Predicting risk of future acute coronary syndromes, 1-year survival and the need for coronary angiography in unstable angina: a nationwide machine learning study

Abstract

Abstract Background We developed machine learning models to predict the need for coronary angiography, recurrent acute coronary syndromes (ACS) and 1-year survival in patients with chest pain and normal high-sensitivity troponins. We also studied whether hs-troponin levels within the normal range convey predictive information on these outcomes. Purpose We studied whether machine learning could reliably predict survival, risk of future ACS and rule out unnecessary angiographies in patients with chest pain and normal hs-troponins. We aimed to deploy these models as open-sourced web applications, to provide clinicians with individualized predictions. We also studied whether normal hs-troponin levels may serve as predictors of these outcomes. Methods We used the SWEDEHEART registry to include patients admitted due to chest pain, with normal high-sensitivity troponin T or I (hs-TnI, hs-TnT), who underwent angiography and did not receive a final diagnosis of acute myocardial infarction. We studied angiographic findings on segmental level, developed machine learning models for future ACS and death (within 1-year, modelled separately) and unnecessary coronary angiography, which was defined as angiography that did not lead to any intervention. Models predicting future ACS and 1-year survival included 130 candidate predictors and models for unnecessary angiography included 110 predictors. We built approximately 50'000 models, using gradient boosting, extreme gradient boosting, random forest, artificial neural networks and logistic regression. Results We included 9'314 patients. The 1-year mortality rate was 0.9% (n=78), rate of future ACS was 2.7% (n=251), and rate of unnecessary angiography was 61.5% (n=5455). Up to 40% had normal angiography. There was a strong association between troponin levels (within normal range) and severity of coronary atherosclerosis; e.g 32.4% in patients with hs-TnI 26–35 ng/L had >50% stenosis in segment 6, as compared with 12.6% in those with hs-TnI 0–5 ng/L. All segments displayed similar associations with troponin levels. Mortality increased at hs-TnI levels above 10 ng/L for men, but not women. Age and sex adjusted hazard ratios for hs-TnI 25–35 vs hs-TnI 0–5 was 5.73 (2.14–15.35) for 1-year mortality. The strongest predictors of 1-year mortality were C-reactive protein, body mass index, estimated glomerular filtration rate, age, time from symptom onset to CCU admission, systolic blood pressure and hs-TnI. Extreme gradient boosting was the best performing model for all prediction tasks; AUC ROC in the test data sets were 0.77 for 1-year mortality, 0.77 for future ACS and 0.78 for unnecessary angiography, with excellent calibration. Conclusion Machine learning models can reliably predict 1-year risk of death or ACS, as well as predict unnecessary angiographies. Troponin levels within normal range constitute a strong predictor of all these outcomes, questioning the definition of normal troponin. Funding Acknowledgement Type of funding sources: Public hospital(s). Main funding source(s): Sahlgrenska University hospital