Abstracts

Rationale: Machine learning algorithms have shown patient-specific, seizure detection effectiveness (high sensitivity with low false detection rates), but the performance degrades substantially when these approaches are applied across different patients. This work presents a single-channel electroencephalography (EEG) machine-learning framework capable of generalizing across diverse patient demographics while detecting common seizure types with high sensitivity and specificity.

Methods: Data were collected from 169 patients who wore Epitel’s wearable single-channel EEG sensors alongside standard-of-care 19+ wired-channel video-EEG monitoring. 58 of these patients had epileptologist-noted seizure activity in their wired EEG records. The raw, single-channel EEG data was processed through a pipeline that included denoising, detrending, and artifact rejection, standardized using a sliding window across the many-day record, and then reshaped into 2-s, non-overlapping segments. A total of 61 features were extracted for each data segment in the time domain (e.g., variance, line-length), frequency domain (e.g., FFT power in specific frequency bins), and time-frequency domain (e.g., summed power of wavelet-transformed data). Additional feature-level normalization was applied using Deep Adaptive Input Normalization (DAIN) which adaptively changes the applied normalization scheme based on the current input. DAIN is an adaptive version of the z-score normalization scheme and can be applied as any other layer in a neural network. We used a muti-layer perceptron (MLP) network with 3 hidden layers to classify each data segment. The output of each hidden layer was also normalized using DAIN. The Area under Curve (AUC) of a Receiver Operating Characteristics curve was used to evaluate the performance of the classifier.

Results: Training and testing were done on balanced datasets that contained features extracted from 2-s segmented data with equal number of ictal and non-ictal segments. The train/test sets were created by randomly shuffling the segmented data for all the patients and then extracting a 70/30 split that was stratified to ensure similar class distributions in the sets. As a baseline measure, with no data standardization or feature normalization, the classifier resulted in an AUC of 0.84. When the standardization and normalization was applied, the AUC improved to 0.94.

Conclusions: This work explored a machine learning framework that has the ability to detect seizures across a varied patient population from just a single-channel EEG wearable sensor. Our initial results of the segmented-data detection show promise that a non-patient-specific seizure detection algorithm can support those with seizure disorders in their everyday lives.

Funding: This work is supported by NIH NINDS grant U44NS121562.