Diagnostic Deep Learning Framework for Heart Failure

Abstract

Abstract Background In the precision medicine era, leveraging advanced technology, including deep learning, has the potential to enhance diagnostic accuracy across various diseases. However, chest radiography (CXR), pivotal for heart failure (HF) diagnosis, currently has limited precision. Methods Through a retrospective cohort study encompassing 144 participants from the RAMA dataset at Ramathibodi Hospital (spanning January 1, 2010, to December 31, 2019), 240 HF CXR images were scrutinized and annotated by cardiologists and radiologists. Clinical diagnosis was confirmed by cardiologists using HF signs and symptoms, pulmonary capillary wedge pressure, natriuretic peptide, and ejection fraction. The developed model, HFNet, was trained on the RAMA dataset and incorporated these clinical factors. Results This study evaluated the performance of the HFNet model in predicting radiographic findings related to heart failure and achieved excellent results. The model demonstrated high precision (for cardiomegaly, 1.0; for pulmonary edema, 0.9; for pleural effusion, 0.8) and good accuracy (for cardiomegaly, 0.9; for pulmonary edema, 0.6; and for pleural effusion, 0.7), coupled with respective AUC values of 1.00, 0.96, and 0.49. Concomitant recall figures stood at 0.9, 0.6, and 0.7, while F1 scores were 0.9 for cardiomegaly and 0.7 for both pulmonary edema and pleural effusion. These findings highlight the potential of HFNet to aid clinicians in the precise detection and diagnosis of HF-associated radiographic cues. Conclusions The development of the HFNet model introduces a promising tool for clinicians, facilitating accurate and precise diagnosis of HF-related radiographies.