interpretable deep learning model for uncertainty-based left ventricle ejection fraction estimation in echocardiography

Introduction: ejection fraction (ef) is a critical metric in diagnosing and managing heart diseases, providing essential insights into cardiac function. recently, deep learning models have emerged to evaluate ef from echocardiographic videos, demonstrating promising accuracy. however, these models often lack clinical explainability, limiting their integration into clinical workflows. additionally, existing approaches predominantly focus on point estimation of ef, whereas range estimation would better align with clinical needs by offering uncertainty quantification. this study presents a deep learning model that not only estimates ef within a confidence range but also incorporates frame-level and pixel-level explainability to enhance clinical trust and usability.methods: we utilized the camus dataset, comprising echocardiographic videos, which were divided into 320 training samples, 40 validation samples, and 40 test samples. the proposed model integrates a convolutional neural network (cnn) with a recurrent neural network (rnn) architecture, along with an attention mechanism to identify the most relevant frames contributing to ef estimation. to provide range estimates for ef, we employed conformal prediction, a statistical technique that quantifies prediction uncertainty by constructing confidence intervals. for interpretability, gradient-weighted class activation mapping (grad-cam) was applied to visualize pixel-level importance, highlighting the specific regions within frames that influence the model's predictions.results: our deep learning model achieved an 87.5% coverage probability on the test set, indicating that 87.5% of true ef values fell within the predicted confidence intervals. the mean interval width was 12%, demonstrating a balance between prediction reliability and precision. attention maps effectively pinpointed key frames corresponding to significant phases of the cardiac cycle, while grad-cam visualizations revealed that the model focused on anatomically and functionally relevant regions, such as the left ventricular cavity and myocardial walls, during ef estimation.conclusions: integrating deep learning models into clinical settings necessitates not only high predictive performance but also robust interpretability to foster clinician trust and facilitate informed decision-making. our study advances ef evaluation by providing range estimates accompanied by detailed frame and pixel-level explanations, bridging the gap between ai-driven predictions and clinical relevance. this approach underscores the importance of developing interpretable models that align with clinical workflows, ensuring that ai tools can be effectively and confidently utilized in cardiovascular care.