Abstracts

Rationale: The unpredictability of seizures is debilitating and dangerous for people with epilepsy. Accurate seizure forecasters could improve quality of life but must be practical for long-term use. We recently launched a wearable and mobile app to forecast seizure risk using established cycles of seizure likelihood (Figure 1). This study validated forecast accuracy in a prospective, blinded study.

Methods: The seizure forecasting pipeline was deployed on a cloud platform capable of updating and displaying seizure risk in a scalable, performant, real-world mobile application. Individual forecast performance was assessed for a cohort of participants blinded to seizure risk. A subset of users with sufficient data were selected to validate performance of two potential implementations of risk forecasts: a diary-based forecast, using just self-reported seizure times, and a diary+wearable forecast which also included signals recorded from a smartwatch. _x000D_  _x000D_ The diary-based forecaster used self-reported seizure times to detect circadian and multiday seizure cycles. The diary+wearable forecaster augmented these seizure cycles with cycles recorded from resting heart rate. Individuals’ seizure cycles were converted to a continuous score of seizure likelihood (between 0 and 1) and seizure risk (low, medium and high). Forecasts were initially trained on 10 seizures, then iteratively tested, and updated with subsequent reported events.

Results: Seizure risk forecasts were run continuously for 460 mobile app users over a 6-month period (deployed on a secure cloud platform) Forecasts were updated weekly and after each seizure, with update time between 1 - 10s per user._x000D_  _x000D_ The diary-based risk forecaster was assessed for 43 participants (mean test seizures: 123.2). During prospective evaluation 41/43 (95%) of participants had forecasting performance above chance. The mean AUC was 0.69 [IQR: 0.66 - 0.74] with an average sensitivity of 52% and average specificity of 75%._x000D_  _x000D_ The diary+wearable risk forecaster was assessed for 27 participants (mean test seizures: 155.8). During prospective evaluation 26 of 27 (96%) of participants had forecasting performance above chance. The mean AUC was 0.67 [IQR: 0.65 - 0.72] with an average sensitivity of 48% and average specificity of 76%._x000D_  _x000D_ For both forecasting methods, higher AUC was significantly correlated (p< 0.05 using linear regression test) with longer recording durations (Figure 2), suggesting that forecast performance improves over time.