NeuroECG: ECGFounder-Based Deep ECG Representation for EEG-Free Neurological Prognostication After Cardiac Arrest
Jiaju Gao, Yi Zhao, Chenyang Xu, Yuxi Zhou, Hao Wang
cs.LG · cs.AI
Quick verdict
Fine-tunes ECGFounder on bedside ECG for post-arrest prognostication (AUROC 0.8077), but the gain over static covariates is statistically inconclusive.
Summary
This paper proposes NeuroECG, a framework that adapts the pretrained ECGFounder foundation model to predict neurological outcome after cardiac arrest from bedside monitoring ECG and basic clinical variables, avoiding resource-intensive EEG. From 412 patients with usable ECG in the multicenter I-CARE database, recordings from the first 72 hours after ROSC are segmented into 10-second single-channel windows, quality-controlled with a flatline mask, and encoded by an ECGFounder backbone fine-tuned with gradual unfreezing. Segment embeddings are aggregated per patient via feature-wise quantile pooling (q=0.24), compressed to 64 dimensions with PCA, concatenated with six static covariates, and classified by CatBoost for poor outcome (CPC 3-5). The adapted backbone achieves the best ECG-only AUROC (0.7333) among five tested backbones; the fused model reaches test AUROC 0.8077 and AUPRC 0.8970 versus 0.7689 for static covariates alone. Ablations favor quantile pooling over mean/max pooling and gradual over immediate full fine-tuning. Main limitations: the AUROC gain over static covariates (0.0465, 95% CI -0.0293 to 0.1263) is statistically insignificant, all ablations are reported on one small test set (n=63), there is no external validation, and 195 of 607 patients were excluded for missing ECG without selection-bias analysis.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- An ECGFounder backbone adapted via gradual unfreezing achieves the best ECG-only performance among five tested backbones, with test AUROC 0.7333 on 412 I-CARE patients.
- Fusing PCA-compressed deep ECG features (64-d) with six static clinical covariates yields test AUROC 0.8077 and AUPRC 0.8970 for poor neurological outcome (CPC 3-5), versus 0.7689 AUROC for covariates alone.
- Feature-wise quantile pooling (q=0.24) is the best patient-level aggregation of segment embeddings, outperforming mean and max pooling in the Static + Deep PCA configuration.
- Gradual unfreezing outperforms both head-only fine-tuning (0.6968) and immediate full-backbone unfreezing (0.6213) for adapting ECGFounder to bedside ECG.
- Deep ECG representations contribute complementary prognostic information, accounting for 58.04% of total SHAP importance in the fused model despite being 64 of 70 features.
Strengths
Leakage-aware evaluation design
Patient-level splits stratified jointly by hospital site (A-F) and outcome; standardization, PCA, and CatBoost fitted on training data only (Eq. 11, Sec. IV-A); the pooling quantile q was chosen by five-fold CV confined to the training cohort (Table V), not on the test set.
Standardized QC pipeline for messy continuous ICU ECG
Deterministic channel-selection rule across five I-CARE labels, 0.5-40 Hz bandpass, robust percentile normalization (Eq. 3), flatline rejection at segment std >= 0.05 (Eq. 4), 20 s stride, and per-patient capped sampling of 256 segments/epoch (Eq. 9) to prevent long recordings from dominating - a reusable benchmark for future waveform studies.
Controlled backbone comparison under one identical pipeline
Transparent reporting of an unfavorable statistic
Thorough ablation coverage
Weaknesses
Central incremental-value claim is not statistically supported
All ablations and design comparisons reported on one small test set
Single database and 32% cohort attrition without bias analysis
Pipeline benefits appear configuration-specific and fragile
No EEG-based comparison on the same cohort
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- 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
Can bedside ECG predict neurological outcome after cardiac arrest without EEG?
Partially, as an auxiliary signal. In this study, deep learning on bedside ECG alone reached an AUROC of 0.7333 for predicting poor outcome (CPC 3-5) in 412 I-CARE patients, and combining ECG features with six routine clinical variables (age, sex, time to ROSC, out-of-hospital arrest, shockable rhythm, TTM temperature) reached 0.8077. However, the improvement over clinical variables alone (0.7689) was not statistically significant in the authors' bootstrap analysis, so ECG is promising but not yet a proven substitute for EEG-based prognostication.
What is the NeuroECG model and how does it work?
NeuroECG fine-tunes the pretrained ECG foundation model ECGFounder on 10-second single-channel bedside ECG segments from the first 72 hours after return of spontaneous circulation, using a gradual unfreezing schedule over 25 epochs. Segment-level embeddings are aggregated per patient with feature-wise quantile pooling (q=0.24), compressed from 1024 to 64 dimensions with PCA, concatenated with six static clinical covariates, and fed to a CatBoost classifier that predicts poor neurological outcome.
What data was the study based on and is the code available?
The study used the public multicenter I-CARE cardiac arrest database; 412 of 607 patients had usable bedside ECG within 72 hours after ROSC, split 288/61/63 into training/validation/test. The source code is stated to be available at github.com/goddream66/NeuroECG, while the I-CARE database itself requires PhysioNet access approval.
How accurate is NeuroECG?
On a 63-patient held-out test set averaged over five seeds, the fused ECG-plus-covariates model achieved AUROC 0.8077, AUPRC 0.8970, F1-score 0.8218, recall 0.8524, and CPC regression MAE 1.5156. Static covariates alone already achieved AUROC 0.7689, confidence intervals are wide, and the ECG-related gain was statistically inconclusive.
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 1-2 are available only as captions in the extracted text; the SHAP figure and workflow diagram could not be independently inspected.
- Tables were reconstructed from linearized PDF text; although values are internally consistent across sections (e.g., 0.7333 and 0.8077 recur correctly), individual cell alignment could contain extraction artifacts.
- Review is based on arXiv v1 (13 Jul 2026); no peer-review history or author response exists yet, so flagged issues are questions for revision rather than settled criticisms.
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.