PaperlayerView exact revision on arXiv

Source-linked AI summary

NeuroECG: ECGFounder-Based Deep ECG Representation for EEG-Free Neurological Prognostication After Cardiac Arrest

Jiaju Gao, Yi Zhao, Chenyang Xu, Yuxi Zhou, Hao Wang

arXiv:2609.18891v1cs.LGcs.AI

TL;DR

EEG is a resource-intensive cornerstone of post-arrest neurological prognostication, while bedside ECG is routinely available but underexplored for this purpose. NeuroECG adapts ECGFounder to bedside ECG and fuses its patient-level representations with static clinical covariates. The fused model achieves the best overall classification performance in the reported experiments.

  • Problem

    EEG is a cornerstone of post-arrest neurological prognostication but requires substantial clinical resources, while bedside ECG's prognostic value remains underexplored.

    Source: §I · p. 1

  • Method

    NeuroECG gradually fine-tunes pretrained ECGFounder on bedside ECG segments, pools and compresses their embeddings, then combines the patient-level representation with static clinical covariates.

    Source: §I · p. 2

  • Results

    NeuroECG achieved the best overall classification performance, with mean test AUROC 0.8077 and AUPRC 0.8970, while ECGFounder led ECG-only backbones with test AUROC 0.7333.

    Source: B. Experimental Result · pp. 5–6

  • Takeaways & Limitations

    The results suggest bedside ECG provides useful auxiliary prognostic information, and combining its deep representation with static clinical covariates improves prediction over clinical covariates alone.

    Source: §VI · p. 7

    Experiments used only the I-CARE database, with no external validation; independent institutional cohort validation remains future work.

    Source: §V · p. 7

Abstract

from arXiv · show

Neurological prognostication after cardiac arrest commonly relies on electroencephalography (EEG). However, EEG demands high clinical resources. Bedside electrocardiography (ECG) is standard and low-cost. Yet, its value for predicting neurological outcomes remains underexplored. In this study, we propose NeuroECG, an ECGFounder-based deep representation framework for EEG-free auxiliary prognostication. NeuroECG adapts a pretrained ECG foundation model via task-specific fine-tuning. We implement a gradual unfreezing strategy on single-channel bedside monitoring ECG. Multiple ECG segments per patient are encoded into segment-level deep features. These embeddings are aggregated via quantile pooling ($q = 0.24$) and compressed using principal component analysis (PCA). Experiments on 412 ECG-available patients from the multicenter I-CARE database show that the adapted ECGFounder backbone achieves the best performance among ECG-only backbone baselines, with a test AUROC of 0.7333. We further combine the learned deep ECG representation with static clinical covariates. The proposed NeuroECG model achieves a test AUROC of 0.8077 and an AUPRC of 0.8970. These results support deep bedside ECG representations as a useful source of auxiliary prognostic information. Their integration with static clinical covariates improves prediction in an EEG-free setting. The source code is available at https://github.com/goddream66/NeuroECG_

I. INTRODUCTION

EEG is a cornerstone of neurological prognostication after cardiac arrest, but its resource demands motivate bedside ECG as an accessible auxiliary signal. NeuroECG adapts a pretrained ECG foundation model and combines patient-level ECG representations with static clinical covariates.

I. INTRODUCTION

  • EEG is a cornerstone of post-arrest prognostication, but it requires specialized hardware, trained technicians, and expert interpretation, constraining availability in some settings.Bedside ECG is routinely acquired in ICU patients at low cost and may provide an accessible auxiliary signal, although it does not directly measure cortical activity.Source: §I · p. 1
  • Conventional ECG descriptors can be noise-sensitive and weak as standalone predictors, while small, heterogeneous, imbalanced cohorts make training complex deep models from scratch difficult.These constraints motivate representation learning from raw bedside ECG and the use of pretrained foundation models.Source: §I · p. 1
  • NeuroECG adapts pretrained ECGFounder to bedside monitoring ECG through gradual unfreezing, then combines patient-level deep ECG representations with static clinical covariates.The framework uses 10-second segments, feature-wise quantile pooling at q = 0.24, PCA, and CatBoost for prediction.Source: §I · p. 2
  • The study establishes a standardized preprocessing and flatline-oriented quality-control workflow for continuous bedside ICU ECG in I-CARE.The authors describe it as a practical benchmark for future waveform studies.Source: §I · p. 2

II. RELATED WORK

Prior work has explored ECG-based neurological prediction and pretrained ECG representations, alongside transfer learning and multimodal ICU prediction. NeuroECG brings these strands to continuous bedside ECG with static clinical covariates.

