Coronary artery bypass risk prediction using neural networks

Area of Science:

• Cardiovascular Surgery
• Machine Learning in Medicine
• Health Informatics

Background:

• Neural networks are powerful nonparametric tools for identifying complex patterns.
• Their application in medical risk prediction is an area of active research.

Purpose of the Study:

• To evaluate the effectiveness of multilayer perceptron neural networks (MLP) for predicting mortality risk in patients undergoing coronary artery bypass grafting (CABG).
• To compare MLP performance against traditional logistic regression and Bayesian analysis.

Main Methods:

• Utilized The Society of Thoracic Surgeons database (80,606 patients).
• Compared single, two, and three-layer MLP networks trained with stochastic gradient descent against logistic regression and Bayesian analysis.
• Models were trained on 40,480 patients and validated on 40,126 patients.

Main Results:

• All models, including MLPs, achieved similar areas under the receiver operating characteristic curve (approx. 76%) for mortality prediction.
• A committee classifier combining MLP and logistic regression demonstrated superior calibration, accurately predicting risk for both low and high-risk patients.
• Individual models sometimes overestimated or underestimated risk, particularly for high-risk individuals.

Conclusions:

• A committee classifier integrating MLP and logistic regression offers the best model calibration for CABG risk prediction.
• While calibration improved, the overall predictive accuracy (76% ROC area) remained consistent across different modeling approaches.