Abstract Body: Introduction: Pressure-volume (PV) loops are central to characterizing ventricular function but typically require invasive catheterization. Apical four-chamber (4CH) echocardiography is noninvasive and widely available, yet routine metrics such as ejection fraction (EF) omit pressure information and limit mechanistic insight.

Hypothesis: Building on the existing physics-informed self-supervised learning framework (Med-Real2Sim) for medical digital twins, we hypothesize that our targeted explainable improvements in synthetic data generation with physiologic filtering, pretext/interpolator modeling, and the PSSL framework optimization would reduce EF error while yielding physiologically consistent and clinically interpretable parameter attributions.

Methods: We conducted a retrospective model development using the public CAMUS dataset with reference end-diastolic volume (EDV), end-systolic volume (ESV), and EF. First, a “pretext” stage generated 1,450 synthetic cardiovascular parameter sets that were retained; a differentiable interpolator was then trained to map these parameters to EDV/ESV. Next, a deep video network ingested each 4-CH cine and predicted seven patient-specific parameters—cycle duration, starting ventricular volume, maximum and minimum elastance, mitral and aortic valve resistances, and the zero-pressure volume. The primary endpoint was mean absolute error (MAE) in EF compared with the echocardiographic reference. Model interpretability was assessed with Shapley additive explanations (SHAP) at both the interpolator and full pipeline levels to quantify the contribution of each parameter to EDV, ESV, and EF.

Results: In the synthetic pretext set, the interpolator achieved an EF MAE of 1.34% (as compared to 2.30% reported in the previous work). In end-to-end testing on CAMUS, we achieved a test EF-MAE of 6.37%, which was modestly better than the supervised 3D-CNN baseline (6.58%). Explainability analyses were clinically coherent: maximum elastance and cycle duration were the principal drivers of EF, zero-pressure volume ranked next, and valve resistances contributed modestly. These patterns align with cardiovascular physiology.

Conclusions: our proposed explainable digital-twin pipeline built on 4-CH echocardiography reconstructed PV loops and estimated EF fraction with competitive accuracy. These findings support the feasibility of bedside hemodynamic phenotyping from echocardiography in settings where invasive data are unavailable.