Improving prognostic risk assessment of cardiovascular events with machine learning: An evaluation using positron emission tomography myocardial perfusion imaging

Background Machine learning (ML) holds potential for improving risk assessment in patients with suspected or confirmed coronary artery disease (CAD). However, certain approaches offer greater benefit than others for this task, particularly to capture non-linearity between variables as well as case-by-case explainability. Methods We included consecutive patients who underwent clinically indicated positron emission tomography (PET) imaging. Using automated machine learning (AutoML) and unseen data for performance testing, clinical and PET variables were used to train the predictive models. A logistic regression (LR) and a deep feed-forward neural network (DNN) were trained on the same data for comparison. Major adverse cardiovascular events (MACEs) included death, myocardial infarction, or coronary revascularization >90 days after imaging. Results We included 8,357 patients (80% for development and 20% held out for testing), 46.3% females, with a mean (standard deviation) age of 67.2 (11.7) years. The median (interquartile range) myocardial flow reserve (MFR) was 2.1 (1.6 to 2.6). After an average follow-up of 589 days, a total of 852 patients (10.2%) experienced MACEs. The AutoML achieved an area under the receiver operator curve of .82 (95% confidence interval: .79 to .85) versus .79 (.76 to .82) and .76 (.73 to .80) for the LR and the DNN models, respectively. Model explainability showed that MFR topped the list of most impactful features, followed by total perfusion defects, serum creatinine, and diastolic blood pressure. Conclusions An AutoML model integrating clinical and PET data discriminated MACE risk in CAD more accurately than LR or DNN models and provides interpretable patient-level explanations that can inform personalized care.