Wearable Sensors Feasible for Automatic Seizure Detection
Preliminary results showed that a fusion of accelerometry and blood volume post modalities achieved the best AUC-ROC for 9 combined seizure types.
Jianbin Tang
Findings from a recent study indicate automatic seizure detection using machine learning (ML) from multimodal sensor data is feasible to detect a broad range of epileptic seizures. Pediatric participants wore sensors on wrists or ankles, with ML models trained to detect all seizure types performing better than models trained to detect specific types.
Of the 94 patients included in the study, investigators reported a total of 548 epileptic seizures over the course of 11,066 hours of sensor data collection, resulting in 930 seizures in total and 9 seizure types. A combination of accelerometry (ACC) and blood volume post (BVP) data with convolutional neural network algorithms achieved the best overall area under the receiver operating characteristic curve (AUC-ROC [0.752]). Algorithm 1, which was trained with seizure-type specific detection models of 9 seizure types, detected 8 out of 9 seizure types better than chance (AUC-ROC = 0.648-0.976). Algorithm 2, which had a general type-agnostic seizure detection system for all seizure types, detected all 9 seizure types better than chance (AUC-ROC = 0.642-0.995).
“Results show the feasibility of seizure detection across a broad spectrum of seizure types, including but not limited to generalized tonic-clonic seizures [GTCSs], with wrist-worn wearable sensors in a large cohort of patients,” lead author Jianbin Tang, data scientist and technical leader, IBM Research Australia, et al wrote. “Our study expands current literature by demonstrating that noninvasive, wrist- and ankle-worn sensors and custom-developed deep learning techniques can automatically detect a variety of epileptic seizure types.”
READ MORE: FDA Accepts Resubmission of Diazepam Buccal Film NDA for Seizure Clusters
Participants had a median age of 9.9 years and were asked to wear sensors, which recorded body temperature, electrodermal activity, ACC, and photoplethysmography to provide BVP, during long-term admission to the epilepsy monitoring unit at Boston Children’s Hospital. Investigators utilized electroencephalographic (EEG) seizure onset and offset as a standard comparison, which remains the current gold standard for diagnosing and evaluating epilepsy. Criteria for exclusion included clusters of 4 our more seizures in 15-minute time windows and high hourly frequency of seizures. The 9 included seizure types were focal to bilateral tonic-clonic seizures, GTCSs, focal tonic seizures, focal subclinical seizures, focal automatisms, focal behavior arrest, focal clonic seizures, generalized tonic seizures, and generalized epileptic spasms.
The study faced certain limitations, including the differences in behavior for pediatric patients during long-term hospital stays, leading to a difference in device signals, as well as patients’ ability to be mobile and partake in activities. Additionally, not all seizure types were investigated, and the possibility of inconsistency resulted from wearing devices on different parts or sides of the body.
“A user-friendly, portable, noninvasive, nonstigmatizing tool that reliably detects seizures can improve patients' quality of life and their health outcomes and may improve the evaluation of treatment outcomes based on seizure frequency,” Tang et al wrote. “Our results suggest that wrist-worn and commercially available sensors, running advanced deep learning models and data preprocessing techniques, could be a feasible out-of-the-box starting alternative to custom-developed monitoring devices. Our results also suggest that individualized customization of detection modalities based on clinical features, including seizure semiology, may improve detection performance in selected patients.”
According to investigators, while preliminary results show promise for the use of ML and wearable data in detecting automatic epileptic seizure, further research should consider clinical chronoepileptological variables. Detection performance may also be improved with future adjustment, and investigators noted the next steps will be validation of results in larger data sets, evaluation of the tool with other seizure types, as well as integrating additional clinical information.
Jianbin and colleagues discussed surprising aspects of the study, telling NeurologyLive, "Interestingly, though our model detected seizures with good accuracy for most of the individuals in the study, the framework was not as accurate for a small subset of individuals. This points to the need for further study into the wide range of epileptic seizure types and their expression in specific patients and could indicate the prevalence of understudied seizures in these individuals."
Investigators' research is part of a larger effort to change how clinicians use artificial intelligence. By evaluating daily markers, Jianbin et al hope to accurately monitor, treat, and eventually prevent epileptic seizure, as well as other diseases.
