Predicting surgical outcome in drug-resistant epilepsy by combining interictal biomarkers within a machine learning framework

Abstract

Delineating the epileptogenic zone (EZ) is essential for achieving seizure freedom in drug-resistant epilepsy (DRE). Conventionally, seizure onset derived from ictal intracranial EEG (iEEG) approximates the EZ, but acquiring ictal data can be challenging. Interictal iEEG abnormalities offer abundant, non-seizure-dependent markers of the epileptogenic tissue; however, these biomarkers offer limited specificity. Here, we propose a machine learning framework that integrates interictal spike and ripple features to predict the epileptogenic contacts targeted for surgical removal and the patient’s surgical outcome. We retrospectively analyzed iEEG data from 62 children with DRE [34 with good outcome (Engel I)], automatically detected spikes and ripples, and computed temporal, spectral, and spatial features for each channel. Using combinations of these features and the resected contacts as targets, we trained Random Forest classifiers using only good outcome patients to estimate epileptogenic contacts. Spike-based and combined spike and ripple features outperformed individual biomarkers in predicting the epileptogenic contacts with an area under the receiver operating characteristic curve of 0.89 and 74% spatial overlap with resection. Although most individual features and classifiers predicted outcome, the combined feature model performed best (i.e., sensitivity 88%, specificity 68%, and accuracy 79%). Our findings demonstrate that integrating multimodal interictal features improves the identification of epileptogenic contacts providing valuable prognostic insights for epilepsy surgery.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-45177-x.

Introduction

Precise delineation of the epileptogenic zone (EZ) remains one of the most significant challenges for surgery planning and achieving seizure freedom post-surgery in patients with drug-resistant epilepsy (DRE)^1,2. This hypothetical zone is often approximated by the seizure onset zone (SOZ), which is identified by manual review of ictal activity captured by intracranial EEG (iEEG) to detect the brain region(s) that initiate clinical seizures¹. Yet, increasing evidence suggests that interictal electrophysiological abnormalities observed in iEEG recordings (i.e., brief pathological events occurring between seizures) may also approximate the EZ^3,4. Importantly, identifying such biomarkers from readily available interictal data offers a significant advantage by eliminating the need to wait for a seizure to occur^5,6.

Among the proposed interictal epilepsy biomarkers, spikes and high-frequency oscillations (HFOs) are the most extensively studied^3,4,7. Interictal spikes, brief high-amplitude deflections typically occurring in the 1–70 Hz frequency band, are routinely used in clinical practice as markers of the irritative zone^6,8,9. HFOs, which are considered more specific biomarkers of the EZ compared to spikes^10,11, are commonly classified into ripples (80–250 Hz) and fast ripples (250–500 Hz)^12–14. Fast ripples, although closely linked to the EZ^15–17, are less frequently observed in patients with DRE, and their low sensitivity may limit their value for EZ localization^18,19. On the other hand, ripples are present in most patients with DRE but are also observed in physiological brain regions^6,10,20,21. Thus, the presurgical value of interictal biomarkers, such as spikes and HFOs, as biomarkers of the EZ remains largely debated since the cortical areas that generate them often extend beyond the EZ into brain regions that do not need to be resected (or ablated) for the patient to become seizure-free^11,20,22,23.

Because of these limitations, temporal (i.e., timing and temporal occurrence), spectral (i.e., frequency content), and spatial (i.e., distribution across brain regions) features of spikes and ripples have been examined as potential indicators of epileptogenicity. In the temporal domain, the spike and ripple rates have been linked to the underlying brain tissue’s epileptogenicity^22,24,25. Yet, spike-generating regions correspond to the broader irritative zone, which can extend beyond the EZ and overlap with physiological brain regions^26,27. Moreover, not all ripples are pathological, as physiological ripples also occur, particularly in brain regions such as the hippocampus, occipital cortex, and paracentral areas^28–30. Consequently, spike and ripple rates alone are unreliable predictors of the EZ. Recent studies have shown that the temporal co-occurrence of spikes and ripples may predict the EZ more effectively, or with comparable efficacy, compared to spikes or ripples alone^22,25,31. This improved performance likely reflects the co-occurrence of spikes and ripples as a more pathologic pattern of neuronal synchronization than either event in isolation²⁵. Yet, the interpretation of ripples remains challenging due to their susceptibility to contamination by artifacts arising from the Gibbs phenomenon, particularly in recordings containing epileptiform spikes^32,33. In the spectral domain, frequency-based features have been shown to identify abnormal oscillatory activity associated with epileptogenic networks^34–36. However, they are limited by their sensitivity to sensor noise and biological artifacts^37,38. Spatial analyses have demonstrated that spikes and ripples often originate in specific brain regions and propagate to adjacent and distant areas within a few tens of milliseconds^10,11,39,40. Recent studies have also shown that the onset generator of spikes and ripples (i.e., regions where discharges show the earliest peaks) are correlated with the EZ^10,39,41,42, often outperforming the clinically defined SOZ in predicting the surgical outcome^11,39,43. However, the spatial distribution of spikes and ripples fluctuates over time, shifting among multiple regions rather than arising from a single consistent source^44,45. Therefore, no single biomarker fully defines the EZ; rather, each reflects only a partial and variable aspect of epileptogenicity, highlighting the need for complementary approaches.

It is reasonable to presume that the integration of temporal, spectral, and spatial features of interictal biomarkers will yield a more comprehensive and precise estimation of the EZ^46–48. Yet, the practical implementation of this approach is technically challenging. iEEG recordings are often performed with several dozens of contacts sampled at high rates (> 1 kHz) and collected over several hours. As a result, they generate a massive amount of data, making their analysis time-consuming, labor-intensive, and prone to bias and error. This underscores the need to develop novel automated approaches that integrate the complementary strengths of these biomarkers and features into a unified framework, enabling a more objective, efficient, and accurate characterization of the EZ.

Machine learning (ML) offers a robust framework for automating the analysis of large-scale data and integrating heterogeneous, high-dimensional features that are difficult to analyze manually or with conventional statistical methods^49,50. Using an ML framework, our group has recently shown that interictal iEEG recordings can delineate the EZ with high precision and predict the surgical outcome of patients with DRE, eliminating the need for manual data inspection, reducing prolonged monitoring, and enhancing surgical planning³⁶. Yet, our ML framework was relying on iEEG signal decomposition that has several limitations. Decomposition methods primarily summarize intracranial EEG in terms of spectral or basis-function content and may therefore fail to fully capture very brief or highly transient epileptiform discharges, especially when not explicitly tuned for such events⁵¹. Moreover, because these approaches are applied on multichannel short time windows, they may not characterize epileptiform propagation dynamics or the spatial extent of epileptiform activity across contacts.

