Kenny Walter is an editor with HCPLive. Prior to joining MJH Life Sciences in 2019, he worked as a digital reporter covering nanotechnology, life sciences, material science and more with R&D Magazine. He graduated with a degree in journalism from Temple University in 2008 and began his career as a local reporter for a chain of weekly newspapers based on the Jersey shore. When not working, he enjoys going to the beach and enjoying the shore in the summer and watching North Carolina Tar Heel basketball in the winter.
Researchers test 3 different machine learning models to aid in expanding the number of variables needed to better predict outcomes.
It is often difficult to forecast who might be more susceptible to an adverse outcome from an acute myocardial infarction.
However, a new machine learning technique could better predict these outcomes, guiding triage of care services and shared decision-making.
A team, led by Rohan Khera, MD, MS, Section of Cardiovascular Medicine, Department of Internal Medicine, Yale School of Medicine, evaluated whether contemporary machine learning techniques could facilitate better risk prediction by including a larger number of variables, while identifying complex relationships between predictors and outcomes.
In the cohort study, the investigators identified all acute myocardial infarction hospitalizations between 2011-2016 using the American College of Cardiology Chest Pain-MI Registry. They identified a total of 755,402 patients, with a mean age of 65 years old.
In addition, 65.5% (n = 495,202) of the patient population was male.
The investigators developed a trio of machine learning models and validated each model to predict in-hospital mortality based on patient comorbidities, medical history, presentation characteristics, and initial laboratory values.
Each model as developed based on extreme gradient descent boosting (XGBoost, an interpretable model), a neural network, and a meta-classifier model, with their accuracy compared against the current standard developed using a logistic regression model in a validation sample.
“In independent validation, 2 machine learning models, gradient descent boosting and meta-classifier (combination including inputs from gradient descent boosting and a neural network), marginally improved discrimination compared with logistic regression (C statistic, 0.90 for best performing machine learning model vs 0.89 for logistic regression),” the authors wrote.
The investigators found nearly perfect calibration in independent validation data in the XGBoost (slope of predicted to observed events, 1.01; 95% CI, 0.99-1.04) and the meta-classifier model (slope of predicted-to-observed events, 1.01; 95% CI, 0.99-1.02). In addition, there was more precise classification found across the risk spectrum.
The XGBoost model reclassified 27% (n = 32,393) and the meta-classifier model reclassified 25% (n = 30,836) deemed at moderate to high risk for death in logistic regression as low risk, which were more consistent with the observed event rates.
“In this cohort study using a large national registry, none of the tested machine learning models were associated with substantive improvement in the discrimination of in-hospital mortality after AMI, limiting their clinical utility,” the authors wrote. “However, compared with logistic regression, XGBoost and meta-classifier models, but not the neural network, offered improved resolution of risk for high-risk individuals.”
Better Testing Plans
In the initial assessment of acute coronary syndrome, biomarker testing with troponin levels could improve screening for the disease in clinical practice, but patients can also be discharged if they are deemed at a low risk following a single negative troponin test result.
Recently, researchers examined the clinical outcomes of patients discharged following a single negative troponin test result compared to patients discharged after serial troponin measurements.
When the researchers adjusted for cardiac risk factors and comorbidities they found no statistically significant difference in the primary outcome of acute myocardial infarction or cardiac mortality within 30 days between the 2 groups (single troponin, n = 56; 0.4%; serial troponin, n = 52; 0.4%; aOR, 1.41; 95% CI, 0.96-2.07).
They also found patients discharged following a single troponin test had lower rates of coronary artery bypass graft (aOR, 0.24; 95% CI, 0.11-0.48) and invasive coronary angiography (aOR, 0.46; 95% CI, 0.38-0.56).
The study, “Use of Machine Learning Models to Predict Death After Acute Myocardial Infarction,” was published online in JAMA Cardiology.