Transfer learning for echocardiographic detection of heart failure with preserved ejection fraction: preliminary results of TALE-HFpEF Study

Abstract

Abstract Background Heart failure with preserved ejection fraction (HFpEF) is a heterogeneous syndrome with increasing prevalence (1). The diagnosis of HFpEF is a complex one that has not yet reached a consensus in current guidelines, and attempts are being made to diagnose it through various algorithms and scoring systems (2, 3). However, the uncertainties in the diagnostic process and the inherent complexity continue to pose significant barriers to practical implementation. The use of artificial intelligence on single apical 4-chamber transthoracic echocardiograhy video clips for HFpEF detection has shown success (4), but knowledge from readily available models trained for different tasks is not utilized. Purpose This study aims to utilize transfer learning, an artificial intelligence method, to detect HFpEF using echocardiography images. Methods In this preliminary anaylsis, echocardiography video clips were collected from 40 healthy volunteers and 53 HFpEF patients, all over 18 years old. The diagnosis of HFpEF was made in accordance with the current ESC guidelines (3). Apical 4-chamber transthoracic echocardiography images of the patients and volunteers included in the study were obtained and analyzed. Patients with chronic obstructive pulmonary disease, recent myocardial infarction (last 6 months), or recent stroke/cerebrovascular disease (last 3 months) were excluded. Transfer learning was applied using a video ResNet model (6), adapted for left and right ventricle ejection fraction (LVEF and RVEF) prediction tasks, along with a non-medical video classification task (Kinetics 400) (6-8). A 5-fold cross-validation schema was used, and models were compared using balanced accuracy with a right-tailed t-test. Results When comparing with the control group, the HFpEF group shows higher rates of hypertension, diabetes, and atrial fibrillation, as well as higher NT-proBNP levels. The paired one-tailed t-test confirmed significant superiority of all transfer learning models over the baseline model (p < 0.005). The model transferred from the LVEF regression task achieved an AUC of 0.95 ± 0.04 and F1 score of 0.93 ± 0.04 (Figure 1), demonstrating superior performance. Statistical analysis indicated no significant variation in balanced accuracy among models (p > 0.05). Figure 1 also depicts ROC curves of the models initialized with different task weights. Figure-2 illustrates the locations where models focus before and after training using the Grad-CAM method (9). The LVEF model has achieved 92% accuracy in identifying HFpEF patients with 95% sensitivity and 90% specificity. Conclusion The preliminary results of our study are promising in the diagnosis of HFpEF patients through echocardiographic clips with transfer learning. Throughout our study, as the sample size grows, this model could become a key tool in clinical practice for detecting HFpEF patients, potentially enhancing AI's role in diagnosing this challenging patient group.