Previous studies have incorporated temporal, spectral, and spatial features using ML to estimate the EZ in patients with DRE^{36,46,52–60}. Yet, only few of them have further compared the performance of individual features with multi-feature ML-based frameworks’ findings^46,53. This comparison is critical, as the epileptogenic brain regions identified by individual spike and ripple features may differ across patients, or map incompletely to the EZ or extend beyond it. Furthermore, event-based biomarkers such as spikes and ripples capture localized pathological activity and provide physiologically interpretable features that may improve the identification of clinically relevant brain regions. These limitations highlight the need for methods that systematically integrate heterogeneous biomarkers for EZ prediction. ML-based frameworks are well-suited for this task, as they can fuse multiple feature types, thereby improving accuracy, robustness, and interpretability in EZ mapping and surgical outcome prediction.

In this study, we propose an ML-based multimodal framework that integrates diverse interictal spike and ripple temporal, spectral, and spatial features to automatically estimate epileptogenic contacts and predict postsurgical outcome in patients with DRE. We hypothesize that the proposed framework outperforms single-feature analyses in estimating epileptogenic contacts and predicting the outcome by capturing nonlinear and cross-domain relationships among epileptogenic biomarkers. To evaluate this framework, automatic spike and ripple detectors were applied to iEEG recordings from children and young adults with DRE to extract multidomain electrode-level features. These features were then used to train Random Forest classifiers, with different combinations of features as predictors and resection as the target, to estimate the epileptogenic contacts to resect and define the predicted EZ (PEZ). The PEZs derived from individual spike and ripple features and from the proposed feature-combining approach were analyzed to evaluate their ability to predict the EZ. We then assessed whether resection of each PEZ could predict surgical outcome.

Materials and methods

Patient cohort

We retrospectively analyzed iEEG data of patients with DRE, from Boston Children’s Hospital (BCH, 51 patients, admitted between June 2011 and June 2023) and Cook Children’s Medical Center (CCMC, 11 patients, admitted between December 2019 and August 2021). The patients had long-term monitoring with iEEG and underwent resection neurosurgery or ablation. The selection of the patients was based on the following criteria: (i) availability of at least 5-min interictal iEEG epochs with no biological or technical artifacts; (ii) availability of post-implantation computerized tomography (CT); (iii) availability of pre-operative and post-operative MRIs (or intraoperative MRIs) or resected electrodes from operative notes; (iv) accurate information about the clinically defined SOZ; (v) availability of post-surgical outcome at least one year after surgery; (vi) data containing spikes and ripples; (vii) sampling frequency > 1,000 Hz. The surgical outcome was determined by a pediatric epileptologist after multiple follow-up visits. We used the Engel score⁶¹ to classify patients as good (Engel I) or poor outcome (Engel II-IV). The patient demographic and the clinical information of the cohort are provided in Table 1. The protocol was approved by the 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.

Table 1.

Patient demographic information categorized by post-surgical outcome.

Characteristic	Total	Good outcome (Engel = I)	Poor outcome (Engel > I)	p
Patients	62	34	28	-
Gender				0.17a
Male	31	18	13
Female	31	16	15
Age at surgery, yr., median (IQR)	12 (8–15.5)	12 (7–15.5)	12 (8–15)	0.8b
Age at seizure onset, yr., median (IQR)	4.5 (2–9)	6 (2–10.25)	4 (2–6.5)	0.08b
Follow up, yr., median (IQR)	2.5 (2–5)	2.5 (2–5)	2.5 (1–5)	0.8b
Number of contacts, no., median (IQR)	124 (102–151)	112 (88–136)	138 (106–168)	0.09b
Type of implantations				0.73a
Subdural (ECoG)	13	8	5
sEEG	36	20	16
Subdural and depth	13	6	7
Implantation side				0.96a
L	40	22	18
R	16	9	7
L + R	6	3	3
Eplilepsy localization				0.16a
Temporal	26	17	9
Extra-temporal	36	17	19
MRI Findings				0.73a
NL	17	9	8
MCD	34	20	14
ACQ	11	5	6

ACQ = Acquired (i.e., stroke, neoplasm, and traumatic brain injury); ECoG = Electrocorticography; IQR = Inter-Quartile Range; L = Left; MCD = Malformation of Cortical Development (i.e., focal cortical dysplasia, polymicrogyria, gliosis, and tuberous sclerosis complex); NL = Non-Lesional; no. = Number; p = p-values; R = Right; sEEG = stereotactic electroencephalography; yr. = Year. ^a Chi-squared test; ^b Wilcoxon rank-sum test (two-tailed).

Interictal iEEG recordings

At BCH, long-term iEEG data were collected using stereotactic electrodes, subdural grids, or combined subdural grids and depth electrodes utilizing XLTEK Quantum NeuroWorks (Natus Inc., USA). The subdural electrodes had a diameter of 2–3 mm with a 10 mm inter-contact spacing, while the depth electrodes consisted of 6 to 16 contacts with a diameter of 0.8 mm and were linearly arranged approximately 3–5 mm apart. At CCMC, long-term iEEG data were collected using stereotactic electrodes utilizing Quantum NeuroWorks (Natus Inc., USA). The electrodes consisted of 8 to 16 contacts with a diameter of 0.8 mm and were linearly arranged approximately 3–5 mm apart.

The data were recorded with a sampling frequency of 1,000 to 2,048 Hz. We specifically chose 5-min interictal segments during non-REM slow-wave sleep (when applicable), occurring at least one hour before or after clinical seizures or half an hour before or after an electrographic seizure⁶². This approach ensured the incorporation of epochs with the highest spike rate and minimal motion artifacts^63,64. Noisy channels and short segments with artifacts were excluded.

Structural imaging and iEEG recordings

MRI and iMRI scans utilizing standard magnetization-prepared rapid acquisition gradient-echo sequences were performed with a 3 T scanner (TIM TRIO, Siemens AG at BCH and Magnetom Skyra at CCMC). Post-implantation CT scans, with voxel size set at 0.5 × 0.5 × 0.5 mm³, were performed after iEEG electrode implantation to document electrode locations. We aligned the electrode coordinates by registering the post-implantation CT scans with the preoperative MRI using Brainstorm⁶⁵. We then projected the electrode locations onto the patient’s cortical surface reconstructed from their preoperative MRI with Freesurfer⁶⁶. We further adjusted the stereotactic and depth electrode positions to accommodate for brain shifting and pneumocephalus post-electrocorticography implantation⁶⁷. The number and type of implanted electrodes for each patient are reported in Supplementary Tables 1 and 2.

