Using machine learning to predict acute myocardial infarction and ischemic heart disease in primary care cardiovascular patients

Background Early recognition, which preferably happens in primary care, is the most important tool to combat cardiovascular disease (CVD). This study aims to predict acute myocardial infarction (AMI) and ischemic heart disease (IHD) using Machine Learning (ML) in primary care cardiovascular patients. We compare the ML-models' performance with that of the common SMART algorithm and discuss clinical implications.Methods and results Patient-level medical record data (n = 13,218) collected between 2011-2021 from 90 GP-practices were used to construct two random forest models (one for AMI and one for IHD) as well as a linear model based on the SMART risk prediction algorithm as a suitable comparator. The data contained patient-level predictors, including demographics, procedures, medications, biometrics, and diagnosis. Temporal cross-validation was used to assess performance. Furthermore, predictors that contributed most to the ML-models' accuracy were identified. The ML-model predicting AMI had an accuracy of 0.97, a sensitivity of 0.67, a specificity of 1.00 and a precision of 0.99. The AUC was 0.96 and the Brier score was 0.03. The IHD-model had similar performance. In both ML-models anticoagulants/antiplatelet use, systolic blood pressure, mean blood glucose, and eGFR contributed most to model accuracy. For both outcomes, the SMART algorithm was substantially outperformed by ML on all metrics.Methods and results Patient-level medical record data (n = 13,218) collected between 2011-2021 from 90 GP-practices were used to construct two random forest models (one for AMI and one for IHD) as well as a linear model based on the SMART risk prediction algorithm as a suitable comparator. The data contained patient-level predictors, including demographics, procedures, medications, biometrics, and diagnosis. Temporal cross-validation was used to assess performance. Furthermore, predictors that contributed most to the ML-models' accuracy were identified. The ML-model predicting AMI had an accuracy of 0.97, a sensitivity of 0.67, a specificity of 1.00 and a precision of 0.99. The AUC was 0.96 and the Brier score was 0.03. The IHD-model had similar performance. In both ML-models anticoagulants/antiplatelet use, systolic blood pressure, mean blood glucose, and eGFR contributed most to model accuracy. For both outcomes, the SMART algorithm was substantially outperformed by ML on all metrics.Conclusion Our findings underline the potential of using ML for CVD prediction purposes in primary care, although the interpretation of predictors can be difficult. Clinicians, patients, and researchers might benefit from transitioning to using ML-models in support of individualized predictions by primary care physicians and subsequent (secondary) prevention.