Forecasting intraoperative hypotension during hepatobiliary surgery.

Area of Science:

• Anesthesiology
• Medical Informatics
• Machine Learning

Background:

• Intraoperative hypotension (IOH) is linked to increased postoperative complications.
• Machine learning (ML) shows promise for predicting IOH.
• Current ML models may benefit from enhanced feature integration.

Purpose of the Study:

• To develop and evaluate an ML model for predicting IOH in adult patients undergoing hepatobiliary surgery.
• To investigate the impact of demographic and physiological features on IOH prediction accuracy.
• To introduce a novel adjustable feature for optimizing prediction sensitivity and specificity.

Main Methods:

• A multivariate random forest (RF) algorithm was trained using 13 physiological time series and patient demographic data (age, sex, BMI).
• A novel multi-model voting (MMV) approach with an adjustable "dial" was implemented for dynamic sliding window predictions.
• The model was evaluated on a cohort with 85% experiencing at least one IOH event.

Main Results:

• The multivariate model achieved an average AUC of 0.97 for static predictions up to 8 minutes before IOH, outperforming a univariate MAP-only model (AUC 0.83).
• The MMV model demonstrated strong performance in a dynamic sliding window context (40 minutes prior), with AUC=0.96, PPV=0.89, and NPV=0.98.
• The developed ML model offers adjustable sensitivity and specificity for predicting the first IOH episode.

Conclusions:

• Integrating demographic and physiological data significantly improves ML-based IOH prediction.
• The novel MMV approach and adjustable feature enhance predictive performance in dynamic surgical settings.
• This ML model provides a valuable tool for proactive IOH management during liver resection surgeries.