Machine learning-based cardiovascular risk calculator for non-cardiac surgery

Area of Science:

• Cardiovascular medicine
• Machine learning applications in healthcare
• Surgical risk assessment

Background:

• 4% of the global population undergoes non-cardiac surgery annually.
• 30% of these patients have cardiovascular risk factors, with 0.5%-2% 30-day mortality.
• Need for accurate cardiovascular risk prediction in this population.

Purpose of the Study:

• Develop an interpretable machine learning model for cardiovascular risk scoring.
• Predict risk for patients >50 years old undergoing non-cardiac surgery.
• Assess risk from surgery date to 30 days post-surgery.

Main Methods:

• Utilized the NSQIP 2022 dataset (4,970,011 patients).
• Defined primary endpoint as 30-day death, myocardial infarction, cardiac arrest, or stroke.
• Trained and evaluated multiple machine learning algorithms (Logistic Regression, Naive Bayes, Random Forest, boosting trees) using AUROC.

Main Results:

• LightGBM achieved the highest AUROC of 0.9009 (95% CI: 0.8889-0.9126).
• The best model identified six key predictors: surgery type, ASA classification, BUN, sepsis, emergent surgery, and mechanical ventilation.
• The model demonstrated strong predictive accuracy and generalization.

Conclusions:

• LightGBM classifier is optimal for cardiovascular risk scoring in this context.
• The model balances prediction accuracy and generalization effectively.
• Identified key factors for cardiovascular risk assessment in non-cardiac surgery patients.