP)]; (iii) precision or PPV [TP/(TP + FP)]; (iv) NPV [TN/(TN + FN)]; (v) accuracy [(TP + TN)/(TP + TN + FP + FN)]; and (vi) AUC. The terms precision and PPV are used interchangeably.

我们随后计算了每个频段内每位患者的以下性能指标：(i) 灵敏度 [TP/(TP + FN)]；(ii) 特异性 [TN/(TN + FP)]；(iii) 精确率或 PPV [TP/(TP + FP)]；(iv) NPV [TN/(TN + FN)]；(v) 准确率 [(TP + TN)/(TP + TN + FP + FN)]；以及 (vi) AUC。术语精确率和 PPV 可互换使用。

We also determined the optimal threshold in each band by averaging the thresholds determined by the maximum Youden index across patients. We used this threshold for optimal thresholding of networks. We also performed ROC AUC analysis using the SOZ as the target. We then compared the AUCs of the entire, interictal epileptogenic, and background networks in predicting the resection and SOZ in different bands (.

我们还通过平均跨患者最大 Youden 指数确定的阈值，找出了每个频段的最佳阈值。我们使用该阈值对网络进行最佳阈值处理。我们还使用 SOZ 作为目标进行了 ROC AUC 分析。随后，我们比较了整个网络、发作间期致痫网络和背景网络在不同频段预测切除区域和 SOZ 的 AUC 值。

Wilcoxon signed-rank

威尔科克森符号秩

test).

测试）。

We then estimated

我们随后估计了

net

网

，

res

结果

, and

，以及

res

结果

of the IENs of good- and poor-outcome patients in each band. We finally compared these properties between patients with different outcomes (

良好和不良转归患者的IEN在每个频段内。我们最终比较了不同转归患者之间的这些特性 (

Wilcoxon rank-sum

威尔科克森秩和检验

test). Using outcome as the actual class and each of the three IEN properties (

测试）。将结果作为实际类别，并使用三个IEN属性中的每一个（

net

网

，

res

结果

, and

，以及

res

结果

) as predictors, we performed ROC analysis to find for each band the Youden index and the corresponding threshold for each of the three properties. We then averaged the thresholds across all bands to estimate thresholds for: (i)

) 作为预测因子，我们进行了ROC分析，以寻找每个频段的Youden指数及对应的三个属性的阈值。然后，我们对所有频段的阈值进行平均，以估算以下阈值：(i)

${F}_{{net}}$

${F}_{{净}}$

(

$t{h}_{F}$

); (ii)

res

结果

(

$t{h}_{O})$

; and (iii)

；以及（iii）

res

结果

(

$t{h}_{D})$

that discriminate good- from poor-outcome patients. We then evaluated the variation of the three IEN properties (

区分良好和不良预后患者的差异。然后我们评估了三个IEN属性的变化 (

net

网

，

res

结果

，

res

结果

) at the patient level in the seven bands. For each patient, we also reported whether their IEN properties were above (

）在七个频段的患者水平上。对于每位患者，我们还报告了他们的IEN属性是否高于（

$t{h}_{F},t{h}_{O})$

or below (

或以下 (

$t{h}_{D}$

) the thresholds.

）阈值。

Robustness of IEN across variable IED rates and implantation types

不同IED速率和植入类型下IEN的稳健性

To validate the applicability of our framework to segments with sparse or absent IEDs, we investigated the ability of the IENs to predict resection and SOZ for both segments with frequent and sparse IEDs. Specifically, we selected for each patient two data segments of 1-min duration with interictal activity: (i) one with frequent IEDs (≥20 IEDs per minute), and (ii) one with sparse or absent IEDs (<20 IEDs per minute).

为了验证我们的框架对于IED稀疏或缺失片段的适用性，我们研究了IEN预测频繁和稀疏IED片段的切除区和发作起始区（SOZ）的能力。具体来说，我们为每位患者选择了两个持续1分钟且具有发作间期活动的数据片段：(i) 一个具有频繁IED（每分钟≥20个IED），以及 (ii) 一个具有稀疏或无IED（每分钟<20个IED）。

We estimated the frequency of IEDs for each data segment and tabulated the findings in Supplementary Table .

我们估计了每个数据段的IED频率，并将结果汇总在补充表中。

. Considering only data from good-outcome patients, we analyzed data segments with both frequent and sparse IEDs. Only 16 out of the 27 good-outcome patients had segments of 1-minute duration with both frequent and sparse IEDs. For these patients, we estimated the IENs and calculated the AUC of the IENs with the resection and SOZ.

仅考虑来自良好预后患者的数据，我们分析了频繁和稀疏IEDs的数据段。在27名良好预后患者中，只有16名患者有同时包含频繁和稀疏IEDs的1分钟时长数据段。对于这些患者，我们估算了IENs，并计算了IENs与切除区域和SOZ的AUC。

We then compared the AUC values derived from the segments with frequent IEDs with those derived from the segments with sparse IEDs..

我们随后比较了从频繁出现IED的片段和稀疏出现IED的片段得出的AUC值。

To examine whether our framework provides consistent findings among different implantation types, we studied the ability of the IENs to predict the resection and SOZ in patients with different implantation types in our cohort. For good-outcome patients, we initially estimated the IENs and then calculated their AUC with the resection and SOZ.

为了检查我们的框架是否在不同植入类型之间提供一致的发现，我们研究了IENs在我们队列中不同植入类型的患者中预测切除区域和SOZ的能力。对于预后良好的患者，我们首先估计了IENs，然后计算了它们与切除区域和SOZ的AUC。

We then compared the AUC of IENs with the resection and SOZ among the three implantation types..

我们随后比较了三种植入类型中IENs与切除和SOZ的AUC。

Predicting surgical outcome

预测手术结果

We developed automated outcome predictors with linear SVM

我们使用线性SVM开发了自动结果预测器

using the three IEN properties (

使用三个IEN属性 (

net

网

，

res

结果

, and

，以及

res

结果

) as features and outcome as target. We set the regularization parameter of the SVM to 1. An SVM was trained in each band separately (i.e., the properties of IENs in each band as predictors and outcome as target). We also trained another linear SVM (denoted by SVM-all) using 21 IEN properties (seven bands .