Defining the SOZ and resection contacts

For each patient, pediatric epileptologists determined the SOZ by visually examining ictal data. iEEG contacts that exhibited altered dynamics at the onset of seizures were designated as SOZ contacts. This process was conducted independently of the current study. To delineate the resection volume, we registered the preoperative MRI to the postoperative MRI [or the intraoperative MRI (iMRI) when a postoperative MRI was unavailable] and manually outlined the resection (or ablation) boundaries on sequential slices using Brainstorm⁶⁵ and MATLAB’s volumeSegmentortoolbox. Postoperative MRI was available for 33 patients and was acquired at a median of 28.4 months after surgery (IQR 2.3–54.6 months); in one patient, the postoperative MRI was obtained within one week of surgery and was used for the delineation of the resection-cavity. In 13 patients, iMRI was used to generate the resection volumes. In cases relying on iMRI or postoperative MRI obtained within one-week post-surgery, a neuroradiologist reviewed (M.B.) and confirmed the resection segmentation boundaries to minimize confounds from postoperative edema and uncertain tissue interfaces. Subsequently, contacts ≤ 10 mm from the delineated resection (or ablation) volume were considered as resected (RES). The selection of the 10 mm cut-off was based on studies that showed that the gyral width is between 11–21 mm⁶⁸. For 15 patients, operative notes were used to identify resected contacts, and no resection volume was constructed (Supplementary Tables 1 and 2).

Overall processing pipeline

The proposed framework automatically identifies the epileptogenic contacts from interictal iEEG data through automated detection of spikes and ripples, spike and ripple features extraction, and ML (Fig. 1). We initially preprocessed a 5-min interictal iEEG segment to remove artifacts and noisy channels (Fig. 1a). The data were then bandpass filtered into the spike (1–70 Hz)³⁹ and ripple (80–250 Hz)^44,69 frequency bands. Spikes and ripples were detected using previously validated fully automated algorithms^10,70. Following automated spike detection, an expert reviewer (C.P.) performed a post hoc quality-control review to retain isolated single spikes and to exclude detections corresponding to bursts of poly-spikes and/or persistent periodic/rhythmic discharges. Ripples were detected using a previously published fully automated algorithm with predefined criteria requiring ≥ 4 oscillations, a local maximum ≥ 5 SD above the Hilbert-envelope baseline, the presence of an isolated island of activity in the time–frequency plane across the ripple duration¹⁰, and a minimum inter-event interval of 10 ms between ripple events. From the detected spikes and ripples, we extracted a set of temporal, spectral, and spatial features. We used the RES as the gold standard for defining the epileptogenic contacts: contacts inside the resection were labeled as epileptogenic (EP), and those outside as non-epileptogenic (non-EP). We first evaluated the discriminatory power of each feature for classifying the EP vs. non-EP contacts. We then trained Random Forest classifiers to automatically identify the EP contacts. The set of EP contacts identified by the classifiers were defined as the PEZ. Finally, we assessed whether resecting the PEZ, which was mapped by each feature and the classifiers, could predict surgical outcome (Fig. 1a).

Fig. 1.

Processing pipeline for automatic identification of epileptogenic contacts from interictal iEEG. (a) Five-minute interictal iEEG segments were preprocessed, cleaned, and filtered in two frequency bands: spike band (1–70 Hz) and ripple band (80–250 Hz). Automatic spike and ripple detectors identified spike and ripple events, from which temporal, spatial, and spectral features were extracted. The seizure onset zone (SOZ) and resection zone (RES) were defined based on clinical reports, intraoperative and postoperative MRI-defined resection volumes. Multiple Random Forest classifiers were trained using different feature sets for mapping the predicted epileptogenic zone (PEZ), and the correspondence between the PEZ and surgical outcome was evaluated. (b) In the spike band, the automatic spike detector identified spikes (blue). Detected propagation events (p1–p3) were analyzed to compute a spike propagation index per channel. Channels exhibiting earliest activation within each event were defined as onset spikes (red signal in p2). (c) In the ripple band, the automatic ripple detector identified ripples (yellow). Similar propagation analysis of detected events (p4–p7) yielded a ripple propagation index per channel, and channels showing earliest activation were defined as onset ripples (red signal in p7). (d) Spectral power features were computed for spikes and ripples across all channels and normalized within [0,1]. (e) Spike and ripple rates were calculated per channel to characterize interictal activity. (f) Spatial distribution features were derived using Euclidean distances between each channel involved in the events and the contacts with highest event rate, maximum power, and onset activity. (g) Coregistration of preoperative MRI with CT scans containing intracranial implantations was used to localize iEEG electrode coordinates, with the contacts located within the resection defined as resected contacts (RES), which served as the gold standard for epileptogenic zone (EZ) prediction. (h) Multiple Random Forest classifiers were trained using three different feature sets [spike features (SF), ripple features (RF), and combined features (CF)] to classify contacts as epileptogenic (EP) or non-epileptogenic (non-EP).

Spike and ripple features

Table 2 summarizes the features adopted in this study. To generate the features used in our classifiers, we focused on quantifying the spectral and spatiotemporal characteristics of spike and ripple propagations. We developed an in-house MATLAB algorithm to identify propagation sequences of spikes and ripples separately across all iEEG contacts for each patient¹¹. After detecting the spikes, we employed our algorithm that marks the earliest spike within each sequence as onset (Fig. 1b)¹¹. Then, it adds the next spike to the propagation sequence if it occurred within a predefined time window, i.e. the inter-event latency. If the time delay between two successive spikes exceeded the inter-event latency threshold, that spike and any later spikes were not included in the propagation sequence. Contacts in every propagation were indexed sequentially, starting from one, denoting the onset contact, followed by increasing numbers for the subsequent contacts. We did not impose additional morphology, frequency, or duration criteria for sequence membership because interictal spike morphology and duration can vary across contacts^11,71and interictal ripples vary in duration and dominant frequency across events and recording cites^14,72. Finally, the spike propagation index (SPI) was calculated for each channel as the median of the per-propagation participation scores assigned to that channel. Median was used for robustness to occasional extreme scores. A similar approach was used for mapping ripple propagations and assigning a ripple propagation index (RPI) to the channels (Fig. 1b-c). Previous studies have shown that the inter-event latency for spikes and ripples is in the range of 3–15 ms^10,73 and 10–32 ms^10,41, respectively. Based on these findings, we set here the inter-event latency to 10 and 30 ms, to map spike and ripple propagations respectively¹¹. We also discarded propagations if more than 50% of their events occurred within 2 ms from each other^52,74. Using spikes and their propagation events, we computed the following features for each contact (Fig. 1d-f): (i) spike rate (SR); (ii) average spike power (SP); (iii) SPI; (iv) the Euclidean distance of the contact from the contact with the highest spike rate (ED-HSR); (v) average Euclidean distance of the contact from the contact with the highest average spike power (ED-HSP) in each propagation; and (vi) average Euclidean distance of the contact from spike onset contact (ED-SO) in each propagation. Similarly, using ripples and their propagation events, we computed for each channel the: (i) ripple rate (RR); (ii) average ripple power (RP); (iii) RPI; (iv) the Euclidean distance of the contact from the contact with the highest ripple rate (ED-HRR); (v) average Euclidean distance of the contact from the contact with the highest ripple power (ED-HRP) in each propagation; and (vi) average Euclidean distance of the contact from ripple onset contact (ED-RO) in each propagation. We finally computed the spike and ripple cooccurrence rate (SRC). For the power calculation, we computed the signal power for each detected spike using the Welch method within a 500 ms window centered on the spike detection timestamp. Specifically, we chose a 500 ms window to minimize the likelihood of including adjacent spikes or other signal artifacts that could contaminate the measurement. For ripple events, we employed a shorter 100 ms window^9,75. For each channel, the powers of spikes and ripples were then averaged across the 5-min data segments.

