Development and validation of 'Patient Optimizer' (POP) algorithms for predicting surgical risk with machine learning

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Surgical Risk Assessment

Background:

• Pre-operative risk assessment is crucial for surgical patient preparation.
• It helps reduce perioperative complications, hospital stay, readmissions, and mortality.
• It also supports collaborative decision-making and operational planning.

Purpose of the Study:

• To develop Machine Learning (ML) algorithms, named Patient Optimizer (POP), for pre-operative risk assessment.
• To predict post-operative complications and inform future large-scale prospective studies.
• To utilize ML for enhanced prediction of surgical outcomes.

Main Methods:

• Developed a base model combining patient and procedure risk, then automated with additional variables.
• Employed 10-fold cross-validation and feature selection for model optimization.
• Utilized logistic regression (LR) and extreme gradient-boosted trees (XGBoost) for model creation.
• Evaluated model performance using Area Under the Receiver Operating Characteristic Curve (AUROC) and Area Under the Precision-Recall Curve (AUPRC).

Main Results:

• XGBoost and LR models showed similar performance across various endpoints.
• POP with XGBoost achieved AUROC of 0.755 for any post-operative complication, 0.869 for kidney failure, and 0.841 for length of stay.
• For 30-day readmission and mortality, POP with XGBoost achieved AUROC of 0.610 and 0.866, respectively.

Conclusions:

• The developed POP algorithms demonstrate effectiveness in predicting post-operative complications, kidney failure, and length of stay.
• A larger study is warranted to refine the algorithms for improved prediction accuracy.
• Expanded datasets could enhance predictions for specific complications, readmission, and mortality.