Improving the Diagnosis of Occlusion Myocardial Infarction with Machine Learning
A recently published study in the journal Nature Medicine introduces an intelligent model that enhances the diagnosis of occlusion myocardial infarction (OMI) using machine learning. This innovative model has the potential to detect OMI in patients without ST-elevation in their electrocardiograms (ECGs), a condition that is often missed by current commercial interpretation systems and practicing clinicians.
Occlusion myocardial infarction (OMI) poses a significant challenge in terms of timely diagnosis. ST-segment elevation (STE) in an ECG is indicative of acute coronary syndrome (ACS), including ST-elevation myocardial infarction (STEMI), which is a severe and life-threatening type of heart attack requiring immediate catheterization. Accurate interpretation of ECG readings is crucial for promptly treating OMI.
However, recent studies have shown that not all patients experiencing chest pain have access to on-demand ECGs. Even when ECG reports are available, a significant portion of patients (24-35%) with OMI do not exhibit ST-elevation, resulting in misdiagnosis. Biomarker-based diagnosis is also limited, as interpretation can vary greatly among clinicians. Consequently, delays in diagnosis and treatment lead to higher mortality rates in OMI patients.
To address this issue, researchers built upon their previous work in developing artificial intelligence (AI) algorithms for automated ACS screening before hospital admission. In this new study, they evaluated the diagnostic accuracy and risk evaluation of AI models using machine learning for STEMI diagnosis.
The study cohort consisted of 7,313 patients reporting chest pain, with ages ranging from 43 to 75 years. Among them, 47% were female, and 5.2% were eventually diagnosed with OMI. The cohort was divided into a derivation group (4,026 individuals) and a validation group (3,287 patients). The validation group aimed to include a higher representation of Black and Hispanic individuals, as well as an increased prevalence of ACS and OMI.
The AI model was trained using 12 pre-hospital reports for each patient in the derivation group. It identified 554 spatiotemporal metrics, of which 73 were selected based on expert recommendations. These metrics were used to create ten classifiers to distinguish between ACS and non-ACS patients and determine the likelihood of OMI in ACS patients.
The random forest (RF) model, one of the ten classifiers, outperformed currently available commercial ECG systems and practicing clinicians in preliminary testing. A risk metric called the OMI score was developed to categorize patients into low-, medium-, and high-risk groups for OMI.
The model was then tested using data from the validation cohort. It exhibited high classification performance and outperformed commercial ECG systems and clinicians. It classified 74.4% of the patients in the validation group as low-risk for OMI, with a score of less than five. Approximately 21% were classified as intermediate risk (score between five and 20), and 4.6% were classified as high-risk (score above 20). When compared to the HEART metric, which combines age, risk factors, troponin values, ECG data, and medical history, the model significantly outperformed the gold standard.
Importantly, the model’s diagnostic accuracy remained consistent across various demographic factors and baseline ECG readings, indicating a lack of bias. It also identified ECG variables that are often overlooked in clinical guidelines but are indicative of future OMI onset, contributing to a better understanding of ACS.
This study has significant clinical implications as it aids clinicians in real-time ECG evaluation, reducing visual errors and biases. Moreover, it enables the ultra-early identification of OMI even without ST-elevation, which was not possible until now. By rapidly assessing patient risk, this model allows for timely medical intervention, potentially reducing mortality among OMI patients.
In conclusion, the study introduces machine learning AI models for the clinical diagnosis and risk assessment of potential OMI patients. The models accurately classify patients into ACS and non-ACS groups and further categorize ACS patients according to their risk of impending OMI. This breakthrough outperforms current commercial metrics and practicing clinicians in assessing OMI risk, even in the absence of ST-elevation in ECG reports. Additionally, the identification of previously overlooked OMI risk factors improves diagnostic recommendations.