Table 2.

Summary of spike and ripple features.

Biomarker	Feature	Description	Type	Normalization	Physiological Meaning
Spike (1–70 Hz)	SR	Spike rate (min⁻1)	Temporal	Min–Max	Interictal spike excitability
	SP	Spike power	Spectral	Min–Max	Signal energy
	SPI	Spike propagation index	Temporal	Max–Min	Degree of spike spread across channels
	ED-HSR	Euclidean distance from highest spike rate contact in each propagation	Spatial	Max–Min	Proximity to maximal spike-generating region
	ED-HSP	Average Euclidean distance from highest spike power contact in each propagation	Spatial	Max–Min	Proximity to strongest spike source
	ED-SO	Average Euclidean distance from spike onset contact in each propagation	Spatial	Max–Min	Distance from earliest spike onset
Ripple (80–250 Hz)	RR	Ripple rate (min⁻1)	Temporal	Min–Max	High-frequency excitability
	RP	Ripple power	Spectral	Min–Max	Signal energy
	RPI	Ripple propagation index	Temporal	Max–Min	Spatial spread of ripple activity
	ED-HRR	Euclidean distance from highest ripple rate contact in each propagation	Spatial	Max–Min	Proximity to dominant ripple-generating region
	ED-HRP	Average Euclidean distance from highest ripple power contact in each propagation	Spatial	Max–Min	Proximity to strongest ripple source
	ED-RO	Average Euclidean distance from ripple onset contact in each propagation	Spatial	Max–Min	Distance from earliest ripple onset
Spike–Ripple	SRC	Spike–ripple co-occurrence rate	Temporal	Min–Max	Pathological coupling of spikes and ripples

Normalization

We normalized SR, RR, SP, RP and SRC between 0 and 1 using Min–Max scaling ensuring that the minimum value is scaled to 0 and the maximum to 1. For a feature , the min–max normalized value is defined by:

1

where is the minimum value and is the maximum.

The SPI, RPI, ED-HSR, ED-HRR, ED-HSP, ED-HRP, ED-SO, and ED-RO were scaled using Max–Min scaling ensuring that the maximum value is scaled to 0 and the minimum to 1. For a feature , the max–min normalized value is defined by:

2

Performance of individual features in EZ and outcome prediction

For each feature, we estimated and compared the normalized values inside and outside the SOZ and resection for good and poor outcome patients, separately (Wilcoxon signed-rank test). Then, to further evaluate the predictive power of individual features, we used the normalized values of each feature as predictors and RES as target. We performed the analysis only for good outcome patients since their resected tissue is assumed to contain the actual EZ. First, we assessed feature importance using one-way ANOVA. For each feature, we quantified the proportion of variance explained by the class labels (η²)⁷⁶, which provides a measure of how strongly a feature differentiates between classes (i.e. epileptogenic vs. non-epileptogenic). Values of η² greater than 0.06 generally indicate that a feature has a meaningful influence on distinguishing between epileptogenic and non-epileptogenic classes. Next, for each feature, we used a leave-one-patient-out cross-validation (LOPO-CV) approach to assess its ability to predict the RES. At each step, one patient was left out, and the remaining patients’ feature values were collectively used to compute the optimal thresholds based on the maximum Youden index (J) obtained from the ROC curves (Supplementary Note 1). Then, the left-out patient’s features were thresholded using the computed optimal threshold to identify PEZ. We then defined the following: (i) true positives (TP), the number of EP contacts in RES; (ii) false positives (FP), the number of EP contacts not in RES; (iii) false negatives (FN), the number of non-EP contacts in RES; and (iv) true negatives (TN), the number of non-EP contacts not in RES. We then calculated the following performance metrics: (i) sensitivity [TP/(TP + FN)]; (ii) specificity [TN/(TN + FP)]; (iii) precision [TP/(TP + FP)]; (iv) negative predictive value (NPV) [TN/(TN + FN)]; (v) accuracy [(TP + TN)/(TP + TN + FP + FN)]; (vi) F1-score [2 TP/(2 TP + FP + FN)]; (vii) area under the receiver operating characteristic curve (AUC-ROC); and (viii) area under the precision-recall curve (AUC-PR). The overlap with resection ( was defined as the percentage of PEZ contacts from RES. We hypothesized that the resection of PEZ is linked to favorable surgical outcome, thus, we considered as predictor and the dichotomized patient’s outcome as target. We considered: (i) TP, the number of good outcome patients who had ≥ 50% of the PEZ resected (≥ 50%); (ii) FN, the number of good outcome patients who had < 50% of the PEZ resected (< 50%); (iii) FP, the number of poor outcome patients who had ≥ 50% of the PEZ resected; and (iv) TN, the number of poor outcome patients who had < 50% of the PEZ resected. We then constructed the confusion matrix for each feature and computed the Fisher’s exact test p-values to find significance. We further evaluated whether features in RES differ across various pathological subtypes: non-lesional MRI (NL), malformation of cortical development (MCD), and acquired brain injury (ACQ). In addition, we compared RES features of patients with temporal lobe epilepsy (TLE) with non-temporal lobe epilepsy (non-TLE).

Training and evaluation of Random Forest classifiers

We trained and compared the performance of different Random Forest classifiers using different subsets of predictors. We considered three sets of features as predictors: spike features only (SF), ripple features only (RF), and combined features (CF). We used RES to approximate the EZ. The setup resulted in training and evaluating three different classifiers with different combination of predictors (i.e., SF, RF, and CF). We exclusively trained and evaluated our classifiers using only patients with good outcome, since these patients remained seizure-free post-surgery, indicating successful removal of the EZ. Regarding the Random Forest classifier’s number of trees parameter, we conducted a five-fold cross-validation, varying the number of trees from 10 to 500 and selecting the value that yielded a stable AUC-ROC performance. Based on this analysis, we set the number of trees to 100 (Supplementary Note 2).

