Machine learning versus guideline recommended pathways for the diagnosis of myocardial infarction

Abstract Background CoDE-ACS (Collaboration for the Diagnosis and Evaluation of Acute Coronary Syndrome) is a validated clinical decision-support tool that uses machine learning to calculate the probability of myocardial infarction (MI) for an individual patient. Current European guidelines recommend the use of 0/1h- and 0/2h-algorithms optimised to rule-out or rule-in MI and the High-STEACS pathway, which enables rapid discharge or admission in patients with suspected MI. However, it is unknown how CoDE-ACS performs against these guideline recommended pathways. Methods We enrolled adult patients presenting with suspected acute coronary syndromes at 12 sites in 5 European countries. Patients with ST-elevation MI and those with missing high-sensitivity (hs)-cTnI concentrations at 0h, 1h or 2h were excluded. Final diagnoses were adjudicated by two independent cardiologists according to current guidelines and the 4th universal definition of MI using all information including serial hs-cTnI measurements and imaging. We compare the diagnostic performance and effectiveness of CoDE-ACS with the 0/1h- and 0/2h-algorithms and the High-STEACS pathway. Results In 4,105 patients (median age 61 [IQR 50-74], 32% women), the prevalence of type 1 MI was 14%. At presentation, CoDE-ACS identified 56% (2,280/4,105) of patients as low-probability, with a 99.7% (99.5-99.9%) negative predictive value (NPV) and 99.0% (98.6-99.2%) sensitivity, which ruled out twice as many patients as other pathways. CoDE-ACS incorporating 1h hs-cTnI measurements identified 65% (2,671/4,105) patients as low-probability with a 99.7% (99.5-99.8%) NPV and 98.6% (98.2-98.9%) sensitivity and 19% (770/4,105) patients as high-probability with a 64.9% (63.5-66.4%) PPV and 92.4% (91.5-93.1%) specificity, with 16% classified as intermediate probability. In comparison, the 0/1h-algorithm identified 49% (2,014/4,105) patients as low-risk with a 100% (99.9-100%) NPV and 100% (99.9-100%) sensitivity and 20% (829/4,105) patients as high-risk with a 61.5% (60.0-63.0%) PPV and 91.0% (90.0-91.8%) specificity, with 31% triaged to the observe zone. Findings for CoDE-ACS incorporating hs-cTnI measurements at 2h, the 0/2h-algorithm and the High-STEACS pathway were comparable (Table). Conclusion All guideline recommended clinical pathways provide excellent performance for the diagnosis of MI. CoDE-ACS provides comparable diagnostic performance, but identifies twice as many patients as low-probability at presentation and fewer patients require further observation following serial hs-cTnI measurements. Implementation studies are warranted to determine whether CoDE-ACS can further improve the early diagnosis of MI.