Interictal neural fragility predicts seizure onset zone and surgical outcomes in drug-resistant epilepsy

SEEG recording

SEEG data were recorded using the Xltek acquisition system (Natus EEG System pk1171 and Neuracle HEEG NSH0256) at sampling rates of 4 or 2 kHz. Electrode placement was determined by the clinical team for each patient. For analysis, a randomly selected interictal snapshot (average duration of 10 min) was used. Interictal sampling was performed at least 2 h apart from seizure events without specific selection criteria, such as the presence or absence of epileptiform activity.

Clinical delineation of epileptogenic zones

The clinical team independently hypothesized the EZ based on non-invasive and invasive data collected during preoperative evaluation. The SOZE was defined as the SEEG channels exhibiting the earliest electrophysiological changes at seizure onset, typically characterized by low-voltage fast activity. The EZ (SOZC) was defined as the anatomical regions requiring ablation or disconnection treatment (RFTC, LITT, surgical resection), including SEEG channels at the seizure onset and those involved in early seizure propagation. Notably, surgical planning was based on the assessment of the EZ, resulting in considerable overlap between SOZC and the actual treatment area.

Clinical classification of surgical outcomes

Surgical outcomes were classified by epilepsy specialists at each center according to the Engel Surgical Outcome Scale and the International League Against Epilepsy (ILAE) classification system. A successful outcome was defined as freedom from disabling seizures for 12 months or longer post-surgery (Engel Class I and ILAE 1–2). Conversely, a failed outcome was defined as the persistence of disabling seizures (Engel Class II–IV and ILAE 3–6) over the same period. Of the 30 patients, 22 achieved seizure freedom (successful outcome), while eight continued to experience seizures (failed outcome). Current clinical experience indicates that MRI-visible lesions are associated with higher success rates, whereas non-lesional, extratemporal, or multifocal epilepsy patients have lower seizure-free rates (Li et al., 1999; Sheikh et al., 2019; Elsharkawy et al., 2008). To categorize clinical complexity (CC), patients were classified as follows: CC I, single lesion (MRI-visible); CC II, mesial or non-mesial temporal lobe epilepsy; CC III, focal or multifocal extratemporal epilepsy; CC IV, multiple neuroimaging abnormalities potentially distributed across multiple brain regions.

Neural fragility network construction

Neural fragility is predicated on the hypothesis that focal epileptic seizures originate from a small subset of fragile nodes, rendering the cortical epilepsy network on the verge of instability. During interictal or preictal periods, the activity recorded from each channel fluctuates around a baseline value. If the network is “balanced,” it responds transiently to perturbations or stimuli, returning to baseline. In contrast, during ictal events, iEEG data demonstrate increased amplitude, oscillations, and propagation across the brain. This study modeled susceptibility to seizure events within the interictal SEEG network. Raw SEEG signals were initially downsampled to 500 Hz and then bandpass-filtered between 0.1 and 30 Hz. After identifying and excluding bad channels, the filtered SEEG signals were analyzed using a modified neural fragility method (Li et al., 2021). Unlike the original neural fragility method, dynamic mode decomposition (DMD) techniques (Brunton et al., 2021) were employed to estimate the linear operator A (state transition matrix), transitioning the state X(t) at time window t to state X(t + 1) at time window t + 1, as described by X(t + 1) = AX(t) + ϵ(t). Here, X(t) represents the SEEG signal within the current time window (window length = 250 ms, overlap = 125 ms), ϵ(t) represents the For this analysis, the Python implementation of DMD (PyDMD v0.4.0.post2301) was utilized with a regularization parameter of 0.05. Subsequently, the minimum 2-norm perturbation matrix for each node in the network was calculated to shift the network from a stable to an unstable state (i.e., triggering a seizure). These minimum perturbations Δ (defined as neural fragility) were then mapped across time and space, generating neural fragility heat maps (y-axis representing channel indices, x-axis representing time). These maps provide a visual representation of the fragility distribution corresponding to the interictal SOZ and other channels, where nodes with high fragility values indicate regions with a greater predisposition to seizures. Fragility heat maps were computed for each individual patient and simulated patient datasets (Fig. 1).

Summary analysis of patients

Initially, we analyzed the differences in neural fragility distribution between SOZ and NSOZ (channels outside the SOZ). Data from all patients were aggregated and grouped by surgical outcome (successful vs. failed). To further investigate whether SOZ fragility is significantly higher than NSOZ fragility within different groups, paired t-tests were employed, which is appropriate for comparing two related samples within the same group (e.g., SOZ and NSOZ fragility values in the same patient) to determine if SOZ fragility is significantly elevated. If SOZ fragility is significantly higher than NSOZ in the successful outcome group, it suggests that quantitative analysis of interictal fragility may be valuable for predicting the SOZ. Additionally, we examined the distribution within the seizure snapshots of each patient, depicting the differential distribution of neural fragility between SOZ and NSOZ in individual patients. The hypothesis is that patients with successful outcomes should maintain a higher level of neural fragility in the interictal SOZ.