）作为特征，结果作为目标。我们将SVM的正则化参数设置为1。在每个频段分别训练了一个SVM（即，将每个频段的IEN属性作为预测变量，结果作为目标）。我们还使用21个IEN属性（七个频段）训练了另一个线性SVM（记为SVM-all）。

$\times$

three properties per band) simultaneously as features that incorporated information from all bands. Each SVM was validated using five-fold cross-validation by dividing patients into five random splits. We defined as: (i) TP, the number of patients predicted as good outcome who had a good outcome; (ii) FN, the number of patients predicted as poor outcome who had a good outcome; (iii) FP, the number of patients predicted as good outcome who had a poor outcome; and (iv) TN, the number of patients predicted as poor outcome who had a poor outcome.

每频带三个属性）同时作为包含所有频带信息的特征。每个SVM通过将患者分为五个随机分组进行五折交叉验证。我们定义为：(i) TP，预测为良好结果且实际为良好结果的患者数量；(ii) FN，预测为不良结果但实际为良好结果的患者数量；(iii) FP，预测为良好结果但实际为不良结果的患者数量；(iv) TN，预测为不良结果且实际为不良结果的患者数量。

We then calculated the following performance metrics: sensitivity, specificity, PPV, NPV, accuracy, AUC, and .

我们随后计算了以下性能指标：灵敏度、特异性、阳性预测值、阴性预测值、准确率、AUC 和。

Fisher’s exact

费舍尔精确检验

test

测试

values of each classifier. Similarly, we evaluated the ability of the entire network to predict outcome, and its performance was compared to that of the IEN.

每个分类器的值。同样，我们评估了整个网络预测结果的能力，并将其性能与IEN进行了比较。

Finally, ANOVA was performed to rank the IEN properties in the seven bands in decreasing order of importance to predict outcome. Feature importance is defined as:

最后，进行方差分析（ANOVA）以根据预测结果的重要性对七个频段中的IEN属性按降序排列。特征重要性定义为：

$$I=-\log (P)$$

(10)

where

其中

represents the

代表了

values of ANOVA. The analysis identified the IEN properties and bands that can best discriminate between good and poor outcome.

ANOVA值。该分析确定了可以最好地区分良好和不良结果的IEN特性和频段。

Statistical analysis

统计分析

The Kolmogorov-Smirnov test was used to test the normality of the power and network properties (

使用Kolmogorov-Smirnov检验来测试功率和网络属性的正态性 (

net

网

，

res

结果

, and

，以及

res

结果

). Cliff’s

). 克里夫的

measure was used to compute effect sizes. We applied two-sided non-parametric

该指标用于计算效应量。我们应用了双侧非参数

Wilcoxon signed-rank

威尔科克森符号秩检验

test for all paired comparisons (median values inside vs. outside the resection and SOZ, comparing the properties of the entire, interictal epileptogenic, and background networks). We applied two-sided

对所有配对比较进行检验（切除区域内外及癫痫发作区内外的中位数值，比较整个网络、发作间期致痫网络和背景网络的特性）。我们采用了双侧检验。

Wilcoxon rank-sum

威尔科克森秩和检验

test for non-paired comparisons between good- and poor-outcome patients. Bonferroni correction was applied in all multiple comparison tests

对良好和不良结局患者之间进行非配对比较的检验。所有多重比较检验均应用了Bonferroni校正。

. We used one-sided

我们使用了单侧

Fisher’s exact

费舍尔精确检验

test to evaluate the predictive value of the IENs to predict outcome.

评估IENs预测结果的预测价值的测试。

Chi-squared

卡方

test was used to compare different measures (gender, implantation side, epilepsy localization, and MRI findings) in terms of outcome. We assumed statistically significant results if

检验用于比较不同测量指标（性别、植入侧、癫痫定位和MRI结果）的结果。如果结果具有统计学意义，我们假设

$\le$

0.05 and reported measures as median and interquartile range. All analysis was performed using MATLAB 2022b (The MathWorks, Inc.).

0.05，并以中位数和四分位距报告测量值。所有分析均使用 MATLAB 2022b（The MathWorks, Inc.）进行。

Data availability

数据可用性

The data are available from the corresponding author upon request.

数据可应要求从通讯作者处获得。

Code availability

代码可用性

The code is available on GitHub (

代码可以在GitHub上找到 (

https://github.com/Hmayag11/AutoEZ

). A detailed pseudocode of the proposed framework is provided in Supplementary Note

）。拟议框架的详细伪代码见补充说明。

。

References

参考文献

Rosenow, F. & Lüders, H. Presurgical evaluation of epilepsy.

罗森诺，F. & 吕德斯，H. 癫痫的术前评估。

Brain

大脑

124

, 1683–1700 (2001).

，1683-1700（2001）。

Article

文章

PubMed

Google Scholar

谷歌学术

Jehi, L. The epileptogenic zone: concept and definition.

耶希，L. 癫痫发生区：概念与定义。

Epilepsy Curr.

癫痫杂志

, 12–16 (2018).

，12-16页（2018年）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Tamilia, E., Madsen, J. R., Grant, P. E., Pearl, P. L. & Papadelis, C. Current and emerging potential of magnetoencephalography in the detection and localization of high-frequency oscillations in epilepsy.

塔米利亚、E.、马德森、J. R.、格兰特、P. E.、珀尔、P. L.、帕帕德拉基斯、C. 《脑磁图在癫痫高频振荡检测与定位中的当前与潜在应用》。

Front Neurol.

前沿神经病学。

, 14 (2017).

，14（2017）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

MacDonald, D. B., Simon, M. V. & Nuwer, M. R. Chapter 6 - Neurophysiology during epilepsy surgery. in

麦克唐纳，D. B.，西蒙，M. V.，努尔，M. R. 第6章 - 癫痫手术期间的神经生理学。于

Intraoperative Neuromonitoring

术中神经监测

(eds. Nuwer, M. R. & MacDonald, D. B.) vol. 186 103–121 (Elsevier, 2022).

