Predicting depth of anaesthesia from single-channel EEG using a deep TCN-BiLSTM-attention model with EWMA

Accurately determining the depth of anaesthesia (DoA) is crucial for ensuring patient safety and providing individualised anaesthetic management. Widely used monitors, such as the Bispectral Index (BIS), rely on proprietary multichannel EEG algorithms, which limit transparency and accessibility. This study presents a single-lead EEG framework for continuous BIS estimation that is window-level causal, end-to-end, and suitable for near-real-time deployment. The approach integrates a Temporal Convolutional Network (TCN) with power-of-two dilations, a compact bidirectional LSTM for temporal refinement, and additive attention pooling, followed by an exponentially weighted moving average (EWMA) to stabilize predictions over time. This design captures multi-scale temporal dependencies directly from raw EEG while preserving interpretability and low latency. Using a random segment-level split on a public perioperative EEG-BIS dataset, the proposed model achieved a mean absolute error (MAE) of 4.499, root mean square error (RMSE) of 7.499, Pearson's correlation coefficient (r) of 0.903, and Lin's concordance correlation coefficient (CCC) of 0.897 relative to reference BIS values. Under a stricter subject-independent 5-fold GroupKFold evaluation, performance decreased as expected. Still, it remained stable across folds, with the best configuration achieving an MAE of 6.03 ± 0.38 and a CCC of 0.819 ± 0.050 after EWMA smoothing. With approximately 1.07 million parameters and a throughput of about 430 segments per second, the proposed framework offers an efficient and transparent solution for single-sensor DoA monitoring. Overall, this work advances data-driven BIS estimation beyond feature-based methods while explicitly addressing both within-dataset performance and generalization to unseen subjects.