An Attention-Residual Convolutional Network for Real-Time Seizure Classification on Edge Devices

Area of Science:

• * Neuroscience and Artificial Intelligence
• * Medical Signal Processing
• * Machine Learning for Healthcare

Background:

• * Epilepsy impacts over 50 million globally, necessitating precise seizure type classification for effective treatment.
• * Manual electroencephalogram (EEG) interpretation is time-consuming and requires expert knowledge, hindering clinical workflows.
• * Accurate seizure classification is crucial as different types require specific antiepileptic drugs.

Purpose of the Study:

• * To develop and evaluate EEG-ARCNet, an attention-residual convolutional network for automated multi-channel EEG seizure classification.
• * To assess the model's performance in distinguishing between five common seizure types.
• * To validate the feasibility of deploying EEG-ARCNet on edge devices for practical seizure monitoring.

Main Methods:

• * Developed EEG-ARCNet, integrating residual connections and channel attention for temporal and spectral EEG feature extraction.
• * Combined nine statistical temporal features with five frequency-band power measures using Welch's spectral decomposition.
• * Evaluated the model on the Temple University Hospital Seizure Corpus, comprising multi-channel EEG recordings.

Main Results:

• * Achieved high classification accuracy (99.65%) and macro-averaged F1-score (99.59%) across five seizure types.
• * Demonstrated efficient edge deployment on a Raspberry Pi 4 with a 2.06 ms inference time per 10s segment.
• * Reported low resource utilization: 35.4% CPU and 499.4 MB memory consumption.

Conclusions:

• * EEG-ARCNet offers a highly accurate and efficient solution for automated epilepsy seizure classification.
• * The model's performance and low resource requirements support its use in resource-constrained seizure-monitoring applications.
• * This technology holds potential for improving clinical workflows and patient outcomes in epilepsy management.