（编辑：努尔，M. R. 和麦克唐纳，D. B.）第186卷，103-121页（爱思唯尔，2022年）。

Seifer, G. et al. Noninvasive approach to focal cortical dysplasias: clinical, EEG, and neuroimaging features.

Seifer, G. 等。局灶性皮质发育不良的非侵入性方法：临床、脑电图和神经影像学特征。

Epilepsy Res Treat.

癫痫研究与治疗。

2012

2012年

, 736784 (2012).

，736784（2012）。

PubMed

PubMed Central

Google Scholar

谷歌学术

Jacobs, J. et al. High-frequency electroencephalographic oscillations correlate with outcome of epilepsy surgery.

雅各布斯，J. 等。高频脑电图振荡与癫痫手术结果相关。

Ann. Neurol.

神经病学年鉴

, 209–220 (2010).

，209-220页（2010年）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Prasad, A., Pacia, S. V., Vazquez, B., Doyle, W. K. & Devinsky, O. Extent of ictal origin in mesial temporal sclerosis patients monitored with subdural intracranial electrodes predicts outcome.

普拉萨德，A.，帕西亚，S. V.，巴斯克斯，B.，道尔，W. K.，德文斯基，O. 在使用皮层下颅内电极监测的内侧颞叶硬化症患者中，癫痫发作起源的范围可预测治疗结果。

J. Clin. Neurophysiol.

临床神经生理学杂志

, 243–248 (2003).

，243-248页（2003年）。

Article

文章

PubMed

Google Scholar

谷歌学术

Widjaja, E. et al. Cost-effectiveness of pediatric epilepsy surgery compared to medical treatment in children with intractable epilepsy.

Widjaja, E. 等。儿童难治性癫痫手术与药物治疗的成本效益比较。

Epilepsy Res.

癫痫研究。

, 61–68 (2011).

，61-68页（2011年）。

Article

文章

PubMed

Google Scholar

谷歌学术索

Papadelis, C. & Perry, M. S. Localizing the epileptogenic zone with novel biomarkers.

帕帕德尔利斯，C. & 佩里，M. S. 使用新型生物标志物定位癫痫发作区。

Semin. Pediatr. Neurol.

小儿神经病学研讨会

, 100919 (2021).

，100919（2021）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Lai, N., Li, Z., Xu, C., Wang, Y. & Chen, Z. Diverse nature of interictal oscillations: EEG-based biomarkers in epilepsy.

赖, N., 李, Z., 徐, C., 王, Y. & 陈, Z. 癫痫中发作间期振荡的多样性：基于脑电图的生物标志物。

Neurobiol. Dis.

神经生物疾病。

177

, 105999 (2023).

，105999（2023）。

Article

文章

PubMed

Google Scholar

谷歌学术搜索

Emmady, P. D. & Anilkumar, A. C. EEG Abnormal Waveforms. in

Emmady, P. D. & Anilkumar, A. C. 脑电图异常波形。于

StatPearls

(StatPearls Publishing, Treasure Island (FL), 2023).

（StatPearls Publishing，佛罗里达州金银岛，2023年）。

Engel, J. Jr. et al. Epilepsy biomarkers.

恩格尔，J. Jr. 等。癫痫生物标志物。

Epilepsia

癫痫

, 61–69 (2013).

，61-69页（2013年）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Klotz, K. A., Sag, Y., Schönberger, J. & Jacobs, J. Scalp ripples can predict development of epilepsy after first unprovoked seizure in childhood.

克洛茨，K. A.，萨格，Y.，舍恩贝格尔，J.，雅各布斯，J. 头皮涟漪可以预测儿童首次无诱因癫痫发作后癫痫的发展。

Ann. Neurol.

神经学年鉴

, 134–142 (2021).

，134-142页（2021年）。

Article

文章

PubMed

Google Scholar

谷歌学术

Scheuer, M. L. et al. Seizure detection: interreader agreement and detection algorithm assessments using a large dataset.

舍尔，M. L. 等。癫痫发作检测：使用大型数据集的阅片者间一致性及检测算法评估。

J. Clin. Neurophysiol.

临床神经生理学杂志

, 439–447 (2021).

，第439-447页（2021年）。

Article

文章

PubMed

Google Scholar

谷歌学术

Sylolypavan, A., Sleeman, D., Wu, H. & Sim, M. The impact of inconsistent human annotations on AI driven clinical decision making.

Sylolypavan, A., Sleeman, D., Wu, H. & Sim, M. 不一致的人类注释对人工智能驱动的临床决策的影响。

npj Digital Med.

npj 数字医学

, 26 (2023).

，26（2023）。

Article

文章

Google Scholar

谷歌学术搜索

Corona, L. et al. Non-invasive mapping of epileptogenic networks predicts surgical outcome.

科罗纳，L. 等。非侵入性癫痫网络绘图预测手术结果。

Brain

大脑

146

, 1916–1931 (2023).

，1916-1931（2023）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Shah, P. et al. High interictal connectivity within the resection zone is associated with favorable post-surgical outcomes in focal epilepsy patients.

Shah, P. 等。在局灶性癫痫患者中，切除区域内高发作间期连接性与良好的术后结果相关。

Neuroimage Clin.

神经影像临床。

, 101908 (2019).

，101908（2019）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术索

Brunton, B. W., Johnson, L. A., Ojemann, J. G. & Kutz, J. N. Extracting spatial–temporal coherent patterns in large-scale neural recordings using dynamic mode decomposition.

布鲁顿，B. W.，约翰逊，L. A.，奥杰曼，J. G.，库茨，J. N. 使用动态模式分解提取大规模神经记录中的时空相干模式。

J. Neurosci. Methods

神经科学方法杂志

258

, 1–15 (2016).

，1-15页（2016年）。

Article

文章

PubMed

Google Scholar

谷歌学术

Bourien, J. et al. A method to identify reproducible subsets of co-activated structures during interictal spikes. Application to intracerebral EEG in temporal lobe epilepsy.

布尔恩，J. 等。一种在发作间期尖峰期间识别可重复的共激活结构子集的方法。应用于颞叶癫痫的颅内脑电图。

Clin. Neurophysiol.

临床神经生理学

116

, 443–455 (2005).

