📄 medicine

A Counterfactually Evaluated Additive Regional Decomposition of Ejection Fraction

This paper introduces the Regional Cardiac Contribution Score (RCCS), an interpretable deep-learning framework that decomposes left-ventricular ejection fraction into seven anatomically grounded regional contributions, which are validated through counterfactual simulations to accurately explain how specific ventricular regions drive global systolic function while maintaining high predictive accuracy with minimal parameters.

Original authors: Rochak Dhakal, Kritick Bhandari

Published 2026-09-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Rochak Dhakal, Kritick Bhandari

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The heart is a pump, but doctors often judge its strength with a single number. This number, called the ejection fraction, measures how much blood the left ventricle squeezes out with every beat. It is a vital sign for patients with heart failure, guiding decisions about medication, devices, and survival. For decades, this measurement has been the gold standard, yet it has a blind spot. Because it collapses the entire pumping action into one global value, it cannot tell the difference between a heart that is weak everywhere and one that is strong in some parts but failing in others. Two patients might have the exact same ejection fraction, yet one could have a uniform weakness while the other has a specific, localized injury that the single number hides. In recent years, artificial intelligence has learned to predict this number from ultrasound videos with great speed, but these smart systems usually act as black boxes, offering a number without explaining how they arrived at it or which parts of the heart mattered most.

A team of researchers has now built a new way to look at this problem, turning the single number into a detailed map of the heart's work. They developed a method that breaks the left ventricle into seven distinct anatomical regions, much like dividing a room into corners and walls. For each of these seven zones, the system calculates how much that specific area contributes to the total pumping power. The result is a set of seven scores that add up perfectly to the final ejection fraction, showing exactly how much the base, the middle, and the tip of the heart are each doing the work. The researchers did not just create this map; they tested it rigorously to ensure it was not just a pretty picture but a true reflection of how the computer model was thinking. They found that when they simulated a failure in one specific region, the model's predicted total strength dropped by an amount that matched the score assigned to that region. This proved that the system was genuinely using the information from each part of the heart to make its decision, rather than just guessing the total and then inventing a story to explain it.

The work began by taking standard ultrasound videos of the heart, specifically the view that shows the four chambers. The researchers first used a separate tool to trace the outline of the left ventricle in every frame of the video, creating a moving mask of the cavity. They then divided this cavity into seven fixed zones: the bottom, middle, and top sections of the wall near the septum, the bottom, middle, and top sections of the side wall, and a shared cap at the very tip where the two walls meet. For each of these seven zones, the system measured sixteen different mechanical features. It looked at how big the zone was, how much it shrank during a squeeze, how fast it moved, and when it reached its smallest size compared to the rest of the heart. These measurements were fed into a compact computer model designed with a specific rule: the output for each zone had to be a positive number representing volume, and the sum of all seven outputs had to equal the total ejection fraction. This design forced the model to be additive by nature, ensuring that the final score was literally the sum of its parts.

To see if this breakdown was real or just an illusion, the researchers performed a counterfactual test. They took a trained model and, for each test video, they artificially turned off the movement of one specific region while leaving the other six exactly as they were. They then asked the model to predict the ejection fraction again. If the model had truly learned that a specific region was important, removing its movement should cause the predicted total strength to drop significantly. The researchers compared this drop to the score the model had originally assigned to that region. They found a strong match: regions that the model had given high contribution scores were the same regions that, when silenced, caused the biggest drop in the predicted total. This agreement held true across thousands of heartbeats and even when the model was tested on a completely different set of patient data it had never seen before.

The study also revealed a surprising truth about how artificial intelligence works in medicine. The researchers built a second model that was just as good at predicting the final ejection fraction number as their new system, but it did not have the rule forcing it to break the heart into seven parts. When they tested this second model with the same counterfactual experiment, it failed. Silencing a region did not cause a predictable drop in the score, proving that the second model had learned to predict the number without truly understanding the contribution of each part. This showed that high accuracy in a prediction does not automatically mean the system has a faithful explanation of how it works. The new method, which uses only a tiny fraction of the computer memory required by standard video-analysis systems, managed to be both accurate and transparent.

