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

Area of Science:

• Neuroscience
• Biomedical Engineering
• Anesthesiology

Background:

• Accurate depth of anesthesia (DoA) monitoring is vital for patient safety and personalized anesthetic care.
• Current Bispectral Index (BIS) monitors use proprietary algorithms, limiting transparency and accessibility.

Purpose of the Study:

• To develop a transparent, single-lead electroencephalogram (EEG) framework for continuous DoA estimation.
• To create a model suitable for near-real-time deployment and efficient processing.

Main Methods:

• Integrated a Temporal Convolutional Network (TCN) with bidirectional LSTM and attention pooling for raw EEG analysis.
• Employed an exponentially weighted moving average (EWMA) for prediction stabilization.
• Utilized a public perioperative EEG-BIS dataset for model training and validation.

Main Results:

• Achieved high accuracy on a random split dataset (MAE: 4.499, CCC: 0.897).
• Demonstrated stable performance in subject-independent cross-validation (best MAE: 6.03 ± 0.38, CCC: 0.819 ± 0.050).
• Framework is computationally efficient with ~1.07 million parameters and high throughput (~430 segments/sec).

Conclusions:

• The proposed framework offers an efficient, interpretable, and low-latency solution for single-sensor DoA monitoring.
• Advances data-driven BIS estimation beyond traditional feature-based methods.
• Addresses both in-dataset performance and generalization to unseen subjects, enhancing clinical applicability.