，443-455页（2005年）。

Article

文章

PubMed

Google Scholar

谷歌学术

Tamilia, E. et al. Surgical resection of ripple onset predicts outcome in pediatric epilepsy.

塔米利亚（Tamilia）等人。ripple发作的外科切除预测小儿癫痫的预后。

Ann. Neurol.

神经学年鉴

, 331–346 (2018).

，331-346页（2018年）。

Article

文章

PubMed

Google Scholar

谷歌学术搜索

Matarrese, M. A. G. et al. Spike propagation mapping reveals effective connectivity and predicts surgical outcome in epilepsy.

马塔雷塞，M. A. G. 等。尖峰传播映射揭示有效连接性并预测癫痫手术结果。

Brain

大脑

146

, 3898–3912 (2023).

，3898–3912（2023）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Jahromi, S. et al. Overlap of spike and ripple propagation onset predicts surgical outcome in epilepsy.

Jahromi, S. 等。癫痫中尖峰和涟漪传播起始的重叠预测手术结果。

Annals of Clin. Transl. Neurol.

临床与转化神经病学年鉴

, 2530–2547 (2024).

，2530–2547（2024）。

Article

文章

Google Scholar

谷歌学术搜索

Diamond, J. M., Chapeton, J. I., Theodore, W. H., Inati, S. K. & Zaghloul, K. A. The seizure onset zone drives state-dependent epileptiform activity in susceptible brain regions.

Diamond, J. M., Chapeton, J. I., Theodore, W. H., Inati, S. K. & Zaghloul, K. A. 发作起始区驱动易感脑区中与状态相关的癫痫样活动。

Clin. Neurophysiol.

临床神经生理学

130

, 1628–1641 (2019).

，1628-1641（2019）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Friston, K. J. Functional and effective connectivity: a review.

Friston, K. J. 功能性与有效连接性：综述。

Brain Connect

大脑连接

, 13–36 (2011).

，13-36页（2011年）。

Article

文章

PubMed

Google Scholar

谷歌学术索

Stam, C. J. et al. The trees and the forest: Characterization of complex brain networks with minimum spanning trees.

Stam, C. J. 等。树木与森林：用最小生成树表征复杂脑网络。

Int. J. Psychophysiol.

国际心理生理学杂志

, 129–138 (2014).

，129-138页（2014年）。

Article

文章

PubMed

Google Scholar

谷歌学术

Bartolomei, F. et al. Defining epileptogenic networks: Contribution of SEEG and signal analysis.

巴托洛梅伊，F. 等。定义癫痫网络：SEEG和信号分析的贡献。

Epilepsia

癫痫

, 1131–1147 (2017).

，1131–1147（2017）。

Article

文章

PubMed

Google Scholar

谷歌学术

Wang, H. E. et al. Delineating epileptogenic networks using brain imaging data and personalized modeling in drug-resistant epilepsy.

王，H. E. 等。利用脑成像数据和个性化建模描绘耐药性癫痫的致痫网络。

Sci. Transl. Med.

科学转化医学

, eabp8982 (2023).

，eabp8982（2023）。

Article

文章

PubMed

Google Scholar

谷歌学术搜索

Lagarde, S. et al. Interictal stereotactic-EEG functional connectivity in refractory focal epilepsies.

拉加德，S. 等。难治性局灶性癫痫的发作间期立体定向脑电图功能连接。

Brain

大脑

141

, 2966–2980 (2018).

，2966-2980（2018）。

Article

文章

PubMed

Google Scholar

谷歌学术

Basiri, R., Shariatzadeh, A., Wiebe, S. & Aghakhani, Y. Focal epilepsy without interictal spikes on scalp EEG: A common finding of uncertain significance.

Basiri, R., Shariatzadeh, A., Wiebe, S. & Aghakhani, Y. 无发作间期头皮脑电图棘波的局灶性癫痫：一个意义不明的常见发现。

Epilepsy Res.

癫痫研究。

150

, 1–6 (2019).

，1-6页（2019年）。

Article

文章

PubMed

Google Scholar

谷歌学术

Alamoudi, O. A., Ilyas, A., Pati, S. & Iasemidis, L. Interictal localization of the epileptogenic zone: Utilizing the observed resonance behavior in the spectral band of surrounding inhibition.

阿拉莫迪，O. A.，伊利亚斯，A.，帕蒂，S. & 雅西米迪斯，L. 癫痫发作间期癫痫源区的定位：利用在周围抑制频谱带中观察到的共振行为。

Front. Neurosci.

神经科学前沿

, (2022).

，（2022）。

Jaakkola, H., Henno, J., Mäkelä, J. & Thalheim, B. Artificial Intelligence Yesterday, Today and Tomorrow. in

Jaakkola, H., Henno, J., Mäkelä, J. & Thalheim, B. 人工智能的昨天、今天和明天。收录于

2019 42nd International Convention on Information and Communication Technology, Electronics and Microelectronics (MIPRO)

2019年第42届国际信息通信技术、电子与微电子会议（MIPRO）

860–867

https://doi.org/10.23919/MIPRO.2019.8756913

(2019).

（2019）。

Xu, Y. et al. Artificial intelligence: A powerful paradigm for scientific research.

徐，Y. 等。人工智能：科学研究的强大范式。

Innovation

创新

, 100179 (2021).

，100179（2021）。

PubMed

PubMed Central

Google Scholar

谷歌学术

Wang, Y. et al. Automatic localization of seizure onset zone based on multi-epileptogenic biomarkers analysis of single-contact from interictal SEEG.

王, Y. 等。基于单接触点间歇性SEEG多癫痫生物标志物分析的癫痫发作起始区自动定位。

Bioengineering (Basel)

生物工程（巴塞尔）

, 769 (2022).

，769页（2022年）。

Article

文章

PubMed

Google Scholar

谷歌学术

Lundstrom, B. N., Brinkmann, B. H. & Worrell, G. A. Low frequency novel interictal EEG biomarker for localizing seizures and predicting outcomes.

Lundstrom, B. N., Brinkmann, B. H. & Worrell, G. A. 低频新型发作间期脑电生物标志物用于癫痫灶定位及预后预测。

Brain Commun.

