In a recently published study in the journal Nature Medicine, researchers developed and tested an intelligent model for diagnosing patients with occlusion myocardial infarction (OMI).
This model was instrumental in diagnosing patients without ST-elevation in their electrocardiograms (ECGs), which practicing clinicians and current commercial interpretation systems may miss. Using data from over 7,000 patients, the model was able to correctly reclassify one in three diagnosis errors made by conventional risk stratification systems.
Study: Machine learning for ECG diagnosis and risk stratification of occlusion myocardial infarction. Image Credit: vchal / Shutterstock.com
Challenges in the timely diagnosis of OMI
ST is a segment of a heart scan trace that, when elevated, a condition otherwise known as ST-segment elevation (STE) indicates an acute coronary syndrome (ACS).
ST-elevation myocardial infarction (STEMI) accompanying acute chest pain is widely considered one of the more severe and life-threatening types of heart attack and requires immediate emergent catheterization. Thus, the accurate interpretation of ECG readings is crucial for the timely treatment of OMI.
Recent studies have shown that not all patients reporting chest pain can access an on-demand ECG. Even in the presence of ECG reports, 24-35% of patients, for unknown reasons, are experiencing OMI, which is not accompanied by STEMI, thereby resulting in misdiagnosis.
Biomarker-based diagnosis is limited, as its interpretation is highly variable and based on visual interpretation by clinicians. This cumulatively results in diagnostic and treatment delays that inevitably lead to higher mortality rates in OMI patients.
Biomarkers that can detect OMI, even in the absence of STEMI, are also limited, as they are only detectable after OMI peak levels have been achieved, by which time the critical period to salvage myocardium has expired. Over 60% of patients admitted to hospitals because of chest pain have inconclusive initial assessment reports, resulting in an estimated 14-22% increased mortality in ACS patients.
About the study
The present study builds upon the authors' previous work, wherein they developed prototype artificial intelligence (AI) algorithms for ECG analysis for automated ACS screening before hospital admission. This study presents the first observational cohort study wherein machine learning diagnostic accuracy was evaluated for use in STEMI diagnosis and risk evaluation.
The study cohort included 7,313 patients with reports of chest pain. Patients were between 43-75 years of age, 47% of whom were female, and 5.2% were eventually found to be OMI positive.
The study cohort was divided into two groups: the derivation group, with 4,026 individuals, and the validation group, with 3,287 patients. While patients in both groups were similar in age, sex, and 30-day cardiovascular mortality, the validation cohort was selected to have increased Black and Hispanic representation and a slightly increased prevalence of ACS and OMI.
The AI model was trained using 12 pre-hospital reports for each of the 4,026 derivation patients. The model identified 554 spatiotemporal metrics, 73 of which were chosen after incorporating recommendations from domain experts. These metrics were used to develop ten classifiers to distinguish between ACS and non-ACS patients and elucidate the likelihood of ACS patients having OMI.
One of the models, the random forest (RF) model, was chosen for testing the validation cohort, as it outperformed currently available commercial ECG systems and practicing clinicians in preliminary testing. The final model development step involved defining a risk metric called the OMI score to categorize patients into low-, medium-, and high-OMI risk groups.
The model was then tested using data derived from the validation cohorts.
Our model generalized well and maintained high classification performance, outperforming the commercial ECG system and practicing clinicians.”
OMI classification identified 74.4% of the 3,287 patients as having low OMI risk, as indicated by a score of less than five. Comparatively, 21% of this cohort were identified as having an intermediate OMI risk with a score between five and 20, whereas 4.6% had a high OMI risk with a score exceeding 20.
When only the OMI score was used, the model significantly outperformed the previously gold-standard HEART metric, the latter of which uses a combination of age, environmental risk factors, troponin values, ECG data, and patient medical history.
Model diagnostic accuracy was consistent irrespective of sex, comorbidities, age, race, or baseline ECG readings, thus demonstrating its lack of aggregation bias. The model also discovered ECG variables that would otherwise be ignored in clinical guidelines as indicative of future OMI onset, thus advancing the researchers' understanding of ACS.
In the present study, researchers developed and validated machine learning AI models for potential OMI patients' clinical diagnosis and risk assessment. This model accurately classified patients into ACS and non-ACS groups, with ACS patients further classified as having a low, intermediate, and high risk of impending OMI.
This model outperforms currently available commercial metrics and practicing clinicians in OMI risk assessment, even in the absence of STEMI in patients' ECG reports. Furthermore, the model identified 73 key detriments of OMI risk, some of which had been largely ignored in clinical diagnostic recommendations.
The clinical implications of this study are numerous, as it can aid clinicians in the real-time evaluation of ECG reports and reduce visual error/bias by practitioners.
Until now, clinicians never had sensitive or highly specific tools that would allow the ultra-early identification of OMI in the absence of a STEMI pattern.”
This type of model can help rapidly assess patient risk, thereby allowing for timely medical intervention and, as a result, significantly reducing mortality in patients with OMI.
- Al-Zaiti, S. S., Martin-Gill, C., Zègre-Hemsey, J. K. et al. (2023). Machine learning for ECG diagnosis and risk stratification of occlusion myocardial infarction. Nature Medicine. doi:10.1038/s41591-023-02396-3