Introduction: Cardiac troponin concentrations vary by age, sex and over time in patients with suspected myocardial infarction.
Hypothesis: Machine learning can combine these variables to improve the assessment of risk for individual patients.
Methods: In a training set of 3,013 patients with suspected myocardial infarction machine learning was used to develop the ‘myocardial-ischemic-injury-index’ [MI3], incorporating age, sex, and paired high-sensitivity cardiac troponin I concentrations at presentation and 1-3 hours later. MI3 determines a value (0-100) reflecting individual likelihood of type 1 myocardial infarction. Calibration and area under the receiver-operator-characteristic curve (AUC) were evaluated in a test set comprising 7,998 patients from 7 international cohorts. Both sets were combined to derive optimal threshold values that identified patients as low- risk and high-risk, and to compare performance to the 99th percentile.
Results: Myocardial infarction occurred in 404 (13.4%) and 849 (10.6%) patients in the training and test sets. MI3 was well calibrated with comparable AUC in the training (0.963 [95% confidence interval 0.957-0.968]) and test (0.963 [0.956-0.971]) sets (Figure). Optimal threshold values were 1.1 and 57.1, which identified 51.6% of patients as low-risk (sensitivity 99.3% [98.9%-99.7%]) and 10.3% as high-risk (PPV 74.9% [72.5%-77.4%]) respectively. Performance was similar in patients with known coronary artery disease or myocardial ischemia, and was better than the 99th percentile (sensitivity 91.1% [89.3%-92.6%], PPV 52.8% [50.7%-55.0%]).
Conclusions: Using machine learning, MI3 provides a personalised and objective assessment of likelihood of myocardial infarction, identifying low-risk patients who may be suitable for early discharge, and high-risk patients who may benefit from earlier treatment.