A. ECG-Based Prognostication After Cardiac Arrest

  • Takahashi et al. modeled neurological outcomes from lead-II ICU ECG and reported a best AUROC of 0.688 for favorable outcome, while continuous ICU ECG remains vulnerable to noise and missing signals.The cited work used 71 diagnostic statement probabilities from a PTB-XL-based premodel; direct deep representation learning from raw bedside ECG remains underexplored.Source: A. ECG-Based Prognostication After… · p. 2

B. Pretrained ECG Foundation Models

  • Pretrained ECG models use approaches including masked and contrastive pretraining and joint-embedding predictive learning [14], motivating their use for bedside ECG outcome prediction.Other examples include spatio-temporal masked reconstruction in ST-MEM [15] and models connecting ECG with clinical knowledge or language [16]–[18].Source: B. Pretrained ECG Foundation Models · p. 2

C. Transfer Learning and Multimodal ICU Prediction

  • Transfer learning and multimodal ICU models address small, heterogeneous cohorts, but post-arrest multimodal studies often rely on EEG or multiple biosignal modalities.NeuroECG instead adapts a pretrained ECG model and aggregates bedside ECG segment embeddings with static clinical covariates.Source: C. Transfer Learning and Multimodal… · p. 2

A. Overview

NeuroECG converts continuous bedside ECG into a compact patient-level representation for prediction. Its preprocessing handles heterogeneous channels and excludes flatline-like segments before patient-level modeling.

A. Overview

  • Recordings from the first 72 hours after ROSC are divided into 10-second single-channel segments, encoded by adapted ECGFounder, pooled feature-wise at q = 0.24, compressed by PCA, and fused with clinical covariates for CatBoost prediction.The overview describes a compact patient-level ECG representation integrated with static covariates.Source: A. Overview · pp. 2–3, Figure 1 · p. 3

B. ECG Preprocessing and Segment Construction

  • A deterministic channel-selection rule handles five channel labels; selected signals are filtered, resampled to 500 Hz, normalized, and divided into 10-second segments with a 20-second stride.A flatline-oriented mask retains valid segments for ECGFounder encoding; when two channels are valid, both are retained as separate samples.Source: B. ECG Preprocessing and Segment… · p. 3
  • Patient-level splits prevent data leakage and use joint stratification by hospital site and outcome; poor outcome is defined as CPC 3–5 and treated as the positive class.Source: B. ECG Preprocessing and Segment… · p. 3

C. ECGFounder-Based Deep ECG Representation

The ECGFounder pathway turns single-channel ECG segments into patient-level deep representations through staged fine-tuning, feature aggregation, and compression. Its design combines hierarchical ECG encoding with safeguards for adapting to the clinical cohort.

C. ECGFounder-Based Deep ECG Representation

  • ECGFounder processes each 10-second single-channel segment through a convolutional stem and hierarchical stages, then global average pooling yields a 1024-dimensional embedding.The stages transform temporal feature maps, with residual bottleneck blocks and channel recalibration in the backbone.Source: C. ECGFounder-Based Deep ECG… · pp. 3–4
  • Gradual unfreezing adapts ECGFounder over 25 epochs, progressing from head-only training to full-backbone fine-tuning with a smaller learning rate.The schedule also uses phase-specific learning rates to preserve pretrained features during adaptation.Source: C. ECGFounder-Based Deep ECG… · p. 4
  • Patients with more than 256 valid segments contribute 256 sampled without replacement per epoch; patients with 256 or fewer contribute all segments.Sampling is repeated each epoch during segment-level fine-tuning to limit the influence of patients with longer recordings.Source: C. ECGFounder-Based Deep ECG… · p. 4
  • Feature-wise quantile pooling aggregates the segment embeddings into a patient-level representation, which PCA compresses to 64 dimensions for fusion.The PCA dimension was selected in preliminary validation experiments to reduce the dimensional gap with the six static covariates without overly aggressive compression.Source: C. ECGFounder-Based Deep ECG… · p. 4

D. Patient-Level Fusion and Prediction

The patient-level prediction pathway concatenates a compact ECG representation with static clinical covariates for outcome prediction. A separate regression model estimates CPC score.

