Leveraging machine learning for rapid rule-out of acute coronary syndrome

Abstract Background Current Acute Coronary Syndromes (ACS) rule-out algorithms rely on a combination of clinical assessment and measuring troponin levels. It can take several hours for troponin levels to rise after a myocardial infarction, so initial testing may not show detectable levels of troponin. In order to rule out a false negative result, troponin levels are typically tested again several hours later to look for rising values meaning patients are admitted for observation which has a large resource implication. Purpose We developed a machine learning model aimed at improving early discharge at initial assessment. Methods The study was conducted using data from the National Institute for Health Research Health Informatics Collaborative Cardiovascular dataset.(1,2) We trained and tuned a machine learning model (Rapid-RO) using patient data from two separate hospitals to rule-out ACS with simple routine demographic or clinical measurements. The model was then tested for its predictive accuracy in cohorts of patients at four different hospitals from separate time periods. The model was assessed against troponin threshold guided management as recommended by the European Society of Cardiology clinical guidelines. Results The patient cohorts of the six derived datasets are presented in Figure 1. On the left side are the training and tuning cohorts and the right side are the cohorts from which the derived model was tested. The Rapid-RO machine learning model included input from 11 inputs that had the highest feature importance, including troponin, age, C-reactive protein, urea, platelet count, eGFR, white cell count, haemoglobin, heart failure, diabetes, and hypertension. The Rapid-RO model identified 12037 (35.69%) very low risk patients on top of standard clinical assessment who could have been discharged early, compared with 8967 (26.58%) identified by a troponin threshold approach alone (Figure 2), with significantly fewer missed ACS cases (27 (0.22%) vs. 108 (1.20%)) and similar mortality rates (2 (0.02%) vs. 4 (0.04%) at 30 days). The Rapid-RO model demonstrated a consistently higher rule-out rate for ACS with a lower missed ACS rate across patient subsets, including patients with and without chest pain or COVID-19. Conclusion The Rapid-RO machine learning model, which uses patient history and initial blood tests, offers a significant advancement in the risk stratification process, presenting a reliable tool for clinicians to rapidly rule out ACS and potentially reduce unnecessary hospital admissions. Its robust performance in diverse patient groups across different time periods, underscores its potential utility in a real-world clinical setting.Figure 1 Figure 2