In a recent article published in the journal Nature Medicine, researchers discuss a novel clinical decision support system based on machine learning (ML) models to predict an individual’s probability of developing myocardial infarction. This system generates Collaboration for the Diagnosis and Evaluation of Acute Coronary Syndrome (CoDE-ACS) scores, which utilized data from the High-Sensitivity Troponin in the Evaluation of Patients with Suspected Acute Coronary Syndrome (High-STEACS) trial population.
The CoDE-ACS integrated cardiac troponin concentrations at presentation or on serial testing with clinical features as a continuous measure to calculate a score between zero to 100. This score reflects how likely it was for an individual to later develop acute myocardial infarction. CoDE-ACS was trained separately but sequentially on consecutive patients with and without myocardial injury at presentation.
For comparison, the researchers used conventional diagnostic pathways based on sex-specific 99th percentile cardiac troponin I thresholds, which also ruled out or predicted an individual’s probability of myocardial infarction.
Study: Machine learning for diagnosis of myocardial infarction using cardiac troponin concentrations. Image Credit: TippaPatt / Shutterstock.com
Conventional cardiac troponin assays are highly sensitive for diagnosing acute myocardial infarction in symptomatic patients. Both national and international clinical practice guidelines recommend these assays; however, they are associated with certain limitations.
For example, these assays use fixed troponin thresholds for all patients that are changeable with age, comorbidities, and gender of the patient. Furthermore, these assays do not account for electrocardiogram (ECG) results or time of symptom onset.
It is also tedious to consistently apply specific time points for serial testing in busy hospital emergency departments (EDs), which is a requirement when performing these assays. Additionally, the 99th percentile diagnostic threshold of cardiac troponin assays is inconsistent across age-, gender-, and comorbidities-based groups.
About the study
In the present study, researchers first trained the ML models-based CoDE-ACS system on derivation cohorts data of 10,038 patients with possible myocardial infarction. The average age of the study cohort was 70 years, 48% of whom were female. All patients had an adjudicated type 1, 4b, or 4c myocardial infarction without ST-segment elevation during the first hospital admission.
Two clinicians independently reviewed all investigations made by CoDE-ACS according to the fourth universal definition of myocardial infarction. A third reviewer resolved any reported disagreements between these two clinicians.
The diagnostic performance on data from 10,286 patients was externally validated from seven cohorts. This allowed the researchers to compare this data with diagnostic methods currently used in clinical practice to elucidate its clinical relevance. Across external healthcare systems, the prevalence of myocardial infarction varied from 4% to 16%.
In patients with no myocardial injury at presentation as compared to those with myocardial injury, a CoDE-ACS score of less than three and 61 or more met the prespecified diagnostic performance criteria, respectively.
Patients without a history of myocardial infection had a negative predictive value of 99.5 and a sensitivity of 90.2. Comparatively, those with a history of myocardial infarction had a positive predictive value of 80.1 and a specificity of 83.4. CoDE-ACS scores performed consistently across all subgroups.
CoDE-ACS successfully discriminated myocardial infarction, as indicated by the area under the curve (AUC) value of 0.953. Furthermore, CoDE-ACS identified 61% of patients at presentation as having a low probability of developing myocardial infarction. This is comparable to the 27% of patients identified when fixed cardiac troponin thresholds with a comparable negative predictive value were used.
CoDE-ACS scores also allowed researchers to identify fewer patients with a high probability of developing acute myocardial infarction with a greater positive predictive value.
Patients with a low probability of myocardial infarction were at low risk of death following discharge. Accordingly, fewer than one in 300 of these individuals suffered cardiac death one year after symptom presentation.
Patients with an intermediate or high probability of myocardial infarction were at higher risk of cardiac death within 30 days and one year from presentation, respectively.
The CoDE-ACS pathway, a novel ML models-based clinical decision support system, incorporated information on the time of testing, serial measurement of cardiac troponin levels at a flexible time point, and time from symptom onset.
Current diagnostic methods require that a patient presents within three hours of symptom onset following an episode of myocardial ischemia for cardiac troponin measurements. Comparatively, CoDE-ACS ruled out myocardial infarction, even in early presenters, using a single cardiac troponin test, thereby reducing the damage due to nonadherence with the measurement timing. In this study, CoDE-ACS ruled out myocardial infarction in 71% of patients with a single test.
In a conservative healthcare system, a lower CoDE-ACS score could identify patients with a very low probability of myocardial infarction. This system’s false-negative rate is one in 500, thus implying that it could guide clinicians to discharge nearly half of the patients with a single test. However, a lower CoDE-ACS score could also identify those with a higher probability of myocardial infarction to decrease the proportion of patients who need observation and serial testing within the ED.
Thus, using CoDE-ACS, clinicians could create an optimal pathway for patient flow according to regional clinical priorities. In the future, this system could be integrated with the 12-lead electrocardiogram, another ML approach, to refine performance and reduce the proportion of patients in need of attention.
Given the flexibility of CoDE-ACS, the authors advocate for its adoption in clinical practice to reduce unnecessary hospital admissions of patients unlikely to have myocardial infarction. Furthermore, this system could allow clinicians to shift their focus to those at a higher risk of cardiac death.
- Doudesis, D., Lee, K.K., Boeddinghaus, J. et al. (2023). Machine learning for diagnosis of myocardial infarction using cardiac troponin concentrations. Nature Medicine. doi:10.1038/s41591-023-02325-4