Machine learning models may not be associated with improvement in predicting patients at higher risk of in-hospital mortality after acute myocardial infarction (AMI), according to a study published March 10 in JAMA Cardiology.
Rohan Khera, MD, MS, et al., used data from ACC's Chest Pain – MI Registry to evaluate whether machine learning methods can facilitate prediction of in-hospital mortality risk after AMI by including a larger number of variables. The researchers developed and validated three machine learning models to predict in-hospital mortality based on comorbidities, medical history, presentation characteristics and initial laboratory values. The models were based on extreme gradient descent boosting (XGBoost), a neural network and a meta-classifier. Accuracy was compared against the current standard using a logistic regression model in a validation sample.
The study cohort consisted of 755,402 patients with a mean age of 65 years, 65.5% of whom were male. Overall, the primary outcome of death from any cause during hospitalization was 4.4%. The machine learning models showed "modest improvements" in discrimination over logistic regression. The XGBoost and meta-classifier models achieved a discrimination of 0.898 (95% confidence interval [CI]: 0.894-0.902) and 0.899 (95% CI: 0.895-0.903), respectively, compared with 0.888 (95% CI: 0.884-0.892) for the logistic regression model.
Calibration slopes improved for the XGBoost and meta-classifier models, but not the neural network model, when compared with logistic regression and applied to a limited or expanded set of variables. In addition, the XGBoost model reclassified 32,393 (27%) individuals and the meta-classifier model reclassified 30,836 (25%) individuals considered to be moderate or high risk of death in logistic regression as low risk. The reclassification was more consistent with observed event rates.
According to the researchers, "improvements in risk prediction for in-hospital mortality with machine learning models were small and likely do not meet the threshold to be relevant for clinical practice." They conclude that "compared with logistic regression, the models offered improved resolution of risk for high-risk individuals."