Heart rate dynamics predict anaesthetic depth: a compact machine learning model.

BACKGROUND: Accurate monitoring of the depth of anaesthesia is essential for patient safety. Although processed electroencephalogram monitoring is widely used, it is not always available. This study aimed to predict episodes of inadequate anaesthesia, defined as a bispectral index (BIS) value >60, by characterising heart rate (HR) dynamics. METHODS: Electrocardiogram data from 3338 surgical patients were analysed. BIS >60 events were identified, and the HR segments preceding these events at 0, 5, 10, and 15 min were extracted. Time-series features were extracted using a highly comparative time-series analysis framework. Gradient boosting models were trained and evaluated using a nested 10-fold cross-validation. A feature-reduction procedure was used to identify a compact feature subset. RESULTS: All models demonstrated strong discriminative performance, with area under the receiver operating characteristic curve values of 0.953 (95% confidence interval: 0.949-0.958) at 0 min, 0.917 (0.909-0.925) at 5 min, 0.910 (0.905-0.916) at 10 min, and 0.903 (0.894-0.912) at 15 min. A compact model comprising 27 features maintained high performance across all time points while achieving a 110-fold improvement in computational speed. These features predominantly captured fractal HR dynamics. CONCLUSIONS: High-dimensional descriptors of heart rate dynamics enabled accurate prediction of BIS >60 events. By identifying a compact subset of 27 features, this approach demonstrates potential as a clinically feasible and computationally efficient tool for monitoring the depth of anaesthesia.