脑通讯。

, fcab231 (2021).

，fcab231（2021）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Burns, S. P. et al. Network dynamics of the brain and influence of the epileptic seizure onset zone.

Burns, S. P. 等。大脑的网络动态及癫痫发作起始区的影响。

Proc. Natl Acad. Sci. USA

美国国家科学院院刊

111

, E5321–5330 (2014).

，E5321–5330（2014）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术搜索

Baud, M. O. et al. Unsupervised learning of spatiotemporal interictal discharges in focal epilepsy.

Baud, M. O. 等。局灶性癫痫中时空发作间期放电的无监督学习。

Neurosurgery

神经外科手术

, 683–691 (2018).

，683-691页（2018年）。

Article

文章

PubMed

Google Scholar

谷歌学术

Liu, S. et al. Stereotyped high-frequency oscillations discriminate seizure onset zones and critical functional cortex in focal epilepsy.

刘，S. 等。刻板高频振荡区分局灶性癫痫的发作起始区和关键功能皮层。

Brain

大脑

141

, 713–730 (2018).

，713-730页（2018年）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Schönberger, J. et al. Distinction of Physiologic and Epileptic Ripples: An Electrical Stimulation Study.

施恩伯格，J. 等。生理性和癫痫性涟漪的区别：一项电刺激研究。

Brain Sci.

脑科学。

, 538 (2021).

，538页（2021年）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Lee, D. & Seung, H. S. Algorithms for Non-negative Matrix Factorization. in

李, D. & Seung, H. S. 非负矩阵分解算法。于

Advances in Neural Information Processing Systems

神经信息处理系统进展

(eds. Leen, T., Dietterich, T. & Tresp, V.) vol. 13 (MIT Press, 2000).

（编者：Leen，T.，Dietterich，T.，& Tresp，V.）第13卷（麻省理工学院出版社，2000年）。

Dice, L. R. Measures of the amount of ecologic association between species.

Dice, L. R. 物种间生态关联程度的测量。

Ecology

生态学

, 297–302 (1945).

，297-302页（1945年）。

Article

文章

Google Scholar

谷歌学术

Pajula, J., Kauppi, J.-P. & Tohka, J. Inter-subject correlation in fMRI: method validation against stimulus-model based analysis.

帕朱拉，J.，考皮，J.-P. & 托赫卡，J. 功能磁共振成像中的主体间相关性：针对基于刺激模型分析的方法验证。

PLOS ONE

, 1–13 (2012).

，1–13（2012）。

Google Scholar

谷歌学术

Cortes, C. & Vapnik, V. Support-vector networks.

Cortes, C. & Vapnik, V. 支持向量网络。

Mach. Learn.

机器学习

, 273–297 (1995).

，273-297页（1995年）。

Article

文章

Google Scholar

谷歌学术

Tripathy, G. & Sharaff, A. AEGA: enhanced feature selection based on ANOVA and extended genetic algorithm for online customer review analysis.

Tripathy, G. & Sharaff, A. AEGA：基于方差分析和扩展遗传算法的增强特征选择用于在线客户评论分析。

J. Supercomput.

超级计算杂志

, 13180–13209 (2023).

，13180-13209（2023）。

Article

文章

Google Scholar

谷歌学术索

Varatharajah, Y. et al. Inter-ictal Seizure Onset Zone localization using unsupervised clustering and Bayesian Filtering. in

Varatharajah, Y. 等。使用无监督聚类和贝叶斯滤波的发作间癫痫发作起始区定位。在

2017 8th International IEEE/EMBS Conference on Neural Engineering (NER)

2017年第八届国际IEEE/EMBS神经工程会议（NER）

533–539

https://doi.org/10.1109/NER.2017.8008407

(2017).

（2017）。

Gunnarsdottir, K. M. et al. Source-sink connectivity: a novel interictal EEG marker for seizure localization.

Gunnarsdottir, K. M. 等。源-汇连通性：一种新的发作间期脑电图标记用于癫痫定位。

Brain

大脑

145

, 3901–3915 (2022).

，3901–3915（2022）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Lagarde, S., Bénar, C.-G., Wendling, F. & Bartolomei, F. Interictal functional connectivity in focal refractory epilepsies investigated by intracranial EEG.

拉加德，S.，贝纳尔，C.-G.，温德林，F.，巴托洛梅，F. 通过颅内脑电图研究局灶性难治性癫痫的发作间期功能连接性。

Brain Conn.

大脑连接。

, 850–869 (2022).

，850-869（2022）。

Article

文章

Google Scholar

谷歌学术索

Varatharajah, Y. et al. Integrating artificial intelligence with real-time intracranial EEG monitoring to automate interictal identification of seizure onset zones in focal epilepsy.

Varatharajah, Y. 等。将人工智能与实时颅内脑电图监测相结合，以自动识别局灶性癫痫发作区的发作间期。

J. Neural Eng.

神经工程杂志

, 046035 (2018).

，046035（2018）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Md. Islam, R., Zhao, X., Miao, Y., Sugano, H. & Tanaka, T. Epileptic seizure focus detection from interictal electroencephalogram: a survey.

Md. Islam, R., 赵欣，苗宇，菅野弘司，田中拓。从发作间期脑电图检测癫痫发作焦点：一项综述。

Cogn. Neurodyn.

认知神经动力学

, 1–23 (2023).

，1-23页（2023）。

Article

文章

PubMed

Google Scholar

谷歌学术

Ying, X. An overview of overfitting and its solutions.

应，X. 过拟合及其解决方案的概述。

J. Phys.: Conf. Ser.

物理学杂志：会议系列

1168

, 022022 (2019).

，022022（2019）。

Google Scholar

谷歌学术

Rijal, S. et al. Functional connectivity discriminates epileptogenic states and predicts surgical outcome in children with drug resistant epilepsy.

里贾尔，S. 等。功能连接性区分癫痫发作状态并预测耐药性癫痫儿童的手术结果。

Sci. Rep.

科学报告

, 9622 (2023).

，9622（2023）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Myers, J. C. et al. The spatial reach of neuronal coherence and spike-field coupling across the human neocortex.