While the system excelled at explaining the model's own logic, the researchers were careful not to overstate its ability to diagnose specific diseases. They tested whether the regional scores could identify heart muscle damage caused by a heart attack, but the results were weak. The system could not reliably distinguish between healthy and damaged tissue on its own, suggesting that while it is excellent at explaining how a model calculates a pump's strength, it is not yet a standalone tool for finding specific injuries. The method relies heavily on the quality of the initial outline of the heart; if the outline jumps or wobbles between frames, the regional scores become less reliable. However, the core achievement remains: the researchers have successfully transformed a single, isolated number into a compact, testable explanation of how different parts of the heart contribute to the whole. This approach offers a new way to see the heart not just as a black box that outputs a score, but as a collection of working parts whose individual efforts can be measured and understood.

Technical Summary: A Counterfactually Evaluated Additive Regional Decomposition of Ejection Fraction

Problem Statement
Left-ventricular ejection fraction (LVEF) is the standard metric for assessing cardiac systolic function, yet it reduces complex ventricular mechanics to a single global value. This aggregation obscures critical clinical distinctions: a normal EF can result from balanced contraction, severe regional impairment compensated by preserved regions, or asynchronous timing differences. While deep learning models have improved the accuracy and reproducibility of automated EF estimation from echocardiographic videos, they typically operate as "black boxes," providing a global prediction without quantifying how specific anatomical regions contribute to that value. Existing interpretability methods, such as saliency maps or attention mechanisms, often fail to express regional contributions in clinically meaningful units (e.g., EF percentage points) or ensure that these contributions sum exactly to the predicted EF. Furthermore, there is a lack of rigorous validation to determine whether a model's regional attributions truly reflect its dependence on those regions, as opposed to being artifacts of the model architecture.

Methodology
The authors introduce the Regional Cardiac Contribution Score (RCCS), a framework designed to decompose model-predicted EF into seven anatomically grounded contributions derived from left-ventricular (LV) cavity motion.

Data and Preprocessing: The model operates on sequences of LV cavity masks (binary segmentation) from the apical-four-chamber (A4C) view, rather than raw video pixels. Using the EchoNet-Dynamic dataset for training and CAMUS for external evaluation, the system processes temporally normalized mask sequences (32 phases per cardiac cycle).
Anatomical Partitioning: A deterministic, parameter-free rule partitions each LV mask into seven regions: basal, mid, and apical septal; basal, mid, and apical lateral; and a shared apical cap. This partition adapts to LV shape changes throughout the cardiac cycle.
Feature Extraction: For each of the seven regions, a 16-dimensional feature vector is extracted from the mask sequence. These features capture regional size, contractile behavior (area change, slope), timing (peak phase, offset), and stroke mechanics (strain, stroke share).
Additive Volumetric Model:
- A shared encoder processes the 16-dimensional feature vector for each region independently to produce a latent representation.
- Two constrained prediction heads estimate non-negative regional end-diastolic volume ( $\widehat{EDV}_i$ ) and end-systolic volume ( $\widehat{ESV}_i$ ) surrogates. The end-systolic volume is constrained to be less than or equal to the end-diastolic volume via a sigmoid gate.
- The global EF is calculated as the sum of regional stroke volumes divided by the total end-diastolic volume.
- The RCCS for region $i$ is defined as $100 \times (\widehat{EDV}_i - \widehat{ESV}_i) / \sum \widehat{EDV}_j$ . By construction, the sum of the seven RCCS values equals the predicted global EF exactly.
Counterfactual Faithfulness Evaluation: To verify that the assigned scores reflect the model's actual dependence on specific regions, the authors employ a feature-space counterfactual test. For each test case, the features of one region are modified to simulate a non-contracting (akinetic) state while keeping other regions unchanged. The model is re-evaluated, and the resulting decrease in predicted EF ( $\Delta \widehat{EF}$ ) is compared to the original RCCS assigned to that region. A faithful decomposition should show a strong correlation between the assigned score and the magnitude of the EF drop upon intervention.