Classifier performance was evaluated using TP, TN, FP, and FN definitions identical to those used for individual feature analyses. Subsequently, we trained and tested the performance of the Random Forest classifier using LOPO-CV (Supplementary Note 3) and computed performance metrics (i.e., sensitivity, specificity, precision, NPV, accuracy, F1-score, AUC-ROC, and AUC-PR). We then quantified the PEZ with spatial measures: (i) focality (); (ii) overlap with resection (), and (iii) average distance from resection ()³⁶. We defined the F_PEZ as the average Euclidean distance (in mm) between the PEZ contacts. We defined the D_RES as the average Euclidean distance of the PEZ contacts to RES contacts. After estimating the PEZ using the trained classifiers, we computed the , , and . While training was performed using only good outcome patients, testing included both good and poor outcome patients.

Robustness of PEZ across implantation types

To evaluate whether our framework provides consistent findings among different implantation types, we examined the ability of PEZ to predict the epileptogenic contacts in good outcome patients with different implantation types. We first divided our good outcome patients into three subsets; (i) ECoG, patients with subdural implantations; (ii) sEEG, patients with stereotactic implantations; and (iii) ECoG and depth, patients with ECoG and a few depth electrodes (Supplementary Table 1). For each subset, we trained and tested using CF with LOPO-CV and computed performance metrics (sensitivity, specificity, precision, NPV, accuracy, F1-score, AUC-ROC, AUC-PR) along with spatial properties (, Dice score, , ) of the PEZ.

Random forest classifiers’ performance to predict surgical outcome

We presumed that the of PEZ predicted from the classifiers SF, RF, and CF could predict outcome. Specifically, we assumed that in patients with good outcome, the PEZ would be located within resection, whereas in those with poor outcome, the resection would have missed the EZ. Subsequently, we performed a prediction analysis where we used the of the PEZ as predictor of outcome. We considered good and poor outcomes as positive and negative classes, respectively. We then defined: (i) TP, the number of good outcome patients who had ≥ 50% of the PEZ resected ( ≥ 50%); (ii) FN, the number of good outcome patients who had < 50% of the PEZ resected ( < 50%); (iii) FP, the number of poor outcome patients who had ≥ 50% of the PEZ resected; and (iv) TN, the number of poor outcome patients who had < 50% of the PEZ resected. We constructed the confusion matrix for each feature set and computed the Fisher’s exact test P-values to assess significance. Finally, we computed the sensitivity, specificity, precision, NPV, F1-score, and accuracy of the outcome predictions. To assess potential differences between short-term and long-term outcome prediction, we performed an analysis including only patients with available Engel scores at both one-year and latest follow-up (Supplementary Tables 1 and 2). We evaluated outcome predictions separately for the two follow-up periods by constructing confusion matrices. We then applied McNemar’s test to compare the classification performance between the two follow-up periods.

Statistical analysis

The Kolmogorov–Smirnov test was used to test the normality of features. Cliff’s d measure was used to compute effect sizes. Demographic data and PEZ spatial measures (, , Dice score, and ) are reported as median (25^th −75^th percentile). The chi-squared test was used to compare the effect of different cohort characteristics (i.e., gender, implantation side, epilepsy localization, and MRI findings) on outcome. We applied two-sided non-parametric Wilcoxon signed-rank test for all paired comparisons (median feature values inside vs. outside RES). We applied the two-sided Wilcoxon rank-sum test for non-paired comparisons between good and poor outcome patients. Bonferroni correction was applied in all multiple comparison tests.

We used the one-sided Fisher’s exact test to evaluate the predictive value of the PEZs to predict outcome. To evaluate the performance of all individual features and trained Random Forest classifier models, we performed a comparative analysis of epileptogenic contact prediction using the AUC-ROC, AUC-PR, and . Pairwise comparisons were performed using the Wilcoxon signed-rank test on each performance metric. False discovery Rate (FDR) was applied to correct for multiple comparisons. The results are displayed in a square matrix where each element indicates whether a significant difference (P < 0.05) between the different predictor models was observed and, if so, identifies the model with the higher median value.

For outcome prediction using individual features, classifier models, and SOZ, we performed a pairwise comparison using McNemar’s test. The elements in the corresponding square matrix show both the presence of significant differences (P < 0.05) and the ID of the model with higher accuracy. McNemar’s test was used to examine differences in the ability to predict outcomes using confusion matrices between different features and the proposed ML classifiers. All analysis was performed with MATLAB 2023b (The MathWorks, Inc.).

Results

Patient cohort

We retrospectively analyzed interictal iEEG data from 62 children and young adults (29 females) with focal onset DRE, having either a good (34 patients, Engel I) or poor (28 patients, Engel I) surgical outcome following epilepsy surgery [median follow-up: 2.5 years (2–5)], who satisfied the study’s inclusion and exclusion criteria. Forty-eight patients underwent resection surgery, and 14 had thermal ablation. Patient demographics are summarized in Table 1. The patients had a median age at surgery of 12 years (range: 8–15.5) with a median seizure onset age of 4.5 years (2–9). Our dataset included patients with three implantation types: (i) subdural electrodes [electrocorticography (ECoG)] (13 patients; eight with good outcome); (ii) stereotactic EEG (sEEG) implantation (36 patients; 20 with good outcome); and (iii) both subdural and depth electrodes (13 patients; six with good outcome). The median number of implanted contacts was 124 (102–151). Seventeen patients (27.4%) had non-lesional MRI, 34 patients (54.8%) had a malformation of cortical development, and 11 patients (17.8%) had acquired brain injury. Twenty-six patients (41.9%) had temporal lobe epilepsy. No significant differences in these characteristics were found between good and poor outcome patients (Table 1).

Histopathology was available for 45 patients (specimens submitted after resection). Sixteen patients did not have histopathology results because 14 of them underwent laser ablation, and two resection cases did not have a specimen submitted for testing. Among the 45 patients with pathology, preoperative MRI and histopathology were concordant in 22 cases (i.e., MRI findings and histopathology fell within the same pathology category). In 13 cases, MRI was abnormal, but the histopathology category differed from the MRI-based diagnosis. In 10 cases, preoperative MRI was reported as normal, but histopathology demonstrated a pathologic abnormality (Supplementary Tables 1 and 2).

