Real-Time EEG-Based Anaesthesia Monitoring Using Neural Networks on Edge-AI Hardware

Background: Monitoring unconsciousness under general anaesthesia typically relies on indirect physiological indicators rather than direct measures of brain activity. This study investigates a compact, real-time electroencephalography (EEG)-based monitoring system that combines machine learning (ML) and edge artificial intelligence (Edge-AI) for assessing anaesthesia-induced unconsciousness. Methods: Multitaper spectral analysis (MSA) was incorporated into the real-time processing pipeline to enable feature extraction from raw EEG prior to inference using LR, LDA, LDA+LR, NN and QNN classifier models. A compact neural network was developed using spectral features from 2 s EEG epochs obtained from an available third-party dataset and evaluated using subject-independent five-fold cross-validation. Across the evaluated classifiers and class-balancing strategies, the neural network trained using the Synthetic Minority Over-sampling Technique (SMOTE) provided the strongest overall performance. The selected model was subsequently quantised to 8-bit precision and deployed as a quantised neural network on an untethered PYNQ-Z2 FPGA module. Results: The neural network achieved an accuracy of 0.90 under subject-independent evaluation. Following 8-bit quantisation, the FPGA-deployed model delivered an accuracy of 0.88, while achieving an inference-stage latency below 0.8 ms and thus a computational speed-up exceeding 2.7×. Conclusions: These results demonstrate the feasibility of combining subject-independent EEG-based classification with efficient, low-latency Edge-AI implementation for real-time anaesthesia monitoring. The proposed system provides a foundation for the development of EEG-based decision-support tools for anaesthesiologists and has potential for future deployment in resource-constrained or remote monitoring environments.