U-Michigan AI Improves Cardiac Diagnosis Across 12 Tasks

Venkatesh L. Murthy, M.D., Ph.D., and his research team at the U-M Health Frankel Cardiovascular Center and U-M Medical School have developed an artificial intelligence model capable of detecting coronary microvascular dysfunction (CMVD) using standard electrocardiograms. The model was trained utilizing data from PET myocardial perfusion imaging—the current “gold standard” for CMVD diagnosis—and a self-supervised learning approach on over 800,000 unlabeled EKG waveforms. This innovation significantly outperforms earlier AI models across nearly every diagnostic task, potentially enabling accurate CMVD identification—a condition often missed in emergency department visits—from a simple 10-second EKG strip. AI Model Detects Coronary Microvascular Dysfunction Researchers at the University of Michigan have developed an AI model designed to detect coronary microvascular dysfunction (CMVD) using standard electrocardiograms (EKGs). This is significant because CMVD requires advanced—and often inaccessible—imaging like PET myocardial perfusion imaging for diagnosis. The model was trained using a self-supervised learning approach, initially analyzing over 800,000 unlabeled EKG waveforms before being refined with labeled PET scan data, allowing it to predict myocardial flow reserve—a key indicator of CMVD. The AI model demonstrated improved diagnostic accuracy across various cardiac prediction tasks, even surpassing previous EKG-AI models. Notably, performance gains were minimal when using EKGs taken during exercise stress tests versus resting EKGs. This suggests the model can effectively identify CMVD using a readily available, simpler diagnostic tool, potentially benefiting hospitals with limited resources or those lacking specialty imaging centers. This new approach addresses a critical gap in CMVD diagnosis, as symptoms can mimic other heart conditions and angiograms may appear “clear” despite the presence of the disease. By predicting CMVD using EKGs, clinicians can more easily identify patients who would benefit from advanced testing, leading to earlier intervention and potentially reducing the risk of heart attack. The study received funding from multiple National Institutes of Health grants and the Department of Veterans Affairs. How the AI Model Was Trained Researchers trained the AI model using a self-supervised learning (SSL) approach, beginning with a deep learning vision transformer pre-trained on over 800,000 unlabeled EKG waveforms. This foundation allowed the model to “understand” the electrical language of the heart without direct human labeling. Subsequently, the model was fine-tuned on a smaller, labeled dataset derived from PET scans—considered the “gold standard” for diagnosing coronary microvascular dysfunction (CMVD). The model’s capabilities extended beyond simply predicting CMVD; it was trained to accurately analyze data across 12 different demographic and clinical prediction tasks. This included tasks not achievable with current EKG-AI models. Importantly, the research showed only minimal improvement in diagnostic accuracy when utilizing EKGs taken during exercise stress tests, compared to resting EKGs, suggesting the model’s core functionality isn’t reliant on strenuous activity data. By leveraging the less accessible PET scan data, the model aimed to extend the predictive power of a standard EKG to identify the more difficult-to-detect CMVD. This is significant because CMVD can be missed in emergency department visits and often appears as a “clear” angiogram, potentially delaying crucial treatment for the approximately 14 million people presenting with chest pain annually. AI Performance with EKG and PET Scans Researchers developed an AI model to diagnose coronary microvascular dysfunction (CMVD) using standard electrocardiograms (EKGs). This is significant because CMVD requires advanced—and often inaccessible—PET myocardial perfusion imaging (MPI) for diagnosis. The model was trained using a “self-supervised learning” approach, initially analyzing over 800,000 unlabeled EKG waveforms before being refined with labeled PET scan data. This allows for potential diagnosis even in hospitals lacking PET scan capabilities. The AI model demonstrated improved diagnostic accuracy across multiple cardiac prediction tasks compared to previous EKG-AI models. It successfully predicted CMVD across different databases and, notably, showed minimal performance increase when utilizing EKGs taken during exercise stress tests versus resting EKGs. The model’s ability to predict myocardial flow reserve is crucial, as this measurement is the “gold standard” for diagnosing CMVD, a condition often missed in emergency department visits. This new technology is particularly valuable because approximately 14 million people annually seek care for chest pain. While angiograms may appear clear in patients with CMVD, this AI model offers a cost-effective and non-invasive way to identify those who could benefit from further, advanced testing. The research team believes this model extends the predictive capabilities of a standard EKG to identify a more challenging-to-detect microvascular condition. Our model creates a way for clinicians to accurately identify a condition that is notoriously hard to diagnose – and often missed in emergency department visits – using a 10-second EKG strip. Venkatesh L. Murthy, M.D., Ph.D. Potential Impact on Cardiac Diagnosis An AI model developed at the University of Michigan shows promise in diagnosing coronary microvascular dysfunction (CMVD), a complex heart condition often missed in emergency department visits. This model utilizes standard electrocardiograms (EKGs) to predict myocardial flow reserve, considered the “gold standard” for CMVD diagnosis, outperforming previous AI models. The technology could be especially valuable in hospitals lacking advanced imaging like PET scans, offering a cost-effective and non-invasive way to identify patients needing further evaluation. The AI model was trained using a novel approach called self-supervised learning (SSL), first learning to ‘understand’ the electrical language of the heart from over 800,000 unlabeled EKG waveforms. It was then refined using a smaller dataset of PET scans. This allows the model to not only predict CMVD accurately across different databases, but also improve diagnostic accuracy for more common cardiac conditions. Researchers found minimal performance gains when using EKGs taken during exercise stress tests compared to resting EKGs. This technology addresses a significant diagnostic challenge, as approximately 14 million people annually visit emergency rooms or clinics with chest pain, and CMVD can be difficult to detect. Currently, CMVD diagnosis requires expensive PET scans, which aren’t widely accessible. By leveraging readily available EKG data, the AI model aims to identify patients who would benefit from these advanced tests, potentially improving outcomes and reducing missed diagnoses of a serious, but often overlooked, heart condition. Source: https://www.michiganmedicine.org/health-lab/ai-model-helps-diagnose-often-undetected-heart-disease-simple-ekg Tags: