OBJECTIVES: Even cost-effective intervention face challenges in terms of adoption if the upfront economic costs are high due to a large target population. A clinically important example is the adoption of early (within 72 hours) non-invasive cardiac testing (NIT) in the triage of patients suspected for acute coronary syndrome (ACS). Our economic evaluation found use of NIT to be cost-effective (<$6,000/QALY). However, if adopted as standard of care in the 8 million annual suspected ACS cases in the US, NIT could increase expenditure by nearly $35 billion annually. Thus, we explore if machine learning algorithms can be developed to identify clinical features that classify patients most likely at risk of death or acute myocardial infarction (MI) especially in those with pre-test low-risk.

METHODS: We used a retrospective cohort study design within the adult ED patient population in whom MI was ruled out, belonging to a large managed care plan. We only included patients with pre-test low-risk using HEART risk score and followed them up to 1-year post ED discharge. First, we used least absolute shrinkage and selection operator techniques to reduce the large number of baseline socio-demographic, cardiac and non-cardiac conditions that could be potential classifiers of death/MI risk. We then used k-fold Classification and Regression Tree Analysis (CART) to identify important features that contribute to the risk of future MI/death.

RESULTS: The cohort included 106,478 patients [mean age 53 (±15) years; female 58%]. CART found age above 65 followed by elevated troponin level as the most prominent features for future MI/death. Peripheral vascular disease was identified as the next important feature ahead of CAD and CHF.

CONCLUSIONS: We implemented machine learning techniques to further classify low-risk patients using smaller set of clinical features and created a decision tree. Our findings may help improve the economic and clinical efficiency of use of NIT.