Validation of interictal fragility features for predicting the epileptogenic zone

To further explore the utility of fragility values derived from interictal SEEG data in localizing the SOZ, we integrated the spatiotemporal fragility heat maps generated from the interictal SEEG snapshots of each patient with a random forest model to establish a binary classification prediction model (SOZ/NSOZ) for individual channels. An effective biomarker should show high concordance with treatment targets in patients with favorable outcomes. Therefore, using the random forest model, we performed binary classification of each patient’s electrode channels (1-SOZ; 0-NSOZ) and compared these results with the clinically identified SOZE and the treatment target SOZC, calculating model prediction precision and recall rates. Subsequently, we constructed a logistic regression model incorporating treatment modality, MRI findings, CC classification, and surgical outcomes to validate the predictive efficacy of fragility features. This involved: (i) performing comparative analyses to validate the predictive performance of neural fragility against clinically identified SOZE (based on EEG) in different outcome groups; (ii) comparing the concordance between model-predicted epileptogenic zones and clinical treatment targets (SOZC) across different surgical outcomes, with further elucidation of the clinical significance of fragility features through specific cases; (iii) analyzing the impact of other clinical variables on the predictive accuracy of fragility features.

High concordance between model predictions and SOZE

Figure 3 illustrates the concordance between the fragility-predicted SOZ and the clinically annotated SOZE (identified as the most likely EZ based on EEG interpretation) across different surgical outcome groups. In the successful outcome group (Engel score = 1), the model achieved an average precision of 0.69 and a recall of 0.93 for SOZE. Conversely, in the unsuccessful outcome group, the model demonstrated a precision of 0.86 and a recall of 0.99. Statistical analysis using t-tests indicated that the difference in precision between the two groups was not statistically significant (p = 0.219), while the difference in recall approached statistical significance (p = 0.057).

These findings suggest that the interictal fragility predictive model captures a significant portion of the information utilized in SEEG interpretation. However, this does not necessarily translate into improved surgical outcomes, as successful surgery may also depend on other factors, including the patient’s specific pathological and physiological characteristics, as well as the precision of the surgical procedure.

Model concordance with SOZC across different surgical outcomes

The distribution of the model’s concordance metrics (precision and recall) for predicting the SOZ relative to the clinical therapeutic target SOZC across different Engel score groups is shown in Fig. 4. For patients with an Engel score of 1 (successful surgery group), the model achieved a recall rate for SOZC close to 100%. The precision distribution for SOZC was also relatively high (median = 1.00). In patients with an Engel score of 2, although data were limited, both metrics (SOZC recall and precision) were relatively high, with quartiles for recall at 0.97 and precision at 0.86. In the Engel score 3 group (moderate surgical outcome), the median recall rate was 0.8793, and the precision distribution was concentrated around 1.00, indicating that the model performed well in identifying SOZC but with some variability. For patients with an Engel score of 4 (failed surgical outcome), the recall rate for SOZC showed considerable variability, with a minimum value of 0.13 and quartiles ranging from 0.44 to 1.00. This suggests that in patients with failed surgical outcomes, the model’s recall rate for SOZC varied significantly, with lower recall rates observed in some patients. Similarly, precision ranged from 0.24 to 1.00, indicating a high degree of variability in the model’s precision for SOZC in the failed surgery group.

Overall, the model demonstrated high precision across different prognostic groups, with higher precision in groups with better outcomes (Engel scores closer to 1), indicating more accurate predictions and a lower false-positive rate. Higher recall rates were also observed in groups with better outcomes, suggesting that the model effectively predicts therapeutic targets in patients with more favorable prognoses.

Interictal fragility as a reliable predictor of clinical treatment outcomes

Among the 30 patients, 11 underwent resective surgery, while 19 did not receive resection and were treated with either RFTC/LITT ablation or neurostimulation therapy. In patients with successful outcomes (12/19, Engel class 1), the model demonstrated a significantly higher concordance between the predicted SOZ and the therapeutic target compared to those with failed outcomes (7/19, Engel classes 2–4). Figure 5A illustrates the relationship between the concordance metrics (recall and precision) of the model-predicted SOZ with SOZE and SOZC and the treatment outcomes (1 = success, 0 = failure). A clear positive correlation can be observed, where higher concordance metrics between the model-predicted SOZ and the clinical therapeutic target SOZC are associated with a higher likelihood of successful surgery/RFTC outcomes.

Logistic regression analysis identified the model’s recall (SOZC recall: regression coefficient = −2.66, p = 0.001) and precision (SOZC precision: regression coefficient = −3.51, p = 0.002) for the therapeutic target as two significant predictors of surgical outcomes (Fig. 5B). These metrics correspond to the model’s ability to capture the actual therapeutic target and the accuracy of predicting the therapeutic target, respectively. The recall and precision for SOZC were significantly negatively correlated with the Engel score, indicating that a higher concordance between the model-predicted SOZ and the therapeutic target SOZC is associated with a lower Engel score (higher surgical success rate). This suggests that the accurate identification of the therapeutic target by the model (high recall and balanced precision) is a key driver of surgical success. RFTC (regression coefficient = 1.20, p-value = 0.03) and the type of Surgery (regression coefficient = 0.66, p-value = 0.04) had a significant positive impact on the Engel score, indicating that different treatment modalities may influence surgical outcomes. While recallE and precisionE were positively correlated with the Engel score, their p-values were greater than 0.05, indicating no significant statistical relevance. MRI and CC treatment modalities did not have a significant impact on the prognostic outcome (p-values > 0.05), suggesting that these variables have a weaker influence on clinical prognosis.