Comprehensive Summary
Ueda and colleagues developed and validated a deep learning model to detect mitral regurgitation (MR) using standard posteroanterior chest radiographs from patients who had undergone echocardiography at a single secondary care hospital in Japan. Chest radiographs obtained within 30 days of echocardiography between 2016 and 2019 were labeled MR-positive (mild–severe) or MR-negative based on echocardiographic reports and split on a patient level into training (8240 images from 4216 patients), validation (1073 images from 527 patients), and test (1054 images from 527 patients) datasets. A ResNet-50 convolutional neural network was trained from scratch with data augmentation, and performance was evaluated on the independent test set. In the test dataset, the model achieved an AUROC of 0.80 (95% CI 0.77–0.82).
Outcomes and Implications
For clinicians, this work suggests that a chest X‑ray–based AI tool could provide a rapid, low-cost, and widely available aid to flag patients with moderate or severe MR, particularly where echocardiography access is limited or patients cannot tolerate prolonged imaging. At the bedside, such a model could be integrated into radiography workflows to prompt earlier echocardiographic evaluation for suspected MR, while its moderate specificity indicates it should be used as a screening or triage aid rather than a standalone diagnostic test pending external, multicenter validation.