Frozen Echocardiography Foundation Model Embeddings Contain Linearly Recoverable LVOT VTI and TAPSE Report-Label Signal in MIMIC-IV-ECHO

Quantitative transthoracic echocardiographic measurements support cardiovascular assessment, but structured report-label values may be unavailable for downstream research. We evaluated whether frozen EchoPrime study-level embeddings encoded signal associated with structured left ventricular outflow tract velocity-time integral (LVOT VTI) and tricuspid annular plane systolic excursion (TAPSE) report-label values. MIMIC-IV-ECHO studies with report labels and frozen EchoPrime embeddings were analyzed. Clip-level 512-dimensional embeddings were mean-pooled into study-level all-clips embeddings. Supervised Ridge regression used deterministic subject-level train/validation/test splits, train-fit standardization, and validation-only alpha selection. LVOT VTI was primary, TAPSE secondary. Comparators were a train-median null model and leakage-safe non-image baselines using age, sex, and acquisition metadata. For LVOT VTI, 3782 studies were split into 2654 training, 573 validation, and 555 test studies. Ridge achieved test mean absolute error (MAE) 3.64 cm versus 4.54 cm for null, root mean squared error (RMSE) 4.66 cm, R2 0.372, and MAE 95% confidence interval (CI) 3.41-3.88 cm. Observed-vs-predicted Pearson r was 0.613 (95% CI, 0.556-0.664). For TAPSE, 1131 studies were split into 787 training, 184 validation, and 160 test studies. Ridge achieved MAE 3.17 mm versus 3.79 mm for null, RMSE 3.94 mm, R2 0.284, and MAE 95% CI 2.80-3.56 mm. Pearson r was 0.539 (95% CI, 0.424-0.632). Non-image baselines were weaker than imaging embeddings. Frozen EchoPrime study-level embeddings encoded signal associated with structured LVOT VTI and TAPSE report-label values. Findings support representation-level report-label prediction experiments, not EchoPrime fine-tuning, direct measurement extraction, measurement replacement, clinical deployment, or decision support.