Predictive power of individual features

In good outcome patients, we observed increased SR (P < 0.001), SP (P < 0.001), SPI (P < 0.01), ED-HSR (P < 0.0001), ED-HSP (P < 0.0001), ED-SO (P < 0.0001), RP (P < 0.001), ED-HRR (P < 0.01), ED-HRP (P < 0.05), ED-RO (P < 0.001), and SRC (P < 0.05) inside (compared to outside) resection (Fig. 2a). In poor outcome patients, we observed higher SR (P < 0.05), SP (P < 0.05), ED-HSP (P < 0.05), ED-SO (P < 0.01), RP (P < 0.05), and ED-RO (P < 0.05) inside (compared to outside) resection (Fig. 2a). Upon comparing RES features across various pathological subtypes, we found no significant differences between features of NL, MCD, and ACQ (Supplementary Note 4). Similarly, we found no differences in RES features between TLE and non-TLE (Supplementary Note 4). To assess the predictive power of each feature, we computed their performance in predicting the epileptogenic contacts in good outcome patients (Fig. 2b). Regarding feature importance, the top-ranked features ED-SO, ED-HSP, SR, ED-HSR, SPI, and ED-RO exhibited strong discriminative power in differentiating between epileptogenic and non-epileptogenic contacts (η² > 0.06) (Fig. 2b). Additionally, SR, ED-HSR, ED-HSP, and ED-SO achieved the highest AUC-ROC and AUC-PR (> 74% and 66%, respectively) and exhibited > 60% overlap with resection. The resection of the PEZ, estimated from most single-feature classifiers, also predicted outcome, except for the SPI, RR, and RPI (Fig. 2b).

Fig. 2.

Predictive power of individual features. (a) Comparison of feature values inside versus outside the resection zone in good and poor outcome patients. For each feature, the red box plot represents the median feature value inside the resection, and the blue box plot represents the median value outside the resection. The notches indicate the median across all patients, while the lower and upper box edges correspond to the 25^th and 75^th percentiles. Whiskers extend to the minimum and maximum values after omitting outliers. Significant differences are indicated by asterisks: *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001 (Wilcoxon signed-rank test, significance levels were corrected for multiple comparisons using the Bonferroni method). The number displayed below the asterisks represents the effect size. (b) Ranked importance (effect size) and performance of individual features in predicting the resection (RES) in good outcome patients. Performance measures were evaluated using Youden index–based optimal thresholding with leave-one-patient-out cross-validation (LOPO-CV). Asterisks denote statistical significance of features in predicting surgical outcome based on the Fisher test (*P < 0.05). Features: SR (spike rate), SP (spike power), SPI (spike propagation index), ED-HSP (distance from highest spike power), ED-HSR (distance from highest spike rate), ED-SO (distance from spike onset), RR (ripple rate), RP (ripple power), RPI (ripple propagation index), ED-HRP (distance from highest ripple power), ED-HRR (distance from highest ripple rate), ED-RO (distance from ripple onset), SRC (spike and ripple co-occurrence).

Predictive power of feature sets

Upon comparing the performance of the Random Forest classifiers using different feature sets (i.e., SF, RF, CF) as predictors of the EZ, SF and CF achieved higher AUC-ROC (0.9 and 0.89, respectively) compared to most individual features, except for the ED-SO (0.9). In terms of AUC-PR, SF, CF, and ED-SO showed higher values compared to all individual features (AUC-PR = 0.76, 0.78, and 0.77, respectively). Many individual features, however, showed higher AUC-ROC and AUC-PR compared to RF. Regarding overlap with resection, SF, RF, and CF all achieved > 74%, which was higher than the individual features ( ≤ 70%) (Fig. 3a).

Fig. 3.

Predicting epileptogenic contacts using individual features and different feature sets. (a) Performance metrics were evaluated for different classifiers in predicting the EZ (PEZ) using individual features and spike features (SF), ripple features (RF), and combined features (CF). Metrics included area under the receiver operating characteristic curve (AUC-ROC), area under the precision–recall curve (AUC-PR), and the overlap with resection (). Classification performance was assessed using a leave-one-patient-out cross-validation (LOPO-CV) approach trained on good outcome patients. The notches represent the median metric value across all patients, while the lower and upper box edges correspond to the 25^th and 75^th percentiles. Whiskers extend to the minimum and maximum values after excluding outliers. The number above each boxplot represents the median value. (b) Pairwise comparison of AUC-ROC, AUC-PR, and for epileptogenic zone prediction, derived from individual interictal features and machine learning classifiers using SF, RF, and CF as predictors. Each red cell indicates a significant difference (P < 0.05, Wilcoxon signed-rank test, FDR corrected) between the corresponding pair of models, with the ID of the better-performing model (higher median AUC-ROC, AUC-PR, and , respectively) displayed inside the cell. Blue cells indicate no differences. Features: SR (spike rate), SP (spike power), SPI (spike propagation index), ED-HSP (distance from highest spike power), ED-HSR (distance from highest spike rate), ED-SO (distance from spike onset), RR (ripple rate), RP (ripple power), RPI (ripple propagation index), ED-HRP (distance from highest ripple power), ED-HRR (distance from highest ripple rate), ED-RO (distance from ripple onset), SRC (spike and ripple co-occurrence).

Statistical pairwise comparisons between the performance measures of individual features and classifiers showed that SF and CF outperformed all individual features in terms of AUC-ROC and AUC-PR, except for the ED-HSR, ED-HSP, and ED-SO. RF showed higher performance than some individual features but did not exceed SF or CF. No differences were observed between SF and CF for the AUC-ROC and AUC-PR. Regarding the , SF and CF outperformed all individual features except the SRC, while no differences were observed among the SF, RF, and CF (Fig. 3b).

Effect of implantation type on EZ predictability

We next examined whether the implantation type (i.e., ECoG, sEEG, or both subdural and depth electrodes) influenced the PEZ’s predictive power. We found no differences between the AUC-ROC, AUC-PR, and of the PEZs with resection among the different implantation types for any of the feature sets (Supplementary Note 5).

Focality, proximity, and overlap of PEZ with resection

To assess the concordance of PEZ with RES, we compared , , and across patients with good and poor outcomes for the three feature subsets. We found that the SF, RF, and CF had higher and lower in good compared to poor outcome patients (P < 0.01, d ). No differences in were observed between good and poor outcome patients for any of the feature sets. Notably in good outcome patients, the SF had a median 74 (56–95)%, RF had a median 75 (50–85)%, and CF had a median 74 (57–94)% (Fig. 4a).

Fig. 4.

Spatial properties of the predicted epileptogenic zone (PEZ) and outcome prediction. (a) Overlap with resection (, %), distance from resection (, mm), and focality (, mm) were computed for PEZ identified by classifiers trained using spike features (SF), ripple features (RF), and combined features (CF) as predictors for good and poor outcome patients. The notches indicate the median values across all patients, while the lower and upper box edges correspond to the 25^th and 75^th percentiles. Whiskers extend to the minimum and maximum values after omitting outliers. Significant differences are indicated by asterisks: *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001 (Wilcoxon rank-sum test, significance levels were corrected for multiple comparisons using the Bonferroni method). (b) Outcome prediction based on resection of the PEZ. Confusion matrices show outcome prediction when 50% of the PEZ was resected. The PEZ was predicted using SF, RF, and CF. The seizure onset zone (SOZ) was also assessed for outcome prediction. Evaluation metrics included sensitivity, specificity, precision, negative predictive value (NPV), F1-score, and accuracy. Significant differences are indicated by asterisk above the confusion matrices: *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001 (Fisher’s exact test). (c) Pairwise comparison of individual feature and feature subset-based classifiers (SF, RF, and CF) for outcome prediction using McNemar’s test. Each yellow cell indicates a significant difference (p < 0.05, McNemar’s test) between the corresponding pair of models, with the ID of the better-performing model (higher average prediction) displayed inside the cell. Blue cells indicate no differences. Features: SR (spike rate), SP (spike power), SPI (spike propagation index), ED-HSP (distance from highest spike power), ED-HSR (distance from highest spike rate), ED-SO (distance from spike onset), RR (ripple rate), RP (ripple power), RPI (ripple propagation index), ED-HRP (distance from highest ripple power), ED-HRR (distance from highest ripple rate), ED-RO (distance from ripple onset), and SRC (spike and ripple co-occurrence).

