Abstract
Abstract
Accurate localization of the seizure onset zone (SOZ) is essential for planning resection or ablation in patients with focal drug-resistant epilepsy (DRE). However, the true SOZ cannot be directly observed during clinical evaluation, so precise electrode-level annotations are rarely available. To address this limitation, we developed a weakly supervised multiple-instance learning (MIL) framework in which labels were constructed from electrode resection status and postsurgical outcome. A shared gated recurrent unit (GRU) with temporal attention encoded each electrode’s peri-onset neural fragility sequence. And attention-based pooling integrated information across variable numbers of electrodes to estimate whether each electrode group contained SOZ-related evidence. The cohort comprised 42 patients with DRE from three datasets, including 29 with successful and 13 with failed postsurgical outcomes. These patients contributed 136 seizure epochs. The cohort was divided at the patient level into training, validation and held-out test sets containing 25, 8 and 9 patients, respectively, with all data from each patient retained in the same subset. Predictions from non-resected regions were evaluated at the seizure-epoch and patient levels to assess postsurgical outcome discrimination for individual seizures and after aggregation across seizures from the same patient, respectively. At the seizure-epoch level, the proposed framework achieved an area under the receiver operating characteristic (ROC) curve (AUC) of 0.828, with a 95% confidence interval (CI) of 0.511-1.000, and performed comparably to Manifold Oblique Random Forests (MORF). At the patient level, the framework achieved an AUC of 0.944, with a 95% CI of 0.677-1.000, and yielded a numerically higher AUC than MORF. However, DeLong’s test did not show a statistically significant difference between the methods, with a P value of 0.398. These findings support the feasibility of extracting clinically meaningful SOZ-related information from electrode-wise neural fragility sequences without precise electrode-level labels. Because the held-out test set included only nine patients and the CIs were wide, these preliminary findings require validation in larger independent cohorts.