Jianbin Tang, data scientist and research leader from IBM Research Australia, and colleagues spoke on findings from their recent study and plans for a newly developed artificial intelligence-based model.
Although manually recorded seizure diaries remain the mainstay for recording and the gold standard for tracking is electroencephalography (EEG), the need for improvement in how seizures are accounted for remains in the epilepsy space.
Jianbin Tang, data scientist and technical leader, IBM Research Australia, and colleagues Stefan Harrer, chief innovation officer, Digital Health Cooperative Research Center, ex-IBM Research; Rima El-Atrache, MD, neurology fellow, Harvard Medical School, Boston Children’s Hospital; and Tobias Loddenkemper, MD, professor of neurology, Harvard Medical School and Boston Children’s Hospital, were part of a recent study on the feasibility of multimodal sensor data for automatic seizure detection. Following data analysis, investigators found that wearable detectors were able to accurately account for when a patient was having a seizure, regardless of seizure type.
Investigators asked pediatric patients at Boston Children’s Hospital to wear wristbands on either their wrist of ankles during their long-term admission to the epilepsy monitoring unit, ultimately concluding that a fusion of accelerometry and blood volume post modalities achieved the best overall area under the receiver operating characteristic curve for 9 combined seizure types. Tang et al shared insight on motivations that prompted the study, clinical benefit, and surprising aspects of results in this interview with NeurologyLive.
Jianbin Tang, et al: In the offices of pediatric neurologists, epilepsy is 1 of the most common diagnoses, affecting 1-2% of the childhood population.1 Despite epilepsy’s long-term effects and impact on quality of life, there is still a great deal of room for improvement in the tracking and identification of seizures. Recording epileptic events accurately is vital for monitoring, evaluation, and effective treatment.
Yet, clinicians today do not have a standard, objective, and reliable way to track when and how often seizures occur in their patients. They primarily rely on manually recorded seizure diaries maintained by patients themselves. This method can often leave a large amount of room for error and is subjective to the patient’s memory. For major seizures, this has often worked, but many epilepsy patients also suffer from smaller, less noticeable seizures that can be hard to detect and easy to forget to track.
The other alternative, but less accessible, way to track seizures is through EEG data, which is currently the gold standard for seizure diagnosis. However, tracking and analyzing EEG data usually requires a hospital stay, and can be costly and disruptive to a patient’s daily life.
A combined team of clinicians, neuroscientists and researchers from IBM Research and Boston Children’s Hospital has tackled the issues inherent in seizure tracking with AI and wearable devices. The goal is to build a framework, founded in technology, that can open up noninvasive and more accurate seizure detection to a wide population. Better seizure detection and tracking could lay the groundwork for a greater understanding of the broad range of seizures suffered by epileptic patients, as well as more personalized and precise treatment and management.
Our initiative began with enrolling a broad group of patients in the epilepsy monitoring unit at Boston Children’s Hospital, one of the largest groups yet studied in the field of pediatric epilepsy. We asked them to wear a smart wristband that collected autonomic biomarkers such as electrodermal activity, temperature, photoplethysmography (from which heart rate can be derived), and accelerometer.
After manually annotating the epileptic seizures that occurred in this monitoring data, we trained deep learning algorithms to automatically recognize seizure segments across a broad spectrum of seizure types from this dataset. Learning from this data, our model detected and identified when many of the individuals in our study were experiencing a seizure, regardless of which specific seizure type they were having across a wide range of possibilities.
Published in the peer-reviewed journal Epilepsia, our team is unveiling a new artificial intelligence (AI)-based model that can automatically detect a wide variety of epileptic seizures by collecting and analyzing data from widely available commercial sensors, such as smartwatches and wristbands.2 This is a feat that has not yet been accomplished in the study of epilepsy and proves AI’s ability to detect a broad spectrum of seizures with simple devices such as relatively low-cost wearables. If this research is expanded upon, this could mean that one day epileptic patients could work with their doctors to automatically monitor seizures with noninvasive, accessible devices in their regular daily lives and environments, freeing them from the burden of manual recording and tracking.
Our best performing detection models reached an AUC-ROC of 75.2% across all seizure types and up to 99.5% for specific seizure types such as generalized tonic-clonic seizures.
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.
When managing epilepsy, the accurate monitoring of seizures is critical to assessing risk, preventing injury, and evaluating patients’ response to treatment. This work is part of our broader vision to transform how clinicians can use AI and cloud to understand how individuals’ daily markers such as for example speech, movement, sleep, and biometrics can help to better monitor, treat, and even prevent a wide range of diseases.
Transcript edited for clarity.