Abstract Body: INTRODUCTION: Left ventricular ejection fraction (LVEF) is central to heart failure assessment, yet most echocardiographic AI models require complete video clips and provide limited online reliability assessment, constraining real-time point-of-care use.
HYPOTHESIS: We tested whether a causal state-space model combined with teacher-guided distillation and conformal martingale auditing, can preserve quantitative accuracy while identifying unreliable estimates during streaming inference.
AIMS: To develop and evaluate a self-auditing framework for continuous volume tracking, and LVEF estimation from echocardiographic frames, and to determine whether automated reliability monitoring can trigger beneficial adaptation in uncertain cases.
METHODS: Two pretrained multiview echocardiography foundation models generated dense pseudo-segmentation masks and volume targets for unlabeled frames. A causal student model was distilled for 50 epochs and then fine-tuned for 50 epochs on expert end-diastolic and end-systolic annotations. At inference, frames were processed sequentially, with hidden-state propagation across time. A conformal martingale monitored nonconformity online, and flagged frames were routed for teacher-guided adaptation. Evaluation used EchoNet-Dynamic (n=10,030 apical four-chamber videos).
RESULTS: The model achieved an LVEF mean absolute error of 5.0%, root mean square error of 6.64%, and R² of 0.72. Left ventricular segmentation Dice was 0.92, with end-diastolic volume mean absolute error of 11.78 mL and end-systolic volume mean absolute error of 18.94 mL. Inferred volume trajectories were physiologically coherent across the cardiac cycle and reproduced expected systolic emptying and diastolic filling patterns. Reliability auditing remained stable in clearly preserved or reduced EF studies and rose appropriately in borderline studies. Adaptation improved end-systolic volume R² from 0.79 to 0.85 and end-diastolic volume R² from 0.62 to 0.68, demonstrating that the auditing layer supported active recalibration.
CONCLUSIONS: This approach supports point-of-care deployment by pairing frame-level predictions with transparent uncertainty signals and automated escalation when confidence degrades.