Quantifying surgical complexity with machine learning: looking beyond patient factors to improve surgical models
Alexander Van Esbroeck1, Ilan Rubinfeld2, Bruce Hall3
1Department of Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, MI.
Surgery
|August 11, 2014
Summary
Machine learning accurately estimates surgical procedure risk, improving models for predicting complications like mortality and infections. This data-driven approach enhances patient care by identifying individual procedural risks.
Area of Science:
- Medical Informatics
- Machine Learning in Healthcare
- Surgical Outcomes Research
Background:
- Accurate prediction of surgical risk is crucial for patient safety and resource allocation.
- Existing models often lack precision in assessing individual procedural risks.
- Data-driven approaches offer potential for improved risk stratification.
Purpose of the Study:
- To develop and validate machine learning models for empirically determining individual surgical procedure risk.
- To enhance existing surgical risk prediction models using machine learning-derived procedural risk scores.
- To investigate the relationship between Current Procedural Terminology (CPT) descriptions and surgical outcomes.
Main Methods:
- Support Vector Machine (SVM) classifiers were trained using American College of Surgeons National Surgical Quality Improvement Program (ACS NSQIP) data (2005-2009).
- Models learned the association between CPT descriptions and 30-day outcomes: mortality, morbidity, Clavien 4 complications, and surgical-site infections (SSI).
- Procedural risk scores were validated on 2010 data using univariate and multivariate analyses.
Main Results:
- SVM classifiers achieved moderate-to-high discrimination for mortality (AUC 0.871), morbidity (AUC 0.789), SSI (AUC 0.791), and Clavien 4 complications (AUC 0.845).
- Incorporating risk scores significantly improved multivariate prediction models (Net Reclassification Improvement up to 0.68, P < .05).
- Improvements in discrimination and calibration were consistent across statistical measures and surgical subcohorts.
Conclusions:
- Machine learning effectively generates clinically useful, data-driven estimates of individual surgical procedure risk.
- These risk scores substantially enhance multifactorial models for predicting postoperative complications.
- This approach offers a novel method for improving surgical risk assessment and patient safety.
