Explainable AI for wearable seizure logging: Impact of data quality, patient age, and antiseizure medication on performance

Abstract

Objective: Objective seizure count estimates are crucial for ambulatory epilepsy management. Wearables have shown promise for the detection of tonic-clonic seizures but may suffer from false alarms and undetected seizures. Seizure signatures recorded by wearables often occur over prolonged periods, including increased levels of electrodermal activity and heart rate long after seizure EEG onset, however, previous detection methods only partially exploited these signatures. Understanding the utility of these prolonged signatures for seizure count estimation and what factors generally determine seizure logging performance, including the role of data quality vs. algorithm performance, is thus crucial for improving wearables-based epilepsy monitoring and determining which patients benefit most from this technology.

Methods: In this retrospective study we examined 76 pediatric epilepsy patients during multiday video-EEG monitoring equipped with a wearable (Empatica E4; records of electrodermal activity, EDA, accelerometry, ACC, heart rate, HR; 1983 h total recording time; 45 tonic-clonic seizures). To log seizures on prolonged data trends, we applied deep learning on continuous overlapping 1-hour segments of multimodal data in a leave-one-subject-out approach. We systematically examined factors influencing logging performance, including patient age, antiseizure medication (ASM) load, seizure type and duration, and data artifacts. To gain insights into algorithm function and feature importance we applied Uniform Manifold Approximation and Projection (UMAP, to represent the separability of learned features) and SHapley Additive exPlanations (SHAP, to represent the most informative data signatures).

Results: Performance for tonic-clonic seizure logging increased systematically with patient age (AUC 0.61 for patients 〈 11 years, AUC 0.77 for patients between 11-15 years, AUC 0.85 for patients 〉 15 years). Across all ages, AUC was 0.75 corresponding to a sensitivity of 0.52 and a false alarm rate of 0.28/24 h. Seizures under high ASM load or with shorter duration were detected worse (P=.025, P=.033, respectively). UMAP visualized discriminatory power at the individual patient level, SHAP analyses identified clonic motor activity and peri/postictal increases in HR and EDA as most informative. In contrast, in missed seizures, these features were absent indicating that recording quality but not the algorithm caused the low sensitivity in these patients.

Significance: Our results demonstrate the utility of prolonged, postictal data segments for seizure logging, contribute to algorithm explainability and point to influencing factors, including high ASM dose and short seizure duration. Collectively, these results may help to identify patients who particularly benefit from such technology.

Keywords: Antiseizure medication; Automatic seizure logging; Autonomous nervous system; Deep learning; Epilepsy; Explainable artificial intelligence; Tonic-clonic seizure; Wearables.