Generalisability of epileptiform patterns across time and patients

Abstract Complexity of epileptogenic zone (EZ) localisation contributes to failure of surgical resection to produce seizure freedom. This is to some extent a result of distinct patterns of epileptiform activity between (i.e., interictal) and during seizures (i.e., ictal) and their diversity across patients. This often leads to suboptimal localisation based on inspection of electroencephalography (EEG) features. We asked two open questions. First, whether neural signal reflecting epileptogenicity would be generalisable from interictal to ictal time window within each patient. This would be critical for patients who are monitored in hospital without having a seizure to help with EZ localisation, and more generally for understanding the predictive power of resting state (interictal) EEG data in determining EZ. Second, whether epileptiform patterns would generalise across patients, and if so, which aspects of those patterns are the most generalisable. We used an intracranial EEG dataset that included fifty-five patients with lesional and non-lesional pathology, who had subsequently undergone cortical resection in frontal or temporal lobe with different levels of seizure freedom. We extracted a large set of simple to complex features from stereo-EEG (SEEG) and electrocorticographic (ECoG) neural signals recorded during interictal and ictal time windows. We fed those features to decision tree classifiers for EZ localisation and to quantify the diversity of ictal and interictal epileptiform patterns through a cross-time and cross-patient generalisation procedure. We observed significant evidence ( Bayes factor >> 10) for generalisability of patterns from interictal to ictal time windows across patients, which were dominantly reflected in signal power and high-frequency network-based features. Majority of patients showed consistent patterns of epileptogenicity across interictal and ictal time windows, reflected in above-chance area-under-curve (mean AUC = 0.6). We observed significant evidence ( Bayes factor >> 10) that signal features of epileptogenic regions could generalise across patients in both interictal and ictal time windows with significant evidence for higher generalisability in ictal than interictal time window (mean AUC 0.75 vs. 0.59; Bayes factor >> 10). While signal power and moment features were the most contributory to the cross-patient generalisation in the interictal window, signal complexity features were the most contributory in the ictal window. These results provide new insights about features of epileptic neural activity that generalise across interictal-ictal time windows and patients, which can have implications for both qualitative and quantitative EZ localisation. The explainable machine-learning pipeline developed here can guide future developments in epilepsy investigations.