Learning from Scarce Labels: Multi-View Echocardiography for Ejection Fraction Prediction
Zhiyuan Gao, Dominic Yurk, Yaser S. Abu-Mostafa
eess.IV · cs.CV · cs.LG
Quick verdict
Claims first public PLAX-EF benchmark (MAE 6.86%) and fusion gain to 6.37%; solid artifacts, but cross-cohort comparisons and label noise temper the claims.
Summary
The paper addresses the absence of public PLAX–EF supervision by building the first openly released resource linking parasternal long-axis echocardiography videos to ejection fraction labels. From MIMIC-IV-Echo (~500k unlabeled videos) and MIMIC-IV-Note, the authors train an X3D view classifier (bootstrapped from TMED-2 via a ResNet-34 image model, using EchoNet-Dynamic/LVH for A4C/PLAX), mine ~25,500 PLAX videos across 4,822 studies, and label them with study-level EF from an A4C teacher model (proxy supervision). Evaluation uses an independent cohort of 295 studies whose EF labels are GPT-4-extracted from discharge notes within a 1-day window; three LLM extractors agreed on 99.5% of notes, and the A4C teacher scores 6.95% MAE on these labels. The PLAX R(2+1)D ensemble reaches 6.86% MAE (r=0.670), and unweighted A4C+PLAX late fusion yields 6.37% MAE (r=0.709) on 284 paired studies. Extended statistics show PLAX is statistically comparable, not superior, to A4C, and significant proportional bias affects all views. Main limitations: proxy-label circularity, no human expert EF annotation of test videos, view-classifier fine-tuning on studies overlapping the test cohort, cross-dataset comparisons backing the 'surpasses existing methods' claim, and no significance test for the headline fusion gain.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- This is the first publicly released resource pairing PLAX cine videos with EF labels (25,532 proxy-labeled videos across 4,822 MIMIC-IV-Echo studies).
- A PLAX ensemble trained with A4C proxy supervision achieves 6.86% study-level MAE on an independent note-derived cohort of 295 studies.
- Unweighted A4C+PLAX late fusion improves both single-view baselines to 6.37% MAE (r=0.709) on 284 paired studies.
- Note-derived EF labels are reliable, evidenced by 99.5% exact agreement across three LLM extractors and an A4C-teacher MAE of 6.95% close to its published out-of-sample 6.0%.
- PLAX-based EF prediction matches but does not statistically exceed the A4C teacher (Wilcoxon p=0.66).
Strengths
First open PLAX–EF resource with unusually complete artifacts
Label CSVs for all cohorts, GPT-4 extraction prompts, pretrained A4C/PLAX weights, an interactive HF Space, and a Colab notebook (Sec. 6, App. B.3) address a genuine supervision gap documented in Table 1.
Non-circular evaluation design
The 295-study ground-truth cohort (Sec. 4.3) is note-derived and independent of the proxy-labeled training set; the A4C teacher's 6.95% MAE on these labels approximates its published 6.0% out-of-sample error, supporting label plausibility.
Multi-LLM validation of extracted labels
Candid statistical treatment in the extended version
Error-driven classifier improvement loop
Weaknesses
Cross-cohort superiority claim
No human expert EF annotation of the test cohort
View-classifier fine-tuning overlaps the test cohort
Headline fusion gain lacks a significance test
Unvalidated final view classifier and wide agreement limits
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This MELBA extended version of a MIDL 2025 oral (Gao, Yurk, Abu-Mostafa) targets a real supervision gap: no public dataset pairs parasternal long-axis (PLAX) cine loops with left-ventricular ejection fraction (EF). The authors claim three firsts: (i) the first public PLAX–EF label resource — 25,532 proxy-labeled videos across 4,822 MIMIC-IV-Echo studies (§4.5); (ii) the first reproducible PLAX EF model — 6.86% study-level MAE on a 295-study note-derived cohort (§5.1); and (iii) unweighted A4C+PLAX late fusion — 6.37% MAE, r=0.709, on 284 paired studies (§5.2, Table 3). A fourth claim, that this 'surpasses the performance of existing methods' (abstract, §5.1), is the weakest link: the comparisons to ExoAI (7.29% MAE, Vega et al. 2024) and LVID/landmark pipelines (8.45%, Goco et al. 2022) are cross-dataset and cross-protocol, so ranking is not established.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the first public dataset for predicting ejection fraction from PLAX echocardiograms?
This paper releases what the authors state is the first public PLAX–EF label resource: 25,532 PLAX videos across 4,822 MIMIC-IV-Echo studies with study-level EF labels generated by an A4C teacher model, plus a 295-study test cohort with EF extracted from clinical notes. Label files, prompts, and pretrained models are on GitHub and Hugging Face; the videos themselves require credentialed PhysioNet access.
How accurate is EF prediction from the parasternal long-axis (PLAX) view?
On this paper's note-derived benchmark of 295 studies, the PLAX two-model ensemble achieves 6.86% MAE (Pearson r=0.670), versus 6–7% MAE typically reported for the clinical-standard A4C view. Combining A4C and PLAX predictions by simple averaging reaches 6.37% MAE. The paper's own statistics show PLAX is statistically comparable to A4C, not better.
Is PLAX as good as the apical four-chamber view for estimating ejection fraction?
In this study, yes within measurement noise: the PLAX ensemble scores 6.77% vs 7.01% MAE for A4C on 284 paired studies, a difference that is not statistically significant (Wilcoxon p=0.66). This suggests PLAX can substitute when apical windows are poor, though both views show significant proportional bias (regression toward mean EF) and limits of agreement around ±16%.
Can I run the PLAX EF model myself?
Yes. The authors release Hugging Face model weights (A4C and PLAX), an interactive Hugging Face Space demo, and a Google Colab notebook for single-clip inference. Reproducing the paper's study-level metrics additionally requires credentialed MIMIC-IV-Echo access plus the released label CSV files and instructions.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–12 are not rendered in the extracted text; plot-level observations rely on the in-text descriptions and quoted statistics only.
- The released GitHub/Hugging Face/Colab artifacts were not accessed; reproducibility and artifact claims are assessed from the paper's documentation alone.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.