Machine learning from wristband sensor data for wearable, noninvasive seizure forecasting

TLDR

Recent seizure‑forecasting studies rely on invasive, patient‑specific setups, whereas wristband sensors can continuously capture autonomic and movement signals, offering a noninvasive, low‑stigma alternative. The study aims to develop a broadly applicable, noninvasive seizure‑forecasting method that requires minimal prior tuning. Deep learning models were trained on multimodal wristband data from 69 epilepsy patients (over 2311 h and 452 seizures) to evaluate seizure‑risk prediction. The models achieved better‑than‑chance forecasting in 43 % of patients, with peak performance when all sensor modalities were used, and performance improved with larger training sets, demonstrating feasible, patient‑independent seizure risk assessment from wearables.

Abstract

Abstract Objective Seizure forecasting may provide patients with timely warnings to adapt their daily activities and help clinicians deliver more objective, personalized treatments. Although recent work has convincingly demonstrated that seizure risk assessment is in principle possible, these early approaches relied largely on complex, often invasive setups including intracranial electrocorticography, implanted devices, and multichannel electroencephalography, and required patient‐specific adaptation or learning to perform optimally, all of which limit translation to broad clinical application. To facilitate broader adaptation of seizure forecasting in clinical practice, noninvasive, easily applicable techniques that reliably assess seizure risk without much prior tuning are crucial. Wristbands that continuously record physiological parameters, including electrodermal activity, body temperature, blood volume pulse, and actigraphy, may afford monitoring of autonomous nervous system function and movement relevant for such a task, hence minimizing potential complications associated with invasive monitoring and avoiding stigma associated with bulky external monitoring devices on the head. Methods Here, we applied deep learning on multimodal wristband sensor data from 69 patients with epilepsy (total duration > 2311 hours, 452 seizures) to assess its capability to forecast seizures in a statistically significant way. Results Using a leave‐one‐subject‐out cross‐validation approach, we identified better‐than‐chance predictability in 43% of the patients. Time‐matched seizure surrogate data analyses indicated forecasting not to be driven simply by time of day or vigilance state. Prediction performance peaked when all sensor modalities were used, and did not differ between generalized and focal seizure types, but generally increased with the size of the training dataset, indicating potential further improvement with larger datasets in the future. Significance Collectively, these results show that statistically significant seizure risk assessments are feasible from easy‐to‐use, noninvasive wearable devices without the need of patient‐specific training or parameter optimization.