Surgical outcome prediction across features and classifiers

PEZ derived from SF, RF, and CF predicted outcome (P < 0.01, Fisher Exact test), while the clinically defined SOZ did not predict outcome (Fig. 4c). Among all models, the CF model predicted outcome with the highest sensitivity (88%), specificity (68%), precision (77%), NPV (83%), F1-score (82%), and accuracy (79%) (Fig. 4c). Pairwise comparisons of outcome predictions revealed that SR performed better than RPI and the SOZ, and ED-SO performed better than the SOZ (P < 0.05, McNemar’s test). Additionally, SF showed better predictive performance compared to RPI and the SOZ (P < 0.05, McNemar’s test) (Fig. 4c). Finally, CF had higher predictive power of surgical outcome than all individual features except the SR, ED-HSR, ED-HSP, ED-SO, ED-HRP, SF, and RF (P < 0.05, McNemar’s test). Upon comparing short- and long-term predictions, we observed similar classification performance across the two follow-up periods (P > 0.05, McNemar’s test) (Supplementary Note 6).

Representative cases

To illustrate how the PEZ relates to the EZ, we report findings from four representative cases: two patients with good and two patients with poor surgical outcome. For these patients, we depict the PEZ overlayed on their MRIs and the resection volume (Fig. 5). For good outcome patient #6, the PEZ derived from the CF (O_RES = 71%) and SF (O_RES = 51%) showed greater overlap with the resection volume than that from the RF (O_RES = 41%) (Fig. 5). The PEZs from the CF and SF were also more focal (= 22 mm and 23 mm, respectively) and closer to the resection (= 12 mm and 13 mm, respectively) compared to the RF (= 32 mm, =19 mm). For good outcome patient #19, PEZs from all three classifiers (i.e., SF, RF, and CF) had high overlap with the resection (> 85%), were focal (< 23 mm), and close to the resection (< 4 mm) (Fig. 5).

Fig. 5.

Case studies of good and poor outcome patients. a Predicted epileptogenic zone (PEZ) from spike (SF, purple), ripple (RF, blue), and combined (CF, red) feature-based Random Forest classifiers in two good outcome patients (#6 and #19) and two poor outcome patients (#37 and #45). Green areas in the figures represent resection and white circles represent contacts.

In contrast, for the poor outcome patient #37, the PEZs from all three classifiers showed low overlap with resection (O_RES < ) and low focality (>31 mm). Finally, for poor outcome patient #44, although the PEZs were focal (<20 mm), they showed a low overlap with the resection (O_RES < ) (Fig. 5).

Discussion

In this study, we introduce a novel supervised ML-based framework that predicts epileptogenic contacts and the surgical outcome in patients with DRE by extracting and combining features of interictal biomarkers from iEEG data in a fully automated fashion. Our framework detects spikes and ripples, extracts relevant temporal, spectral, and spatial features, and employs Random Forest classifiers to identify the PEZ. We evaluated the performance of both the individual features and Random Forest classifiers trained on SF, RF, and CF with RES as target. Our findings indicate that the SF and CF classifiers not only outperformed the models based on individual feature sets but also yielded a focal PEZ with high overlap and proximity to the resection in patients with good surgical outcome. Moreover, resection of the PEZ, defined by the CF, achieved an accuracy of ~ 80% in predicting surgical outcome, while resection of the clinically defined SOZ showed limited predictive value. This discrepancy likely reflects fundamental limitations of the SOZ concept in sEEG and depth-electrode-based evaluations, where seizure onset may be rapidly distributed across mesial and lateral temporal structures⁷⁷, or reflect only the early propagation rather than the full epileptogenic network^78,79. In addition, the SOZ is constrained by electrode coverage of the EZ, which may incompletely capture deeper or spatially remote regions contributing to seizure generation. As a result, SOZ localization alone may fail to identify all tissue necessary for resection and achievement of seizure freedom. Our framework generalizes across different implantation types and enables fully automated localization of the EZ from interictal iEEG data, substantially reducing the need for manual identification and evaluation of epileptogenic brain regions.

Spike spatial distribution plays a key role in defining the EZ

Interictal biomarkers such as spikes and ripples, as well as their features including event rate, power, spike and ripple co-occurrence and their onsets, have been previously associated with the EZ^{11,39,58,59,80,81}. Our results were consistent with these previous findings where several spike and ripple features had higher values within the resection (compared to outside) for good outcome patients (Fig. 2a). Notably, some features also showed elevated values within resection for patients with poor outcome but with lower effect size (Fig. 2b). These findings indicate that although each feature independently captures distinct aspects of the EZ, resection of brain regions based on individual features alone does not necessarily lead to seizure freedom. Upon ranking feature importance, distance-based spike features emerged as the most informative (showing higher effect sizes) than other features in distinguishing epileptogenic from non-epileptogenic contacts (Fig. 2b). This notion was further validated by the epileptogenic contact prediction analysis, where performance metrics showed that the ED-SO, ED-HSR, and ED-HSP achieved higher AUC-ROC, AUC-PR, and values compared to the other features (Fig. 2b). This aligns with a recent study showing that the spatial distribution of biomarkers plays a key role in defining the EZ⁴⁷.

Combined features define the EZ better than individual features

While each feature captures a distinct characteristic of the epileptogenic tissue, no single measure fully represents the complex spatial, spectral, and temporal dynamics of the EZ. A substantial body of literature suggests that ML-based integration of multiple electrophysiological biomarkers yields a more accurate localization of the epileptogenic contacts compared to individual features alone^{52,53,56,57,60,82,83}. Such integration facilitates the discrimination between epileptogenic and non-epileptogenic contacts, even when measurements are incomplete or noisy. Our findings support this notion, as SF and CF classifiers outperformed individual features when predicting the epileptogenic and non-epileptogenic contacts. Specifically, SF and CF outperformed all individual features in terms of AUC-ROC and AUC-PR (except for the ED-HSR, ED-HSP, and ED-SO) (Fig. 3a). Moreover, SF and CF exhibited higher than all individual features, except for the SRC (Fig. 3b). Overall, our results highlight the added value of integrating multiple spatial, temporal, and spectral interictal features to improve the prediction of PEZ and support the notion that ML classifiers can capture nonlinear feature relationships of different epileptogenic activity more effectively than single-feature approaches.

