Multimodal Seizure Detection with Long-term Ambulatory ECG and Accelerometry Data

ObjectiveDetecting epileptic seizures in real-world environments remains challenging, as electroencephalography (EEG) is often impractical in chronic ambulatory monitoring. Heart rate and accelerometry, measurable from wearable devices, provide a less obtrusive alternative. Although some studies explored multimodal wearable-based seizure detection, few have been validated on long-term ambulatory datasets reflecting real-world variability. This study investigated the added value of accelerometry for electrocardiography (ECG)-based seizure detection using a long-term, ambulatory video-EEG dataset. Using an established pseudo-prospective framework, we assessed performance across heart rate variability (HRV), accelerometry (ACC), and combined feature sets, and examined how it varied across patient groups with different seizure types. ApproachRecordings from 78 patients (587 seizures, 385 days) undergoing ambulatory monitoring were analyzed. Time and frequency domain features were extracted from ECG and triaxial ACC. A logistic regression classifier trained on data from 47 patients was evaluated on a hold-out set of 31 patients. Seizure types were derived at the patient level through automated keyword extraction from clinical reports, enabling performance comparison across seizure type groups. Main resultsA combined HRV+ACC model achieved the best performance in leave-one-patient-out validation (Improvement over chance AUC, i.e., {Delta}AUC=0.12), while the HRV-only model generalized best to the hold-out set ({Delta}AUC=0.046). Across feature sets, 74% of patients demonstrated better-than-chance performance. Patients diagnosed with focal epilepsy, and those with substantial seizure-related heart rate increase achieved better performance. Patients with motor seizures showed significantly higher performance with ACC features, while HRV features were more informative for non-motor seizure detection. SignificanceThis study validated the use of ACC and ECG-derived HRV for patient-independent seizure detection in long-term ambulatory data. ACC features were less generalizable overall, but improved performance for a subset of patients. Seizure type analysis revealed complementary strengths of autonomic and movement signals, highlighting the value of incorporating seizure type information into future multimodal systems.