Automated Classification of Mitral and Tricuspid Regurgitation With Explainability and Real-Worl

An Explainable, Flow‑Aware AI System for Concurrent MR and TR Assessment

• The work addresses joint evaluation of mitral regurgitation (MR) and tricuspid regurgitation (TR) using echocardiography, with an emphasis on explainability and physiologic constraints to support real‑world workflow integration.
• The authors sought an end‑to‑end pipeline that not only classifies MR/TR severity but also provides view selection, morphology assessment, and Doppler interpretation aligned with the systolic phase.

• Development used a retrospective, multicenter approach with an internal dataset of 5,086 outpatient studies from a tertiary center for training and internal testing.
• External validation employed cohorts from two additional institutions to assess generalizability.
• The pipeline integrates automated acquisition of relevant routine echocardiographic views from DICOM files, leaflet pose estimation for valve morphology, and systolic phase awareness to inform Doppler interpretation.

• The model performs automated view selection, leaflets’ positional analysis, and phase‑aware Doppler interpretation.
• It jointly grades MR and TR severity from transthoracic echocardiography, providing an integrated assessment rather than isolated valve analyses.
• The framework emphasizes explainability within an AI workflow, aiming to enhance clinician trust and adoption.

• In the internal test set, the model achieved outstanding discrimination for clinically significant MR/TR (moderate or greater) with area under the ROC curves exceeding 0.98 for both valves.
• Negative predictive values were high, exceeding 0.97, indicating strong rule‑out capability.
• Performance was generally stable across age, sex, and common valvular comorbidities, with modest reductions observed in patients with atrial fibrillation or reduced left ventricular ejection fraction.
• In external cohorts, AUROCs were modestly lower, particularly for severe regurgitation, but sensitivity and negative predictive value remained high, suggesting useful external applicability and conservative screening properties.

• The system supports automated triage and prioritization of echocardiographic studies in both specialist and noncardiology settings, leveraging its robust rule‑out ability and external performance.