Key Contributions

RCCS Framework: A novel, inherently interpretable model that quantifies the contribution of seven anatomical LV regions to the predicted EF in percentage points, ensuring the contributions sum exactly to the global prediction.
Counterfactual Validation: The introduction of a feature-space "knockout" analysis to experimentally test whether regional scores are faithful to the model's internal logic, distinguishing between exact additivity (guaranteed by design) and regional faithfulness (empirically tested).
Accuracy-Faithfulness Dissociation: The demonstration that high global EF prediction accuracy does not automatically guarantee a faithful regional decomposition. Through feature-group ablation studies, the authors show that removing specific features (e.g., stroke-related features) can preserve global accuracy while significantly degrading the correlation between assigned scores and counterfactual EF changes.

Results

Global Performance: On the EchoNet-Dynamic test set, RCCS achieved a Mean Absolute Error (MAE) of 4.30 with an $R^2$ of 0.781, using only 5,378 parameters in its prediction head. This performance is competitive with much larger video-based models (e.g., R(2+1)D with 31M parameters, MAE 4.06) and superior to an unconstrained MLP with similar inputs (MAE 4.49).
External Transfer: On the external CAMUS dataset, the zero-shot RCCS model achieved an MAE of 6.27, which improved to 4.87 after a simple affine recalibration, maintaining a Pearson correlation of 0.835.
Faithfulness: The RCCS scores showed strong agreement with counterfactual EF decreases on EchoNet-Dynamic (Pearson $r = 0.74$ ) and CAMUS ( $r = 0.82$ ). In contrast, a control model with similar EF accuracy but no additive constraints showed substantially weaker agreement.
Ablation Findings: Removing "Stroke" features (direct measures of systolic excursion) reduced the faithfulness correlation from 0.738 to 0.361 on EchoNet-Dynamic while barely affecting global MAE (4.30 vs. 4.31). This confirms that accurate prediction alone does not ensure a faithful regional explanation.
Robustness: The model is robust to small, systematic segmentation errors (1-pixel dilation/erosion) but sensitive to temporally inconsistent jitter, highlighting the need for temporally stable upstream segmentation.
Exploratory MI Analysis: In an exploratory analysis on HMC-QU, RCCS showed weak association with segment-level myocardial infarction (AUROC 0.561), indicating it is primarily a decomposition of model-predicted EF rather than a direct detector of regional pathology.

Significance and Claims
The paper claims that RCCS transforms EF from an isolated global estimate into a compact, experimentally testable explanation of how anatomically grounded regions contribute to systolic function. The primary significance lies in the decoupling of prediction accuracy from interpretability: the authors demonstrate that a model can be highly accurate at predicting a global value without providing a meaningful breakdown of why that value was reached. By enforcing an additive structure and validating it through counterfactual interventions, RCCS provides a framework where regional contributions are not just visual approximations but are mathematically tied to the prediction and empirically verified to reflect the model's reliance on specific regional mechanics.

The authors maintain a modest stance regarding clinical utility, noting that RCCS is currently a research tool that decomposes model-predicted EF. They explicitly state that the framework does not yet equate to physiological ground truth or a direct diagnostic tool for regional myocardial pathology (such as infarction), and that the relationship between RCCS and specific disease phenotypes (e.g., Takotsubo cardiomyopathy) remains a hypothesis requiring further validation against clinical standards. The work establishes a methodological precedent for building and validating interpretable cardiac models where the explanation is structurally guaranteed and experimentally verified.

Technical Summary: A Counterfactually Evaluated Additive Regional Decomposition of Ejection Fraction

More like this