📄 cardiovascular medicine

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

This paper systematically maps the accuracy-compute trade-off for video-based ejection fraction estimation by evaluating 22 model configurations, revealing that sparse temporal sampling with specific backbones like R3D-18 or R(2+1)D-18 can significantly reduce computational costs while maintaining expert-level accuracy, thereby providing clear deployment guidance for resource-constrained clinical tools.

Original authors: Pandey, A., Sharma, K., Shah, A.

Published 2026-09-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Pandey, A., Sharma, K., Shah, A.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

In the quiet hum of a hospital room, a doctor holds a small probe against a patient's chest, sending sound waves deep into the body to watch the heart beat. The most important number this test produces is the ejection fraction, a percentage that tells how much blood the main pumping chamber squeezes out with every beat. This single figure acts as a vital sign for heart failure, determining whether a patient needs specific medications or if their condition is severe enough to require advanced intervention. Doctors have traditionally measured this by manually tracing the edges of the heart on a video screen, a process that varies from one expert to another and can be slow. In recent years, artificial intelligence has learned to do this automatically, reading the video frames to calculate the number with a level of skill that rivals human specialists. However, these smart programs are often heavy and hungry for computing power, designed to run on powerful hospital servers rather than the small, battery-powered devices doctors carry in their pockets or use at a patient's bedside.

A team of researchers set out to map the landscape of these artificial intelligence models to find the sweet spot between accuracy and cost. They asked a practical question: how much computing power is actually required to get a reliable heart measurement, and where does the quality of the answer begin to crumble? To find out, they trained twenty-two different versions of a video-analysis system on a large collection of heart ultrasound recordings. They systematically changed three things for each version: how many frames of video the system looked at at once, how far apart in time those frames were spaced, and the specific mathematical architecture the system used to process the images. They measured not just how close the computer's guess was to the true value, but also how long it took to run on a standard processor and how much memory it needed.

The study revealed a clear and somewhat surprising rule about how these systems should look at time. The researchers found that spreading the video frames out over a longer period actually produced better results than cramming many frames into a short burst. Imagine watching a movie by looking at one frame every second versus looking at four frames every second; the first method captures the full motion of a heartbeat more effectively than the second, even if the total number of pictures viewed is the same. When the researchers forced the systems to look at frames that were very close together in time, the accuracy dropped significantly, and the cost of running the system went up because the computer had to process many more tiny clips to cover the same amount of heartbeats. The most efficient approach was to sample the video more sparsely, letting the system see the full cycle of a heartbeat without needing to process every single moment.

When it came to the architecture of the systems themselves, the researchers discovered that simpler designs could outperform the complex ones that had been considered the standard. One specific type of network, which treats the video as a solid block of three-dimensional data, achieved the highest accuracy in the entire study. It matched the performance of the best-known model but ran nineteen percent faster on a standard processor. Another finding was that removing the ability to see motion entirely caused the system to fail. A model that looked at each video frame independently, without connecting them in time, produced errors so large that it would be useless for clinical decisions. This established a hard floor: any device that wants to measure heart function accurately must be able to understand motion, and it cannot be made cheap enough to run on the most basic hardware without losing its medical value.

The researchers identified three distinct paths for deployment depending on the available resources. For devices with a generous budget, the fastest and most accurate option was the solid 3D network. For those with a tighter budget, a slightly different version of the standard model could cut the computing cost in half with no statistically significant loss in accuracy. For the most constrained devices, a lightweight mobile network could run at a third of the cost, though it required a specific adjustment to the video input that the researchers noted needed further testing. The study concluded that while the cost of accurate heart measurement can be reduced dramatically, it cannot be eliminated. The savings end where the ability to see motion ends, and any attempt to go cheaper than that results in a tool that misses the very patients who need help the most. By releasing their data and methods, the team provided a clear map for engineers to build the next generation of handheld heart monitors, ensuring they are both smart enough to be trusted and small enough to be carried.

Technical Summary: Ejection Fraction on a Budget

Problem Statement
Deep learning models have demonstrated expert-level accuracy in estimating left ventricular ejection fraction (EF) from echocardiogram videos, yet the computational cost of deploying these models in resource-constrained environments (e.g., handheld probes, point-of-care devices) remains largely unquantified. While previous work established that EF can be regressed directly from video, the trade-off between accuracy and compute cost has not been systematically mapped under a single protocol. Clinicians and engineers lack guidance on which configurations (clip length, sampling period, backbone architecture) offer the best efficiency without sacrificing clinical utility, particularly at critical treatment thresholds (EF <40% and <50%).