迈尔斯，J. C. 等。神经元相干性和尖峰场耦合在人类新皮层中的空间范围。

J. Neurosci.

神经科学杂志

, 6285–6294 (2022).

，6285–6294（2022）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Kane, N. et al. A revised glossary of terms most commonly used by clinical electroencephalographers and updated proposal for the report format of the EEG findings. Revision 2017.

凯恩，N. 等。临床脑电图学家最常用术语的修订词汇表及脑电图结果报告格式的更新建议。2017年修订版。

Clin. Neurophysiol. Pract.

临床神经生理学实践

, 170–185 (2017).

，170-185页（2017年）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Douw, L. et al. Epilepsy is related to theta band brain connectivity and network topology in brain tumor patients.

Douw, L. 等。癫痫与脑肿瘤患者的θ波段脑连接性和网络拓扑结构有关。

BMC Neurosci.

BMC神经科学。

, 103 (2010).

，103页（2010年）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Horstmann, M.-T. et al. State dependent properties of epileptic brain networks: comparative graph-theoretical analyses of simultaneously recorded EEG and MEG.

Horstmann, M.-T. 等。癫痫脑网络的状态依赖性特性：同时记录的EEG和MEG的比较图论分析。

Clin. Neurophysiol.

临床神经生理学

121

, 172–185 (2010).

，172-185页（2010年）。

Article

文章

PubMed

Google Scholar

谷歌学术

Douw, L. et al. Functional connectivity’ is a sensitive predictor of epilepsy diagnosis after the first seizure.

Douw, L. 等。功能连接性是癫痫诊断的敏感预测因子，尤其是在首次癫痫发作后。

PLoS One

PLOS ONE

, e10839 (2010).

，e10839（2010）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Morgan, R. J. & Soltesz, I. Nonrandom connectivity of the epileptic dentate gyrus predicts a major role for neuronal hubs in seizures.

摩根，R. J. & 索特茨，I. 癫痫性齿状回的非随机连接预测神经元枢纽在癫痫发作中起主要作用。

Proc. Natl Acad. Sci. USA

美国国家科学院院刊

105

, 6179–6184 (2008).

，6179-6184页（2008年）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Alowais, S. A. et al. Revolutionizing healthcare: the role of artificial intelligence in clinical practice.

阿尔瓦伊斯，S. A. 等。革新医疗保健：人工智能在临床实践中的作用。

BMC Med Educ.

BMC医学教育。

, 689 (2023).

，689页（2023年）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Tveit, J. et al. Automated interpretation of clinical electroencephalograms using artificial intelligence.

Tveit, J. 等。使用人工智能自动解读临床脑电图。

JAMA Neurol.

神经病学JAMA。

, 805–812 (2023).

，805-812页（2023年）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Abdi-Sargezeh, B. et al. A review of signal processing and machine learning techniques for interictal epileptiform discharge detection.

阿卜迪-萨尔热泽，B. 等。用于发作间期癫痫样放电检测的信号处理与机器学习技术综述。

Computers Biol. Med.

计算机生物学与医学

168

, 107782 (2024).

，107782（2024）。

Article

文章

Google Scholar

谷歌学术

Bernabei, J. M. et al. Electrocorticography and stereo EEG provide distinct measures of brain connectivity: implications for network models.

贝尔纳贝，J. M. 等。皮层电图和立体脑电图提供不同的大脑连接性测量：对网络模型的意义。

Brain Commun.

脑共通。

, fcab156 (2021).

，fcab156（2021）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术索

Lesser, R. P., Crone, N. E. & Webber, W. R. S. Subdural electrodes.

Lesser, R. P., Crone, N. E. & Webber, W. R. S. 硬膜下电极。

Clin. Neurophysiol.

临床神经生理学

121

, 1376–1392 (2010).

，1376-1392（2010）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Halgren, M. et al. The generation and propagation of the human alpha rhythm.

Halgren, M. 等。人类α节律的产生与传播。

Proc. Natl Acad. Sci. USA

美国国家科学院院刊

116

, 23772–23782 (2019).

，23772-23782（2019）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Alhilani, M. et al. Ictal and interictal source imaging on intracranial EEG predicts epilepsy surgery outcome in children with focal cortical dysplasia.

Alhilani, M. 等。颅内脑电图的发作期和发作间期源成像可预测局灶性皮质发育不良儿童的癫痫手术结果。

Clin. Neurophysiol.

临床神经生理学

131

, 734–743 (2020).

，734-743页（2020年）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Dimakopoulos, V. et al. Protocol for multicentre comparison of interictal high-frequency oscillations as a predictor of seizure freedom.

Dimakopoulos, V. 等。多中心比较发作间期高频振荡作为无癫痫发作预测因子的协议。

Brain Commun.

脑共通。

, fcac151 (2022).

，fcac151（2022）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Frauscher, B. et al. Facilitation of epileptic activity during sleep is mediated by high amplitude slow waves.

弗劳施尔，B. 等。癫痫活动在睡眠期间的促进是由高振幅慢波介导的。

Brain

大脑

138

, 1629–1641 (2015).

，1629-1641（2015）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Sammaritano, M., Gigli, G. L. & Gotman, J. Interictal spiking during wakefulness and sleep and the localization of foci in temporal lobe epilepsy.

萨马拉塔诺，M.，吉格利，G.L.，戈特曼，J. 颞叶癫痫在清醒和睡眠期间的发作间期放电及病灶定位。

Neurology

神经学

, 290–297 (1991).

，290-297页（1991年）。

Article

文章

PubMed

Google Scholar

谷歌学术

Tadel, F., Baillet, S., Mosher, J. C., Pantazis, D. & Leahy, R. M. Brainstorm: A User-Friendly Application for MEG/EEG Analysis.

Tadel, F., Baillet, S., Mosher, J. C., Pantazis, D. & Leahy, R. M. Brainstorm：一款用于MEG/EEG分析的用户友好型应用程序。

Comput. Intell. Neurosci.

计算机智能与神经科学

2011

2011年

, 879716 (2011).

