Machine learning to optimise use of cardiac troponin in the diagnosis of acute myocardial infarction

Abstract

Abstract Background Guidelines recommend fixed cardiac troponin thresholds for the assessment of patients with suspected acute coronary syndrome, however, performance varies in important patient groups as concentrations are influenced by age, sex and comorbidities. This limitation can be addressed using machine learning algorithms. Methods Machine learning algorithms were developed that integrate cardiac troponin concentrations at presentation or on serial testing with age, sex and clinical features in 10,038 consecutive emergency patients with suspected acute coronary syndrome. The primary outcome was an adjudicated diagnosis of type 1, type 4b or type 4c myocardial infarction. The best performing algorithm was selected for the CoDE-ACS (Collaboration for the Diagnosis and Evaluation of Acute Coronary Syndrome) decision-support tool, and performance was externally validated in 3,035 patients pooled from three prospective studies. Findings CoDE-ACS had excellent discrimination and calibration using cardiac troponin at presentation (area under curve [AUC] 0.959, 95% confidence interval 0.948–0.971, Brier score 0.040), in the pooled external validation cohort. At presentation, the rule-out score identified 62.1% (1,885/3,035) of all patients as low-probability of myocardial infarction with a 99.5% (99.1–99.7%) negative predictive value and 97.0% (96.3–97.6%) sensitivity. The rule-in score identified 8.3% (252/3,035) of patients as high-probability with an 83.7% (82.4–85.0%) positive predictive value and 98.5% (98.0–98.9%) specificity. Performance of the rule-out and rule-in scores was consistent across patient subgroups (Figure 1 and Figure 2). CoDE-ACS incorporating a second cardiac troponin measurement also had excellent discrimination and calibration (AUC 0.971 [0.962–0.980], Brier score 0.039) and refined the individualised probabilities in the 29.5% (898/3,035) of patients neither ruled-out or ruled-in at presentation to guide further investigation. Conclusions We developed and externally validated the CoDE-ACS decision-support tool using machine learning to aid in the diagnosis of myocardial infarction. CoDE-ACS had excellent diagnostic performance to rule-out and rule-in myocardial infarction at presentation, performed consistently across patient subgroups, and provided individualised probabilities to guide further care in those who require serial troponin measurements. Conclusions We developed and externally validated the CoDE-ACS decision-support tool using machine learning to aid in the diagnosis of myocardial infarction. CoDE-ACS had excellent diagnostic performance to rule-out and rule-in myocardial infarction at presentation, performed consistently across patient subgroups, and provided individualised probabilities to guide further care in those who require serial troponin measurements. Funding Acknowledgement Type of funding sources: Public Institution(s). Main funding source(s): National Institute for Health ResearchBritish Heart Foundation