Methodology
The authors mapped the accuracy-compute trade space using the EchoNet-Dynamic dataset (10,030 apical four-chamber videos). They trained 22 distinct configurations under a fixed training recipe, varying three primary axes:

Clip Length ( $F$ ): 8, 16, 32, and 64 frames.
Sampling Period ( $P$ ): 1, 2, and 4 (frames between sampled inputs).
Backbone Architecture: R(2+1)D-18 (reference), R3D-18, MC3-18, X3D-S, X3D-M, and a 2D ResNet-18 with temporal pooling (serving as a temporal-modeling floor).

Evaluation Protocol:

Accuracy: Measured via Mean Absolute Error (MAE), RMSE, $R^2$ , and Bland-Altman agreement. Clinical utility was assessed via sensitivity and specificity at EF thresholds of 40% and 50%, with error stratified by EF bands.
Cost: Measured under a single frozen protocol on a specific hardware setup (NVIDIA RTX 2070 Super GPU; 4-thread CPU). Metrics included floating-point operations (FLOPs), parameter count, model size, GPU/CPU latency, and peak memory. Crucially, costs were reported per video, accounting for the number of non-overlapping clips required to tile a full video based on the $F \times P$ span.
Robustness: Key comparisons were stress-tested using three different training seeds to distinguish between random variance and consistent architectural advantages.

Key Results

Sparse Sampling Superiority: At a fixed frame count, sparse temporal sampling (higher $P$ ) consistently outperformed dense sampling. For example, at 8 frames, increasing the period from 1 to 4 improved MAE by approximately 1.0 point (5.34 to 4.36) while reducing per-video cost by a factor of four. Dense sampling forces the network to see only a fraction of a cardiac cycle, whereas sparse sampling covers a full systole-diastole excursion within the same frame budget.
Backbone Performance:
- R3D-18: Achieved the best point accuracy in the study (MAE 3.99) using the reference geometry (32 frames, period 2). It was statistically tied with the reference R(2+1)D-18 (MAE 4.04) but achieved a 19% reduction in CPU latency. The factorized design of R(2+1)D-18, which typically benefits action recognition, offered no accuracy advantage here and incurred higher overhead.
- X3D Models: X3D-M (16 frames, period 4) achieved an MAE of 4.14 (statistically indistinguishable from the reference) at roughly one-third of the reference's cost and 12 MB model size. However, these results carry a caveat: X3D requires input upsampling from the dataset's native 112 pixels to its native resolution, introducing a confound regarding input quality.
- 2D Floor: Removing temporal modeling entirely (2D ResNet-18 with pooling) collapsed accuracy (MAE 5.65), demonstrating that motion modeling is essential for this task.
The Pareto Frontier: The accuracy-cost frontier is steep below ~500ms per video and flat above it.
- High Budget: R3D-18 (32×2) offers the best accuracy with lower latency than the reference.
- Medium Budget: R(2+1)D-18 (16×4) halves the reference cost with no statistically confirmed accuracy loss (though seed replication suggests a small, consistent disadvantage).
- Low Budget: X3D-M (16×4) offers the lowest cost among non-inferior models, provided the upsampling confound is addressed.
Clinical Error Structure: Errors are not uniform; they concentrate in the low-EF bands (<40%) where treatment decisions are made. Cheap models that degrade accuracy disproportionately miss patients in these critical bands.

Significance and Claims
The paper claims to provide the first systematic mapping of the accuracy-compute trade space for video-based EF estimation under a unified protocol. Its primary contributions are:

Quantified Trade-offs: It identifies specific configurations that reduce compute costs by up to 50% (or more) without statistically significant accuracy loss, providing concrete deployment options for edge devices.
Temporal Sampling Insight: It establishes that sparse sampling is strictly superior to dense sampling for fixed frame budgets, a finding that holds across all backbones and seed replicates.
Compute Floor: It defines a "compute floor" below which accuracy collapses (the 2D pooling models), indicating that temporal modeling cannot be removed entirely for clinical viability.
Methodological Rigor: By stress-testing claims with multiple seeds, the authors demonstrate that single-seed comparisons can be misleading, particularly when evaluating "cheap" configurations against references.

The authors conclude that while significant savings are possible, they are bounded by the necessity of temporal modeling. They release all code, cost protocols, and per-configuration results to allow future efficient architectures to be evaluated on the same map.

More like this