Shanghai Tenth Hospital’s AI System Excels in ECG Interpretation “Simulated Exams” and “Final Exams”

    A groundbreaking AI system developed by cardiovascular experts at Shanghai Tenth People’s Hospital has achieved remarkable success in interpreting electrocardiogram (ECG) waveforms, particularly in identifying the culprit vessels in ST-segment elevation myocardial infarction (STEMI) patients. The system, developed by Dr. Zhang Yi and Dr. Zhao Yifan’s team, has demonstrated exceptional accuracy, especially in detecting blockages in the left circumflex artery (LCX), which is notoriously difficult to diagnose. The findings were recently published in BMJ Digital Health & Artificial Intelligence, a sub-journal of the renowned British Medical Journal. 

    The adage “Time is muscle, time is life” underscores the critical importance of early detection and intervention in heart attacks. The sooner the symptoms are identified, the culprit vessel is pinpointed, and medical treatment is initiated, the better the patient’s prognosis. However, accurately locating the blocked vessel using ECG waveforms has historically been challenging due to variations in coronary anatomy, collateral circulation, and differences in clinicians’ experience. This new AI system promises to revolutionize emergency decision-making by significantly reducing diagnosis time, optimizing treatment workflows, and improving patient outcomes. 

    Unlike general-purpose AI models, this specialized system is tailored for ECG waveform analysis. It was trained on a vast dataset of high-quality diagnostic data, enabling it to excel in speed, accuracy, sensitivity, and specificity. The team collected 2,957 original ECG records, using coronary angiography as the gold standard for diagnosis. From these, 698 ECG records with clearly identified culprit vessels were selected as the “high-quality textbook” and “real-world exam” for training and testing the AI.  

    The dataset was divided into three parts: 543 records from two top-tier hospitals were split into an internal training set (the “textbook”) and an internal test set (the “simulated exam”) in a 4:1 ratio, while 155 records from another top-tier hospital formed the external validation set (the “final exam”). After rigorous training and testing, the AI system demonstrated outstanding performance: 

    - In the “textbook” (internal training set), the algorithm achieved sensitivities of 92.4%, 93.2%, and 99.7% for identifying blockages in the left anterior descending artery (LAD), right coronary artery (RCA), and LCX, respectively, with specificities of 99.7%, 97.4%, and 95.8%. 

    - In the “simulated exam” (internal test set), it achieved sensitivities of 91.6%, 75.1%, and 97.0% for the same vessels, with specificities of 96.0%, 95.8%, and 88.8%, outperforming both cardiologists and existing ECG algorithms. 

    - In the “final exam” (external validation set), the system maintained high performance, with sensitivities of 72.0%, 90.5%, and 92.9% and specificities of 94.3%, 92.4%, and 91.2%. 

Dr. Zhang Yi highlighted the system’s potential in four key areas: 

1. Wearable Integration: The AI can be embedded in pre-hospital emergency devices or wearable ECG monitors, providing early warnings and preliminary vessel localization for high-risk patients. 

2. Efficiency Enhancement: By identifying the culprit vessel during patient transport, the system allows hospitals to better prepare for procedures, reducing the time from admission to coronary reperfusion. 

3. Precision Improvement: It addresses the clinical challenge of detecting LCX blockages, reducing missed diagnoses and errors. 

4. Empowering Grassroots Healthcare: In resource-limited or remote areas with less experienced cardiologists, the system can serve as a powerful diagnostic aid, improving overall STEMI treatment standards. 

This breakthrough represents a significant step forward in leveraging AI to enhance cardiovascular care, offering hope for faster, more accurate, and accessible heart attack diagnosis and treatment.