A multimodal AI system for out-of-distribution generalization of seizure detection

Abstract

Epilepsy is one of the most common severe neurological disorders worldwide. The International League Against Epilepsy (ILAE) define epilepsy as a brain disorder that generates (1) two unprovoked seizures more than 24 hrs apart, or (2) one unprovoked seizure with at least 60% risk of recurrence over the next ten years. Complete remission has been defined as ten years seizure free with the last five years medication free. This requires a cost-effective ambulatory ultra-long term out-patient monitoring solution. The common practice of self-reporting is inaccurate. Applying artificial intelligence (AI) to scalp electroencephalogram (EEG) interpretation is becoming increasingly common, but other data modalities such as electrocardiograms (ECGs) are simpler to collect and often recorded simultaneously with EEG. Both recordings contain biomarkers in the detection of seizures. Here, we propose a state-of-the-art performing AI system that combines EEG and ECG for seizure detection, tested on clinical data with early evidence demonstrating generalization across hospitals. The model was trained and validated on the publicly available Temple University Hospital (TUH) dataset. To evaluate performance in a clinical setting, we conducted nonpatient-specific inference-only tests on three out-of-distribution datasets, including EPILEPSIAE (30 patients) and the Royal Prince Alfred Hospital (RPAH) in Sydney, Australia (31 patients shortlisted by neurologists and 30 randomly selected). Across all datasets, our multimodal approach improves the area under the receiver operating characteristic curve (AUC-ROC) by an average margin of 6.71% and 14.42% for prior state-of-the-art approaches using EEG and ECG alone, respectively. Our models state-of-the-art performance and robustness to out-ofdistribution datasets can improve the accuracy and efficiency of epilepsy diagnoses.