A unified deep-learning model called PH-ECHO-AI, trained on 8,416 echocardiographic clips from four public datasets, performs simultaneous four-chamber segmentation, biventricular ejection fraction estimation, deformation analysis, and pulmonary hypertension (PH) prediction from a single apical four-chamber view. Right ventricular ejection fraction (RVEF) was estimated with a correlation of r=0.754 and mean absolute error of 4.98% against 3D-echocardiographic reference standards — markedly outperforming conventional geometric RV fractional area change (r=0.278). PH detection reached an AUC of 0.697 using geometry alone, with strong calibration (Brier score 0.061) across 1,076 MIMIC-IV patients.
Right ventricular function is a well-established predictor of survival in pulmonary hypertension and heart failure, yet AI cardiac imaging has disproportionately focused on the left ventricle. This model's ability to extract clinically meaningful right-heart metrics from a routine single view — without geometric assumptions or institutional training data — represents a meaningful methodological advance. The RVEF performance approaches published single-view ceilings, suggesting the architecture is near the practical limit of 2D estimation. However, critical limitations apply: cross-centre RVEF validation is explicitly absent, and PH prediction was developed and tested within the same MIMIC-IV cohort, raising overfitting concerns. The AUC of 0.697 for PH detection is modest and unlikely sufficient for standalone clinical screening. As a preprint not yet peer-reviewed, these results remain provisional. If external validation confirms generalisability, this tool could meaningfully democratise right-heart assessment in lower-resource echocardiography settings.