Machine learning for diagnosing non-ST-segment elevation myocardial infarction: a derivation and validation study
Arnaud Champetier1, Pedro Lopez-Ayala1, Christoph Reich2
1Cardiovascular Research Institute Basel (CRIB) and Department of Cardiology, University Hospital Basel, University of Basel, Basel, Switzerland.
Background:
Previously developed machine-learning (ML)-based decision support tools for patients presenting with suspected non-ST-segment elevation myocardial infarction (NSTEMI) remain proprietary, limiting public accessibility and clinical adoption.
Methods:
To address this limitation, we used two international prospective multicentre diagnostic studies for model derivation and internal validation, and one large prospective European study for external validation, evaluating two open-access ML-based models. The derivation and internal validation cohort comprised 8763 patients (34% women) enrolled across 13 and 12 sites, respectively, in Switzerland, Spain, the Czech Republic, Poland, Belgium, Germany, the UK, Italy and the USA between April 2006 and September 2020, and between August 2011 and June 2013. The external validation cohort included 4882 patients (41% women) from Germany. Patients were excluded if they had a ST-segment elevation myocardial infarction, unclear final diagnosis, renal failure or an absent 12-lead electrocardiogram. A single-high-sensitivity cardiac troponin (hs-cTn) model incorporated information available at emergency department presentation, while a serial-hs-cTn model additionally utilised the second hs-cTn measurement and the time interval between samples. The final diagnosis of NSTEMI was centrally adjudicated by two independent cardiologists in all studies. The diagnostic performance of both models was compared with the European Society of Cardiology (ESC) hs-cTn-0/1 h-algorithm. This study is registered with ClinicalTrials.gov, numbers NCT00470587 and NCT03111862.
Findings:
The single-hs-cTn- and serial-hs-cTn model demonstrated excellent discrimination in internal validation, and external validation, with an area under the receiver-operating-characteristic curve of 0.94 [0.92-0.95], and 0.91 [0.90-0.92], and 0.96 [0.96-0.97], and 0.96 [0.95-0.97], respectively. Calibration was good across all datasets. Compared to the ESC-0/1 h algorithm, FAST-NSTEMI provided comparable safety metrics while triaging more patients to rule-out or rule-in. Triage efficacy improved substantially with the single-hs-cTn model (internal validation 52.4% versus 29.8%, external validation 32.1% versus 14.9%) and modestly with the serial-hs-cTn model (internal validation 80.8% versus 76.9%, external validation 77.1% versus 72.8%, all p < 0.01).
Interpretation:
The FAST-NSTEMI ML-models offer excellent discrimination, good calibration, high safety, and improved triage efficacy compared to the ESC 0/1 h-algorithm. Further external validation or prospective implementations are warranted to confirm the generalisability of the findings and the clinical utility of the models.
Funding:
Swiss National Science Foundation and Swiss Heart Foundation.