Abstracts

Rationale: Epileptic seizures may be influenced by a range of patient-specific, environmental and physiological factors. However, it has recently become clear that there are some generalisable patterns that modulate seizure onset. For instance, a majority of people show a circadian rhythm in their seizure times. Most people (over 50%) also show a slower, multiday rhythm of one week or longer (Karoly et al. 2018). Seizure cycles can be captured from a range of recording modalities, including self-reported seizure diaries. Methods: This study presents a mobile app and web-based framework to generate real-time forecasts that combine data recorded from wearable devices and user inputs from a mobile app. Seizure forecasts based on circadian ('fast') and multiday ('slow') seizure cycles were tested in a pseudo-prospective manner using data recorded from a mobile seizure tracking app (Seer Medical) from 53 users with at least 50 reported seizures (mean of 188.2 seizures). Individuals' strongest circadian and multiday cycles were estimated from their reported seizure times using the mean resultant vector and corresponding significance tests (Karoly et al. 2018). The phase of the fast and slow cycles was computed for the entire recording duration then broken into 20 equally spaced bins. The probability of future seizures with respect to each phase bin was then calculated based on a histogram of previous seizure times. The final probability for the combined slow and fast rhythms was obtained as the product of the log-odds of each probability. High and low-risk warning thresholds were computed using a brute force optimization that maximized the time spent in low risk periods and number of seizures classified in high risk periods (Fig 1). Forecasts were also generated using only the self-reported events from the NeuroVista dataset, an existing database of long-term, individual seizure records that has been widely used to develop forecasting algorithms (Cook et al. 2013). The NeuroVista dataset enabled a comparison of non-invasive forecasting performance to state-of-the-art seizure prediction. Results: Fig. 2 shows forecasting performance for all users. Non-invasive forecasts allowed mobile app users to spend an average of over half (54%) of their time in a low-risk state, with 21% of their time in a high-risk warning. On average, 55% of seizures occurred in the high-risk state (with less than 15% of seizures in the low-risk state). Non-invasive forecasting performance was comparable to a recent Kaggle data science competition using intracranial EEG from three NeuroVista patients to forecast seizures (Kuhlmann et al. 2018). The best published forecast using intracranial EEG from 10 NeuroVista patients (Kiral-Kornek et al. 2017) gave a mean sensitivity of 69% and mean time in warning of 27%. Conclusions: Warnings based on seizure cycles enabled app users to spend less time in high-risk warning than forecasts using invasively recorded EEG. Furthermore, forecast sensitivity for many app users was comparable to forecast sensitivity based on invasive recordings. Importantly, non-invasive forecasts enabled users to spend over half their time in a low-risk state. These results provide a compelling demonstration of the potential for mobile apps to provide personalized information about users' seizure likelihood. In future, mobile diaries may be combined with other non-invasive measures to provide a more accurate forecast of seizure likelihood. This information can give people with epilepsy greater confidence to go about their everyday lives. Funding: No funding