Ambulatory seizure forecasting with a wrist-worn device using long-short term memory deep learning

Mona Nasseri,Nicholas M Gregg,Mark P Richardson,Matthias Dümpelmann,Dean R Freestone,Pedro F Viana,Boney Joseph,Tal Pal Attia,Gregory Worrell,Benjamin H Brinkmann,Ewan S Nurse

doi:10.1038/s41598-021-01449-2

Abstract

The ability to forecast seizures minutes to hours in advance of an event has been verified using invasive EEG devices, but has not been previously demonstrated using noninvasive wearable devices over long durations in an ambulatory setting. In this study we developed a seizure forecasting system with a long short-term memory (LSTM) recurrent neural network (RNN) algorithm, using a noninvasive wrist-worn research-grade physiological sensor device, and tested the system in patients with epilepsy in the field, with concurrent invasive EEG confirmation of seizures via an implanted recording device. The system achieved forecasting performance significantly better than a random predictor for 5 of 6 patients studied, with mean AUC-ROC of 0.80 (range 0.72–0.92). These results provide the first clear evidence that direct seizure forecasts are possible using wearable devices in the ambulatory setting for many patients with epilepsy.

Highlights

The ability to forecast seizures minutes to hours in advance of an event has been verified using invasive EEG devices, but has not been previously demonstrated using noninvasive wearable devices over long durations in an ambulatory setting
Invasive devices may not be acceptable for some patients with epilepsy, and no clinically available invasive device currently has the capability to sample and telemeter data needed for seizure forecasting
The power and capability of deep learning algorithms trained on very large datasets hold promise to enable applications not previously believed possible, and may open the door to seizure forecasting with noninvasive sampling devices

Summary

Introduction

The ability to forecast seizures minutes to hours in advance of an event has been verified using invasive EEG devices, but has not been previously demonstrated using noninvasive wearable devices over long durations in an ambulatory setting. Deep learning approaches have shown promising performance for a variety of difficult applications[15], including seizure forecasting[7] In particular these “end-to-end learning” methods are attractive for seizure forecasting given the challenges of identifying salient features in ultra-long term time-series data, and the heterogeneity in time series data characteristics between different patients. Performing device studies on in-hospital patients with concurrent video-EEG validation is logistically feasible, but such studies are expensive, and limited in duration, and restrict normal daily activities which could produce false alarms, such as exercise, brushing teeth, or other activities Because of these challenges an ILAE-IFCN working group recently published g uidelines[17] for seizure detection studies with non-invasive wearable devices, but few studies achieve phase 3–4 evidence in an ambulatory setting[18]. In studies of seizure forecasting it is imperative that ambulatory data including the full range of normal activities be included in the training, testing, and validation sets

Objectives

Methods

Results

Discussion

Conclusion