Artificial intelligence-enhanced epileptic seizure detection by wearables

Area of Science:

• Biomedical Engineering
• Neuroscience
• Digital Health

Background:

• Epilepsy monitoring traditionally relies on electroencephalography (EEG), which can be intrusive.
• Wearable sensors offer a less invasive alternative for continuous seizure detection.

Purpose of the Study:

• To develop and evaluate deep-learning models for detecting various seizure types using wearable biosensor data.
• To assess the performance of different sensor modalities and model architectures for seizure detection.

Main Methods:

• Collected electrodermal activity, accelerometry (ACC), and blood volume pulse (BVP) from wrist/ankle sensors in 166 epilepsy patients.
• Applied convolutional neural network (CNN) and CNN-long short-term memory (LSTM) models for seizure detection.
• Evaluated models on 28 seizure types using sensitivity, false positive rate, and detection delay.

Main Results:

• A CNN-LSTM model using fused ACC and BVP data achieved 83.9% sensitivity and a 35.3% false positive rate.
• Nineteen of 28 seizure types were detectable by at least one data modality with an area under the curve > 0.8.
• Wearable sensor data fusion outperformed individual modalities in seizure detection.

Conclusions:

• Deep learning models effectively detect diverse seizure types using noninvasive wearable sensors.
• This approach represents a paradigm shift towards more accessible and patient-centered epilepsy monitoring.
• Combines generalizable algorithms with personalized seizure detection for enhanced clinical utility.