Ripple features provide modest improvement over spike features

The Random Forest classifiers trained using the CF slightly outperformed those trained on SF, suggesting that spike features capture most of the relevant epileptogenic activity (Fig. 3). While combining spike and ripple features provided a more accurate representation of the PEZ than any individual feature alone, our results indicate that ripple features provided only a modest improvement over spike-based features. This may be due to the fact that not all ripples are pathological, underscoring the importance of distinguishing pathological from physiological ripples⁶⁹, which also occur in non-epileptogenic brain regions and during normal brain fucntion^84,85. Although efforts have been made to address this distinction suggesting that pathological HFOs tend to co-occur with spikes and exhibit shorter durations and lower power and amplitude^22,72,84,86, other studies reported the opposing findings^87,88. Collectively, our findings suggest that, although integrating multiple biomarkers can offer incremental gains, spike features remain the dominant contributor to accurate identification of the PEZ based on interictal data. These findings are consistent with a recent study showing that low-frequency features alone were sufficient for the accurate identification of epileptogenic contacts and prediction of surgical outcome, with high-frequency features providing little or no additional value⁵³.

Combined features predict outcome with higher accuracy than individual features

Previous studies have reported strong associations between interictal biomarkers and postsurgical outcome^{11,22,36,80,81,89–92}. These results indicate that, in patients with focal epilepsy, seizure-generating regions form a distinct pathological network isolated from surrounding brain tissue. When these pathological regions are completely resected, patients are more likely to achieve postoperative seizure freedom. Indeed, the spatial concordance between the identified epileptogenic regions and the resected zone has previously been used as a reliable predictor of surgical success^{10,36,39,80,91}. Our findings support this notion since we observed a decrease in the as well as increased in poor outcome patients (Fig. 4). These findings may indicate the presence of other epileptogenic regions distant from the resection, which may not have been resected due to overlap of the EZ with eloquent areas or could have been entirely missed [e.g., patient #37 and #44 (Fig. 5)]. Upon assessing the ability of individual features to predict outcome, we found that almost all features were predictive, except for the SPI, RR, and RPI (Fig. 2b). All Random Forest classifiers were predictive of outcome (Fig. 4b). When comparing outcome predictive performance across all individual features and classifiers, CF outperformed many of the features but not SR, ED-HSR, ED-HSP, ED-SO, ED-HRP, SF, and RF. These findings suggest that outcome prediction remains feasible using individual features alone, and that combining features does not necessarily improve performance, as many individual features are already informative (Fig. 4c).

Justification of features selection

The selected features were designed to capture complementary temporal, spectral, and spatial characteristics of interictal spikes and ripples, known to reflect epileptogenic tissue (Table 2). Spike and ripple rates (i.e. SR and RR, respectively) and power (i.e. SP and RP, respectively) quantify local excitability and pathological high-frequency activity^93,94, while spike and ripple propagation indices (i.e. SPI and RPI, respectively) characterize onset and spread channels¹¹. The spatial distribution of interictal epileptiform events has been shown to reflect underlying epileptogenic organization and network spread, with events clustering around putative onset regions and dispersing along pathological propagation pathways⁴⁷. We used several distance-based measures to quantify the spatial distribution of epileptiform events relative to the contact with the highest spike and ripple rate (i.e. ED-HSR and ED-HRR, respectively), the highest spike and ripple power (i.e. ED-HSP and ED-HRP, respectively), and the spike and ripple onset (i.e. ED-SO and ED-RO, respectively) contacts. In addition, we included the spike and ripple co-occurrence (SRC) to quantify pathological coupling between spikes and ripples, which reflects abnormal local network synchronization and has been strongly associated with epileptogenicity^11,22. Together, these features integrate local and network-level information, providing a physiologically grounded and interpretable representation for automated prediction of epileptic contacts.

Limitations

Despite these promising results, several limitations should be acknowledged. Our retrospective study analyzed iEEG data from a cohort of DRE patients obtained from two tertiary centers and heterogeneous pathologies. Because the models were trained using data from good outcome patients, future studies from larger, multicenter cohorts are needed to evaluate its clinical applicability, particularly within more homogeneous patient populations.

Resection volumes from good outcome patients are often used as a gold standard for the EZ, despite the possibility that they include non-epileptogenic tissue^1,95. This may result in either missing critical epileptogenic regions or defining areas larger than the actual EZ. Because ML model performance depends strongly on the accuracy of the ground truth used, improved approximations of the EZ are needed to provide more reliable targets for ML training. Moreover, iEEG is spatially limited since it is based on a subjective hypothesis agreed upon during presurgical evaluation. Consequently, the EZ was likely sampled well in good outcome patients but may have been missed in poor outcome patients. Advances in whole brain iEEG implantations may overcome these drawbacks. Furthermore, future studies on longer data segments may be conducted to enhance the accuracy of PEZ in estimating the EZ. Finally, the retrospective nature of the study warrants further validation in larger, prospective cohorts to confirm the robustness and generalizability of these findings.

Conclusion

In this study, we systematically evaluated the predictive power of individual features and combined feature-based classifiers for localizing the epileptogenic contacts and predicting surgical outcome in patients with DRE. Our results demonstrate that Random Forest classifiers outperform individual features, providing more focal and accurate identification of the epileptogenic contacts. While spike features remain highly informative for outcome prediction, the integration of ripple features offers a modest additional advantage. Importantly, these findings support the use of short segments of interictal data to approximate the EZ and inform resection planning, potentially reducing reliance on ictal recordings. Future prospective studies with larger cohorts are necessary to validate and refine this approach for clinical translation.

Author contributions

H.P., S.J., and C.P. conceptualized the project and experimental design. H.P. and S.J. developed the mathematical formulation and the codes of the proposed framework. H.P., S.J., S.P., D.H., M.B., J.W., E.T., J.M., J.B., S.S., P.P., and C.P., contributed to the acquisition and analysis of data. H.P. and S.J. participated in the development of various mathematical formulations, statistical analysis, and representation of results of the study. H.P., S.J., and C.P. contributed to drafting the text and preparing figures. All authors have read and approved the manuscript.

Funding

This research was supported by the National Institute of Neurological Disorders & Stroke (R01NS104116-01A1 and R01NS134944, PI: C. Papadelis).

Data availability

The data are available from the corresponding author upon request.

Code availability

The code is available from the corresponding author upon request.

Declarations

Competing interests

The authors declare no competing interests.