Ejection fraction on a budget: mapping the accuracy-compute trade space for video-based ejection fraction estimation

Abstract

Deep video networks estimate left ventricular ejection fraction (EF) from echocardiograms with expert-level accuracy, but the compute cost of running them is rarely reported. This leaves anyone building a handheld or bedside tool without clear guidance on what to deploy. We measured the accuracy-versus-compute tradeoff for EF estimation on EchoNet-Dynamic by training 22 configurations that varied clip length (8 to 64 frames), frame sampling period (1 to 4), and backbone: R(2+1)D-18, R3D-18, MC3-18, X3D-S, X3D-M, and a 2D ResNet-18 with temporal pooling. All models used one fixed training recipe. Every configuration was evaluated for accuracy using mean absolute error, R-squared, and Bland-Altman agreement; clinical utility using sensitivity and specificity at the clinically relevant EF cutoffs of 40% and 50%, plus error stratified by EF band; and cost using floating-point operations, parameter count, GPU and CPU latency, and peak memory under a single frozen measurement protocol. We stress-tested the main findings with replicate training seeds. Sparse temporal sampling outperformed dense sampling at matched frame budgets. A sampling period of 4 outperformed a period of 1 at every tested frame count while also reducing per-video cost. In the seed-replicated 8-frame comparison, the advantage averaged one full point lower mean absolute error across all nine cross-seed pairings. A standard R3D-18 achieved the best accuracy in the study, with a mean absolute error of 3.99, while requiring 19% less CPU latency than the reference configuration. A 16-frame, period-4 R(2+1)D-18 cut reference cost in half with no statistically confirmed loss in accuracy. Removing temporal modeling entirely substantially reduced accuracy, with a mean absolute error of 5.65, setting a practical floor for how inexpensive this task can be. We release the code, cost-measurement protocol, and per-configuration results.