Abstracts

Rationale: Seizure forecasting has been established using intracranial EEG, but invasive devices are not appropriate for everyone. Our consortium has evaluated wearable and subcutaneous EEG devices with the potential to provide continuous monitoring in people living with epilepsy and has developed analytical methods to detect and forecast seizures. 

Methods: Patients were recruited for ultra-long term monitoring with a wearable device (Empatica E4, Fitbit Charge HR, or Fitbit Inspire) and concurrent ambulatory EEG monitoring (UNEEG SubQ, EpiMinder Subscalp, NeuroPace RNS) at three sites. Wearable and EEG data from enrolled patients were recorded for 8 months or more. Self-reported electronic seizure diaries and mood and symptom surveys were recorded by participants. Recorded data were analyzed to identify circadian and multi-day cycles, and machine learning methods were used to forecast and detect seizures. 

Results: Thirty-nine patients with epilepsy have recorded over 12,500 days (33.7 years) of ambulatory wearable and EEG data, including over 1700 seizures. Nine patients left the study before completion due to device malfunctions, complications, poor adherence, poor data qualtiy, or unanticipated seizure freedom. Results from preliminary analysis of this data are as follows:  _x000D_ _x000D_ - Heart rate circadian and multi-day cycles measured using the Fitbit device were significantly phase-locked with self-reported seizure likelihood in 10 of 19 patients. _x000D_ - Cyclical patterns in tonic and phasic electrodermal activity, heart rate, temperature, and actigraphy are significantly phase-locked with iEEG-confirmed seizures in 10 patients with at least 8 months of recorded data with the Empatica E4. _x000D_ - Seizure forecasting significantly better than chance in 5 of 6 patients using the wrist-worn Empatica E4 device for more than 6 months, with iEEG confirmation of events. _x000D_ - Seizure forecasting using the Fitbit device in 11 patients with at least 6 months of of data and 20 seizures provided a mean AUC of 0.74 when assessed hourly referenced against an electronic seizure diary. _x000D_ - Seizure forecasting using the UNEEG subscalp EEG system significantly greater than chance using a machine learning algorithm with an intra-subject training/testing approach in 5 of 6 patients with at least four seizures recorded. A cross-patient machine learning algorithm was able to forecast seizures greater than chance in 4 of 6 patients using a cross-validation approach.  _x000D_ - Forecasting in one patient has been demonstrated using the EpiMinder subscalp EEG system over 6 months with 83% sensitivity and 26% of time in a high risk forecast alert. _x000D_ - A seizure diary app with the ability to forecast seizure risk based on patients’ reported events is now freely available (https://seermedical.com/health/).  _x000D_ - A data science competition on the eval.ai platform runs from June to October to crowdsource seizure forecasting from Empatica E4 data  with EEG confirmation of seizures.  _x000D_ _x000D_ Conclusions: Forecasting seizures using long-term cycles, wearable devices, and/or subcutaneous EEG is possible, but progress is needed to improve performance for routine use.  

Funding: The Epilepsy Foundation of America My Seizure Gauge grant