Transfer learning can predict the presence of hibernating myocardium from rest myocardial perfusion images.

Purpose: Hibernating myocardium is a viable but dysfunctional myocardium state caused by chronic ischemia, with potential for recovery postrevascularization. This study evaluates the feasibility of transfer learning for predicting hibernating myocardium from rest myocardial perfusion images.

Methods: Patients who underwent myocardial viability assessment from January 2017 to September 2022 were split into training (70%) and validation (30%) sets, while those from October 2022 to January 2023 formed the testing set. Hibernating myocardium was defined as a mismatched perfusion-metabolism defect with impaired contractility. Rest myocardial perfusion polar maps were embedded using Google's InceptionV3, followed by data normalization and analysis of variance-based feature selection. Three gradient boosting algorithms were trained with stratified 10-fold cross-validation, validated, and tested. Performance was assessed using area under the curve (AUC), classification accuracy (CA), F1 score, specificity, and model interpretability via SHapley Additive exPlanations (SHAP) plots.

Results: The study included 239 patients (214 males, 25 females, mean age 56 ± 11 years); 123 (51.5%) had hibernating myocardium. All models achieved >0.700 in performance metrics across all datasets. Among them, extreme gradient boosting (xgboost) performed best on the test set (F1 score: 0.800, CA: 0.774, specificity: 0.909, AUC: 0.782). Beeswarm SHAP plots revealed a clear pattern of model interpretability for all models.

Conclusion: This study demonstrates the feasibility of transfer learning for predicting hibernating myocardium from rest myocardial perfusion images. The integration of deep convolutional neural networks with gradient boosting models highlights the potential of machine learning-based myocardial viability assessment, contributing valuable early evidence.