，879716（2011）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术索

Roberts, D. W., Hartov, A., Kennedy, F. E., Miga, M. I. & Paulsen, K. D. Intraoperative brain shift and deformation: a quantitative analysis of cortical displacement in 28 cases.

Roberts, D. W., Hartov, A., Kennedy, F. E., Miga, M. I. & Paulsen, K. D. 手术中脑移位和变形：28例皮质移位的定量分析。

Neurosurgery

神经外科手术

四十三

, 749–758 (1998).

，749-758页（1998年）。

Article

文章

PubMed

Google Scholar

谷歌学术

SCHMID, P. J. Dynamic mode decomposition of numerical and experimental data.

施密德，P. J. 数值与实验数据的动态模态分解。

J. Fluid Mech.

流体力学杂志

656

, 5–28 (2010).

，5-28（2010）。

Article

文章

Google Scholar

谷歌学术索

Wu, Z., Brunton, S. L. & Revzen, S. Challenges in dynamic mode decomposition.

吴，Z.，Brunton，S. L.，Revzen，S. 动态模式分解中的挑战。

J. R. Soc. Interface

皇家学会界面期刊

, 20210686 (2021).

，20210686（2021）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Ikeda, S., Kawano, K., Watanabe, S., Yamashita, O. & Kawahara, Y. Predicting behavior through dynamic modes in resting-state fMRI data.

池田，S.，河野，K.，渡边，S.，山下，O.，川原，Y. 通过静息态fMRI数据中的动态模式预测行为。

NeuroImage

神经影像

247

, 118801 (2022).

，118801（2022）。

Article

文章

PubMed

Google Scholar

谷歌学术

Raak, F., Susuki, Y., Mezić, I. & Hikihara, T. On Koopman and dynamic mode decompositions for application to dynamic data with low spatial dimension. in

Raak, F., Susuki, Y., Mezić, I. & Hikihara, T. 关于Koopman和动态模式分解在低空间维度动态数据中的应用。

2016 IEEE 55th Conference on Decision and Control (CDC)

2016年IEEE第55届决策与控制会议（CDC）

6485–6491

https://doi.org/10.1109/CDC.2016.7799267

(2016).

（2016）。

Kutz, J. N., Brunton, S. L., Brunton, B. W. & Proctor, J. L.

库茨，J. N.，布鲁顿，S. L.，布鲁顿，B. W.，普罗克特，J. L.

Dynamic Mode Decomposition: Data-Driven Modeling of Complex Systems

动态模式分解：复杂系统的数据驱动建模

. (SIAM-Society for Industrial and Applied Mathematics, Philadelphia, PA, USA, 2016).

。（工业与应用数学学会，美国宾夕法尼亚州费城，2016年）

Staley, K. J. & Dudek, F. E. Interictal spikes and epileptogenesis.

Staley, K. J. & Dudek, F. E. 发作间期尖波与癫痫发生。

Epilepsy Curr.

癫痫电流。

, 199–202 (2006).

，199-202页（2006年）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Henin, S. et al. Spatiotemporal dynamics between interictal epileptiform discharges and ripples during associative memory processing.

Henin, S. 等。联想记忆处理过程中发作间期癫痫样放电与涟漪之间的时空动态关系。

Brain

大脑

144

, 1590–1602 (2021).

，1590-1602（2021）。

Article

文章

PubMed

PubMed Central

Google Scholar

谷歌学术

Kim, H. & Park, H. Sparse non-negative matrix factorizations via alternating non-negativity-constrained least squares for microarray data analysis.

Kim, H. & Park, H. 通过交替非负约束最小二乘的稀疏非负矩阵分解在微阵列数据分析中的应用。

Bioinformatics

生物信息学

, 1495–1502 (2007).

，1495-1502（2007）。

Article

文章

PubMed

Google Scholar

谷歌学术索

Han, J., Kamber, M. & Pei, J.

韩家炜、坎伯、裴健

Data Mining: Concepts and Techniques

数据挖掘：概念与技术

. (Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 2011).

（摩根考夫曼出版社，美国加利福尼亚州旧金山，2011年）。

Kubik, S. & Chad D., A.

库比克，S. & 查德，D.

Atlas of the Cerebral Sulci

大脑沟回图集

. (Thieme Medical Publishers, 1990).

。（Thieme Medical Publishers，1990年）。

Kaplan, D & Glass, L. Understanding Nonlinear Dynamics. (Springer, 1997).

Kaplan, D & Glass, L. 理解非线性动力学. (Springer, 1997).

Haynes, W. Bonferroni Correction. in

海恩斯，W. 邦费罗尼校正。在

Encyclopedia of Systems Biology

系统生物学百科全书

(eds. Dubitzky, W., Wolkenhauer, O., Cho, K.-H. & Yokota, H.) 154–154 (Springer New York, New York, NY, 2013).

（编辑：Dubitzky, W., Wolkenhauer, O., Cho, K.-H. & Yokota, H.）154–154页（Springer New York，纽约，NY，2013年）。

https://doi.org/10.1007/978-1-4419-9863-7_1213

。

Download references

下载参考文献

Acknowledgements

致谢

This work was supported by the National Institute of Neurological Disorders and Stroke (R01NS104116; R01NS134944; PI: Christos Papadelis).

这项工作得到了国家神经疾病和中风研究所的支持（R01NS104116；R01NS134944；首席研究员：Christos Papadelis）。

Author information

作者信息

Authors and Affiliations

作者与所属机构

Neuroscience Research Center, Jane and John Justin Institute for Mind Health, Cook Children’s Health Care System, Fort Worth, TX, USA

神经科学研究中心，Jane and John Justin心理健康研究所，Cook儿童医疗系统，德克萨斯州沃斯堡，美国

Hmayag Partamian, Saeed Jahromi, Ludovica Corona, M. Scott Perry & Christos Papadelis

Hmayag Partamian, Saeed Jahromi, Ludovica Corona, M. Scott Perry 和 Christos Papadelis

Department of Bioengineering, The University of Texas at Arlington, Arlington, TX, USA

