Background:
The ACC/AHA Pooled Cohort Equation (PCE) for Atherosclerotic cardiovascular disease (ASCVD) has shown modest accuracy. We assessed if machine learning algorithms (MLA) could improve PCE performance with traditional and selected enriched clinical features.
Methods:
We tested MLA in a community cohort of patients >30 years, that sought primary care in Olmsted County, MN. Inclusion criteria was as of the PCE. ASCVD events were validated in duplicate and included fatal and non-fatal myocardial infarction and ischemic stroke at 10 years, analyzed as a composite and individual outcomes. A random sample of 70% of the dataset was used for training and optimal MLA were selected with grid search. Performance was evaluated in an independent testing set with the remaining observations.
Results:
We included 34,831 adults, mean ± SD age 49.8 ± 23.2, 54% female, baseline PCE risk 6.54 ±11.5. There were 4,255 ASCVD events (12.2%) in 2,613 people: 1,850 non-fatal MI, 901 MI deaths, 1,299 ischemic strokes and 205 ischemic stroke deaths. MLA did not perform better than the PCE to predict ASCVD as whole or specific outcomes. Logistic regression provides modest improvement to predict ASCVD when adding enriched features (ROC 0.83361 vs 0.81843, p-value=0.03). Interestingly, all perform better when predicting fatal events vs. non-fatal events. See Table.
Conclusion:
MLA do not perform better than PCE when using the same variables and likely require additional features to enhance predictive capabilities.