Machine-learning prediction of myocardial viability using speckle tracking echocardiography: a PET-validated study

Background Artificial intelligence (AI) and machine learning (ML) are increasingly transforming cardiovascular imaging by integrating complex data into clinically useful predictions. Echocardiographic myocardial strain has emerged as a promising marker of myocardial viability; however, the application of AI to strain-based viability assessment remains limited. Purpose To develop and validate ML models for predicting myocardial viability based on echocardiographic regional longitudinal strain (rLS), using PET-defined viability as the reference standard. Methods We retrospectively evaluated 56 patients (age 65.0 ± 7.9 years, 90% male, LVEF 47.1 ± 12.4%) with recent myocardial infarction (30 patients) or ischemic cardiomyopathy (26 patients). All patients underwent transthoracic echocardiography and PET imaging using either ¹⁵O-water or ¹¹C-acetate tracers. Clinical and laboratory variables, including NT-proBNP and culprit vessel/occlusion site, were collected. Myocardial viability was defined at the segment level using tracer-specific PET criteria and classified as viable (1) or non-viable (0). In patients imaged with ¹⁵O-water PET, viability was defined as a perfusable tissue fraction ≥0.60 [1], while in patients imaged with ¹¹C-acetate PET, viability was defined as preserved tracer uptake (k1 ≥0.5 mL/min/g) [2]. Segment-level analyses were performed, with stratified patient-level data splitting (70% training, 30% testing) to prevent data leakage. Incremental logistic regression models were constructed using rLS alone, with sequential addition of LVEF, NT-proBNP, and angiographic occlusion site. Multiple ML algorithms were evaluated, and the top five were combined into an ensemble model. Expert assessment was performed using segmental wall motion score (WMS), with preserved/mildly reduced motion (score ≤2) considered viable. Results A total of 807 viable and 128 non-viable myocardial segments were analyzed. Supervised ML models including rLS alone or in combination with LVEF and NT-proBNP demonstrated good discrimination for myocardial viability on the test set (AUC 0.92, 95% CI 0.88–0.96), whereas addition of coronary occlusion site did not improve discrimination AUC (0.87, 95% CI 0.80–0.94) (Figure 1). Among the top five supervised ML models using rLS and LVEF as input features, logistic regression achieved the highest discrimination.followed by Gaussian NB and Gradient Boosting (Figure 2). The ensemble model showed good overall performance. In comparison, expert WMS-based assessment demonstrated lower discrimination for PET-defined viability (AUC 0.70, 95% CI 0.66–0.73) than ML models. Conclusion Echocardiographic rLS combined with LVEF enables robust prediction of PET-defined myocardial viability. ML-based models demonstrated higher discrimination than conventional WMS assessment and may provide objective decision-support in clinical viability assessment. External validation in larger cohorts is warranted.