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

Area of Science:

• Anesthesiology and Critical Care Medicine
• Biomedical Engineering
• Data Science in Healthcare

Background:

• Accurate monitoring of anaesthesia depth is vital for patient safety.
• Processed electroencephalogram (EEG) monitoring, like Bispectral Index (BIS), is common but not universally available.
• This study explores heart rate (HR) dynamics for predicting inadequate anaesthesia (BIS >60).

Purpose of the Study:

• To predict episodes of inadequate anaesthesia (defined as BIS >60) using heart rate dynamics.
• To develop a computationally efficient and clinically feasible monitoring tool.
• To characterize HR dynamics that correlate with changes in anaesthesia depth.

Main Methods:

• Analysis of electrocardiogram (ECG) data from 3338 surgical patients.
• Extraction of HR time-series features preceding BIS >60 events at 0, 5, 10, and 15 minutes.
• Training and evaluation of gradient boosting models with nested 10-fold cross-validation and feature reduction.

Main Results:

• Models demonstrated strong predictive performance, with AUC values ranging from 0.903 to 0.953.
• A compact model with 27 features maintained high performance and achieved a 110-fold improvement in computational speed.
• The predictive features primarily captured fractal characteristics of HR dynamics.

Conclusions:

• High-dimensional HR dynamics descriptors accurately predict inadequate anaesthesia events (BIS >60).
• A compact subset of 27 features offers a clinically feasible and computationally efficient method for anaesthesia depth monitoring.
• This approach holds significant potential for improving patient safety during anaesthesia.