美国德克萨斯大学阿灵顿分校生物工程系，阿灵顿，德克萨斯州，美国

Hmayag Partamian, Saeed Jahromi, Ludovica Corona & Christos Papadelis

阿米亚格·帕塔米安，赛义德·贾霍罗米，卢多维卡·科罗纳和克里斯托斯·帕帕迪利斯

Fetal-Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital, Harvard Medical School, Boston, MA, USA

胎儿-新生儿神经影像与发育科学中心，波士顿儿童医院，哈佛医学院，波士顿，马萨诸塞州，美国

Eleonora Tamilia

埃莱奥诺拉·塔米利亚

Division of Epilepsy and Clinical Neurophysiology, Department of Neurology, Boston Children’s Hospital, Harvard Medical School, Boston, MA, USA

癫痫与临床神经生理学部，神经内科，波士顿儿童医院，哈佛医学院，波士顿，马萨诸塞州，美国

Eleonora Tamilia, Jeffrey Bolton & Phillip L. Pearl

埃莱奥诺拉·塔米利亚、杰弗里·博尔顿和菲利普·L·珀尔

Department of Neurosurgery, Boston Children’s Hospital, Harvard Medical School, Boston, MA, USA

美国马萨诸塞州波士顿，哈佛医学院，波士顿儿童医院神经外科 department

Joseph R. Madsen & Scellig S. D. Stone

约瑟夫·R·马德森和斯凯利格·S·D·斯通

Burnett School of Medicine, Texas Christian University, Fort Worth, TX, USA

伯内特医学院，德克萨斯基督教大学，沃斯堡，德克萨斯州，美国

Christos Papadelis

克里斯托斯·帕帕德尔尼斯

Authors

作者

Hmayag Partamian

哈马亚格·帕塔米安

View author publications

查看作者的出版物

You can also search for this author in

您还可以搜索此作者在

PubMed

Google Scholar

谷歌学术索

Saeed Jahromi

赛义德·贾霍罗米

View author publications

查看作者的出版物

You can also search for this author in

您还可以搜索此作者在

PubMed

Google Scholar

谷歌学术索

Ludovica Corona

露多维卡·科罗纳

View author publications

查看作者出版物

You can also search for this author in

您还可以搜索此作者在

PubMed

Google Scholar

谷歌学术

M. Scott Perry

M. 斯科特·佩里

View author publications

查看作者的出版物

You can also search for this author in

您还可以搜索此作者在

PubMed

Google Scholar

谷歌学术

Eleonora Tamilia

埃莱奥诺拉·塔米利亚

View author publications

查看作者出版物

You can also search for this author in

您还可以搜索此作者在

PubMed

Google Scholar

谷歌学术索

Joseph R. Madsen

约瑟夫·R·马德森

View author publications

查看作者出版物

You can also search for this author in

您还可以搜索该作者在

PubMed

Google Scholar

谷歌学术

Jeffrey Bolton

杰弗里·博尔顿

View author publications

查看作者的出版物

You can also search for this author in

您还可以搜索该作者在

PubMed

Google Scholar

谷歌学术

Scellig S. D. Stone

斯凯利格·S·D·斯通

View author publications

查看作者的出版物

You can also search for this author in

您还可以搜索此作者在

PubMed

Google Scholar

谷歌学术搜索

Phillip L. Pearl

菲利普·L·珀尔

View author publications

查看作者的出版物

You can also search for this author in

您还可以搜索此作者在

PubMed

Google Scholar

谷歌学术

Christos Papadelis

克里斯托斯·帕帕德利斯

View author publications

查看作者的出版物

You can also search for this author in

您还可以搜索此作者在

PubMed

Google Scholar

谷歌学术

Contributions

贡献

H.P. and C.P. conceptualized the project and experimental design. H.P. developed the mathematical formulation and the codes of the proposed framework. H.P., S.J., L.C., S.P., E.T., J.M., J.B., S.S., P.P., and C.P., contributed to the acquisition and analysis of data. H.P., S.J., and L.C. participated in the developing various mathematical formulations, statistical analysis, and representation of results of the study.

H.P. 和 C.P. 构思了该项目和实验设计。H.P. 开发了所提出框架的数学公式和代码。H.P.、S.J.、L.C.、S.P.、E.T.、J.M.、J.B.、S.S.、P.P. 和 C.P. 参与了数据的获取和分析。H.P.、S.J. 和 L.C. 参与了研究中各种数学公式的开发、统计分析和结果表示。

H.P., S.J., L.C., and C.P. contributed to drafting the text and preparing figures. All authors have read and approved the manuscript..

H.P.、S.J.、L.C. 和 C.P. 参与了文本的起草和图表的准备。所有作者都已阅读并批准了手稿。

Corresponding author

通讯作者

Correspondence to

致信给

Christos Papadelis

克里斯托斯·帕帕德利斯

。

Ethics declarations

伦理声明

Competing interests

竞争利益

The authors have no competing interests to disclose.

作者没有需要披露的竞争利益。

Additional information

附加信息

Publisher’s note

出版商说明

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Springer Nature 对已发布地图中的管辖权声明和机构隶属关系保持中立。

Supplementary information

补充信息

Supplementary Information

补充信息

Rights and permissions

权利与许可

Open Access

开放获取

This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material.

本文根据知识共享署名-非商业性使用-禁止演绎 4.0 国际许可协议获得许可，该协议允许您在任何媒介或格式中进行非商业性的使用、分享、分发和复制，只要您对原作者和来源给予适当的署名，提供指向知识共享许可协议的链接，并说明是否对授权材料进行了修改。

You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.

根据本许可，您无权分享从本文或其部分内容衍生的改编材料。本文中的图像或其他第三方材料包含在文章的Creative Commons许可中，除非材料的署名行另有说明。如果材料未包含在文章的Creative Commons许可中，并且您的预期用途不被法律法规允许或超出了允许的使用范围，您需要直接从版权持有人处获得许可。

To view a copy of this licence, visit .

要查看此许可证的副本，请访问。

http://creativecommons.org/licenses/by-nc-nd/4.0/

。

Reprints and permissions

重印和许可

About this article

关于本文

Cite this article

引用本文

Partamian, H., Jahromi, S., Corona, L.

帕塔米安，H.，贾霍米，S.，科罗纳，L.