AI-based diagnostic model for functional myocardial ischemia using resting echocardiographic strain analysis

Background

Coronary functional assessment is crucial for identifying ischemic patterns and guiding revascularization, but current techniques have limitations. This study aimed to develop a noninvasive AI-based diagnostic model for functional myocardial ischemia using resting echocardiographic strain analysis.

Purpose

This study aimed to develop a noninvasive artificial intelligence (AI)-based diagnostic model for functional myocardial ischemia using resting echocardiographic strain analysis.

Methods

Patients with suspected coronary artery disease (CAD) who underwent echocardiography (UCG) and coronary angiography (CAG) at our University First Hospital (2019–2021) were enrolled. Exclusions included prior cardiac surgery, segmental wall motion abnormalities, severe myocardial hypertrophy, or poor image quality. Coronary functional parameters (angiography-derived FFR [caFFR] and IMR [caIMR]) were measured via the FlashAngio system. Segmental strain and global longitudinal strain (GLS) were calculated using a self-developed automated left ventricular endocardial tracing tool.

Myocardial ischemia (per-vessel: LAD, LCX, RCA) was defined as follows: 1) Epicardial ischemia: caFFR ≤0.80; 2) Isolated microvascular ischemia: caFFR >0.80 with caIMR >25; 3) Overall ischemia: caFFR ≤0.80 and/or caIMR >25. Non-ischemia was defined as caFFR >0.80 and caIMR ≤25. Inputs included 17-segment strain, GLS, conventional UCG parameters, demographic data, and medical history. Three ML models were constructed for overall, epicardial, and microvascular ischemia, with receiver operating characteristic (ROC) curves and diagnostic indices calculated.

Results

A total of 1,207 patients (61.14% male, mean age 64.45 years) were included: 45.82% had macrovascular ischemia, 27.92% had isolated microvascular ischemia, and 26.26% were non-ischemic. The cohort was split into a training set (2019–2020, n=969, 80.28%) and a testing set (2021, n=238, 19.72%). In the test set: Model A (overall ischemia): Accuracy 79.8%, sensitivity 88.3%, specificity 58.2%, AUC 0.832; Model B (epicardial ischemia): Accuracy 76.9%, sensitivity 67.7%, specificity 82.8%, AUC 0.841; Model C (microvascular ischemia): Accuracy 80.7%, sensitivity 78.2%, specificity 82.8%, AUC 0.879.

Conclusion

Multimodal AI models integrating resting echocardiographic strain and clinical data demonstrate robust diagnostic performance for distinct types of myocardial ischemia, providing a non-invasive, intelligent tool for precision management. Multi-center prospective validation is currently underway to further confirm the models’ clinical applicability and generalizability.Central IllustrationROC curve of Model A