Introduction: Patients at high risk for myocardial infarction (MI) benefit from treatments designed for primary prevention, especially cholesterol lowering therapy. The pooled cohort equations (PCE) are the most commonly used risk predictor for future MI but show only modest performance.

Hypothesis: Electronic health record (EHR)-based, and ECG-based machine learning models are better at predicting MI as compared toPCE.

Methods: We retrospectively included all patients (age 40-79) with clinical encounters in Geisinger, having at least 1 clinically acquired ECG and no prior history of MI. Incident MI within 10 years of an index ECG was identified. The ability of 2 machine learning models to predict MI were compared to the PCE: 1) an XGBoost model with EHR data (age, sex, race, smoking history, vitals, 24 lab tests) as input features; and 2) a deep neural network (DNN) that used ECG traces, age, and sex as inputs. Models were compared based on area under the receiver operating characteristic curve (AUROC).

Results: A total of 103,933 ECGs from 34,932 patients had sufficient follow-up, 21% of ECGs were followed by an MI event within 10 years. The EHR-based XGBoost model had the best performance (AUROC 81%) in comparison with both the ECG-based DNN model (AUROC 68%) and the PCE (AUROC 72%). 9% of the total encounters were predicted to be ‘high risk’ by the EHR-based model and not by the pooled cohort equations. In that subgroup, 60% of patients were not on a statin and the 10-year MI event rate was 26%. 20% of encounters were predicted to be ‘high risk’ by the PCE and not by the EHR-based model. The event rate in that subgroup was 12%.

Conclusions: An EHR-based XGBoost model, but not an ECG-based DNN, is superior to the PCE in predicting future MI. Patients identified as high risk by the EHR-based model but missed by the PCE, have a high rate of future MI. Statin use in that group is low, suggesting ample opportunity for intervention.

VIEW THE PUBLICATION

VIEW THE POSTER