Accepted for/Published in: JMIR mHealth and uHealth
Date Submitted: Feb 2, 2021
Date Accepted: Sep 3, 2021
Date Submitted to PubMed: Nov 22, 2021
Warning: This is an author submission that is not peer-reviewed or edited. Preprints - unless they show as "accepted" - should not be relied on to guide clinical practice or health-related behavior and should not be reported in news media as established information.
Detecting Tonic-Clonic Seizures in Multimodal Biosignal Data from Wearables
ABSTRACT
Background:
Video-EEG recordings, routinely employed in epilepsy monitoring units (EMU), are the gold standard in monitoring epileptic seizures. Monitoring is however also needed in the day to day lives of people with epilepsy, where video-EEG is not feasible. Wearables could fill that gap, providing patients with an accurate log of their seizures.
Objective:
While there are already systems available that give promising results for the detection of tonic-clonic seizures (TCS), research in this area is often limited to detection from one biosignal modality or only during the night when the patient is in bed. In this study, we provide evidence that supervised machine learning is able to detect TCS from multimodal data in a new data set during day- and night-time.
Methods:
An extensive data set of biosignals from a multimodal watch worn by people with epilepsy was recorded during their stay in the EMU at two European clinical sites. From a larger data set of 243 enrolled participants, those that had data recorded during TCS were selected, amounting to ten participants with 21 TCS. Accelerometry (ACC) and electrodermal activity (EDA) recorded by the wearable were used for analysis, and seizure manifestation was annotated in detail by clinical experts. Ten ACC and three EDA features were calculated for sliding windows of variable size across the data. A gradient tree boosting algorithm was used for seizure detection, and the optimal parameter combination was determined in a leave-one-participant-out (LOPO) cross-validation on a training set of ten seizures from eight participants. The model was then evaluated on an out-of-sample test set of eleven seizures from the remaining two participants. To assess specificity, we additionally analyzed data from up to 29 participants without TCS during model evaluation.
Results:
In the LOPO cross-validation, the model optimized for sensitivity could detect all ten seizures with a false alarm rate (FAR) of 0.46 per day, in 17.3 days of data. In a test set of eleven out-of-sample TCS, amounting to 8.3 days of data, the model could detect ten seizures and produced no false positives (FP). Increasing the test set to include data from 28 more participants without additional TCS resulted in a FAR of 0.19 per day in 78 days of wearable data.
Conclusions:
We here show that a gradient tree boosting machine can robustly detect TCS from multimodal wearable data in an original data set, and that even with very limited training data, supervised machine learning can achieve a high sensitivity and low FP rate. This methodology may offer a promising way to also approach wearable-based non-convulsive seizure detection.
Citation
