A Comprehensive Framework for the Prediction of Intra-Operative Hypotension.

Area of Science:

• Anesthesiology
• Medical Informatics
• Machine Learning

Background:

• Intra-operative hypotension (IOH) prediction models often show a performance gap between development and clinical evaluation.
• A need exists for improved collaboration between data scientists and clinicians for robust IOH model assessment.
• Current fixed-time-to-onset prediction methods have limitations in real-world clinical scenarios.

Purpose of the Study:

• To propose a comprehensive framework for intra-operative hypotension (IOH) prediction.
• To address limitations of fixed-time-to-onset approaches by introducing a sliding window method.
• To evaluate IOH prediction models using precision-recall metrics for better clinical relevance.

Main Methods:

• A sliding window approach was developed to predict IOH risk.
• Precision-recall metrics were used for risk prediction formalization, moving beyond traditional ROC curves.
• A standard machine learning model was applied and trained on non-cardiac surgery data, with testing on both non-cardiac and cardiac surgery datasets.

Main Results:

• The proposed framework significantly improved IOH prediction in non-cardiac surgeries compared to a clinical baseline (48% precision vs. 32% at 28% recall, p<0.0001).
• The model demonstrated generalization capabilities, showing improvement in cardiac surgery prediction, though less pronounced than in non-cardiac cases.
• The sliding window approach and precision-recall metrics provided a more effective evaluation of IOH prediction models.

Conclusions:

• The proposed framework offers a more reliable method for early warning of intra-operative hypotension.
• The sliding window approach and precision-recall metrics enhance the clinical utility and evaluation of IOH prediction models.
• This work highlights the importance of tailored evaluation metrics and collaborative development for clinical prediction models.