Deep learning-derived regional strain: an effective echocardiographic predictor of LAD occlusion in revascularized patients

Background Strain echocardiography provides valuable insights into myocardial function, but current approaches for measuring global longitudinal strain (GLS) and regional longitudinal strain (RLS) remain time-consuming and operator-dependent(1). These limitations hinder efficiency and reproducibility, impeding objective assessment of regional pathologies, such as in myocardial infarction. Recent advances in deep learning (DL) now enable fully automated and highly reproducible RLS quantification(2). This may enable more consistent detection of ischaemia and enhance clinical decision-making and patient outcomes in coronary artery disease. Purpose This external validation study compares the diagnostic accuracy of DL-derived RLS with conventional semi-automatic measures, including RLS, GLS, and ejection fraction (EF), in predicting total occlusion of the left anterior descending (LAD) as confirmed by invasive angiography. Methods We included 103 patients (mean age 62 ± 11 years, 73% male) with prior coronary revascularization and available conventional semi-automatic echocardiographic parameters (EF, GLS, RLS, analysed in vendor specific software). An in-house, internally validated DL algorithm analysed the same echocardiograms to produce automated DL RLS measurements. The predictive performance of each echocardiographic metric for total LAD occlusion was assessed via receiver operating characteristic (ROC) curves and odds ratios. Optimal cut-offs were determined using the Youden Index. Statistical significance was defined as p < 0.05. Total LAD occlusion was confirmed by angiography in 40.8% (n = 42) of the 103 revascularized patients. The fully automated measurements of DL RLS demonstrated high predictive performance, with an AUC of 0.812 (95% CI: 0.722-0.802), similar to the conventional RLS with AUC 0.875 (95% CI: 0.801-0.936) (Figure 1). At optimal cut-offs (-12.6% vs -13.5%), DL RLS showed sensitivity/specificity of 79%/74% compared to 84%/78% for conventional RLS. GLS and EF showed lower predictive performance with an AUC of 0.789 (95% CI: 0.697-0.874) and 0.716 (95% CI: 0.612-0.813) respectively (Table 1). Conclusion Deep learning-derived RLS offers a fully automated, reproducible alternative to conventional semi-automatic RLS, with comparable predictive accuracy and the potential to streamline clinical workflows for echocardiographic assessment of coronary artery disease.Comparison of ROC curves Table 1:Results