Localising the epileptogenic zone from single-pulse electrical stimulation responses using cross-trial attention

Abstract Background Analysis of SPES responses often relies on averaging repeated stimulation trials to improve signal quality. However, this may obscure clinically relevant trial-to-trial variation. We tested whether explicitly modelling cross-trial dependencies improves localisation of the epileptogenic zone, using concordance with the clinical SOZ as a proxy endpoint, and explored whether resection of model-positive channels is associated with postsurgical seizure freedom. Methods We developed an interleaved Hierarchical Attention Transformer (HAT), a deep learning architecture that models cross-trial and cross-channel dependencies in SPES responses without averaging. We compared it with two baselines averaging responses or trial embeddings. Models were evaluated with patient-held-out, repeated five-fold cross-validation on SPES data from 35 patients. Robustness to reduced trial availability at inference was assessed by restricting test inputs to 1 or 5 trials. Associations with surgical outcome were assessed using AUROC and patient-level tests on the proportion of model-positive channels resected. Results The HAT had higher SOZ concordance than the trial-averaged baseline (AUROC 0.762 vs 0.721; mean paired difference 0.041; one-sided 95% lower confidence bound 0.009; Holm-adjusted p = 0.0197). Performance changed little when inference was restricted to 1 trial. Outcome analyses did not provide statistical evidence that seizure-free patients had a higher proportion of model-positive channels resected (AUROC 0.634; p = 0.102). Conclusions Modelling cross-trial dependencies improved concordance with the SOZ compared with trial-averaged approaches, while remaining robust to reduced trial availability at inference. Associations with postsurgical outcome were inconclusive, consistent with limited sample size and training on SOZ labels rather than outcome-aligned labels. Key points Cross-trial modelling of SPES responses improved concordance with the clinical SOZ compared with trial-averaged baselines. In a 35-patient cohort, SOZ AUROC was higher than a trial-averaged baseline (0.762 vs 0.721). Performance was stable when inference used as few as 1 trial, with the largest drop in the trial-averaged baseline. Resection of model-positive channels did not provide statistical evidence of an association with seizure freedom (AUROC 0.634; one-sided Mann–Whitney p = 0.102).