D. Patient-Level Fusion and Prediction

  • The fused patient-level feature vector concatenates a 64-dimensional deep ECG representation with six clinical covariates.The covariates are age, sex, time to ROSC, out-of-hospital cardiac arrest, shockable rhythm, and TTM temperature.Source: D. Patient-Level Fusion and Prediction · p. 4
  • A CatBoost classifier predicts the probability of poor neurological outcome from the fused patient-level representation.The table caption identifies comparisons of downstream classification metrics and CPC-MAE under the proposed Deep PCA + Static configuration.Source: D. Patient-Level Fusion and Prediction · p. 4, Table II · p. 5
  • A separate CatBoost regressor estimates CPC score as an ordered numerical target, with performance evaluated using MAE.Table I covers ECG-backbone performance under ECG-only and PCA-compressed settings.Source: D. Patient-Level Fusion and Prediction · p. 5, Table I · p. 5

A. Experimental Setup

The experiments use ECG-available patients from the multicenter I-CARE database and evaluate patient-level prediction on a fixed, stratified split. Cross-validation and shared preprocessing support the model comparisons.

A. Experimental Setup

  • Of 607 I-CARE patients, 412 with usable ECG were included and split into 288 training, 61 validation, and 63 test patients, stratified by site and outcome.The split used seed 42.Source: A. Experimental Setup · p. 5
  • The quantile threshold was selected by five-fold training-cohort cross-validation, with adaptation, standardization, PCA, and CatBoost fitted only on each fold's training patients.Final evaluation was performed at the patient level.Source: A. Experimental Setup · p. 5
  • Downstream models used validation-based early stopping and seeds 42–46, with results reported as mean ± standard deviation and paired bootstrap confidence intervals for AUROC comparisons.Backbone comparisons shared the same patient split and preprocessing settings.Source: A. Experimental Setup · p. 5

B. Experimental Result

NeuroECG combines deep ECG features with static covariates and achieves the strongest reported classification performance. ECGFounder also leads the ECG-only backbone comparison, while bootstrap uncertainty qualifies the fusion comparison with the static-only baseline.

B. Experimental Result

  • 0.7689 test AUROC and 0.8574 test AUPRC came from static covariates alone, versus 0.7333 AUROC and 0.8540 AUPRC for the deep-only representation.The results indicate prognostically relevant information in both static covariates and deep ECG embeddings.Source: B. Experimental Result · p. 5
  • 0.7333 test AUROC and 0.8540 test AUPRC were achieved by ECGFounder in the ECG-only comparison, the highest performance among evaluated ECG backbones.After PCA compression, ECGFounder AUROC decreased to 0.7097, while ECG-JEPA, ECG-FM, and ST-MEM had higher AUROC values than their raw representations.Source: B. Experimental Result · p. 5
  • 0.8077 mean test AUROC and 0.8970 AUPRC were achieved by NeuroECG, the best overall classification model; gains over static-only were 0.0388 AUROC and 0.0396 AUPRC.Static covariates improved performance across all evaluated backbones.Source: B. Experimental Result · pp. 5–6
  • The ensemble AUROC difference favoring NeuroECG was 0.0465 (95% CI: -0.0293–0.1263), with the interval including zero.The reported point estimate favored NeuroECG, but the bootstrap interval leaves uncertainty about the difference from the static-only baseline.Source: B. Experimental Result · p. 6

C. Ablation and Sensitivity Analysis

The ablations show that PCA’s value depends on feature fusion, while quantile pooling and gradual unfreezing are favored in this framework. The selected q = 0.24 lies in a competitive low-quantile region.

Feature fusion and PCA compression

  • 0.8077 AUROC and 1.5156 CPC-MAE were achieved by Static + Deep PCA, which reduced fused feature dimension from 1030 to 70.Static + Deep without PCA achieved an AUROC of 0.7931; the reported PCA benefit was specific to fusion.Source: Table III · p. 6, C. Ablation and Sensitivity Analysis · p. 6
  • PCA reduced ECG-only AUROC from 0.7333 to 0.7097, so the held-out results showed no standalone ECG discrimination advantage from compression.The Deep-only configuration achieved an AUROC of 0.7333 and a similar AUPRC of 0.8540 to the static-covariate baseline’s 0.7689 AUROC.Source: C. Ablation and Sensitivity Analysis · p. 6

Temporal pooling strategies

Quantile-threshold sensitivity

  • 0.8093 CV validation AUROC and 0.8945 AUPRC were obtained at q = 0.24; nearby low-quantile settings also remained competitive.Performance generally decreased at larger quantile values, and q = 0.24 was selected based on cross-validation.Source: Table V · p. 6, C. Ablation and Sensitivity Analysis · p. 6

Fine-tuning strategies

D. Model Interpretability and SHAP Analysis

SHAP analysis ranked static and deep ECG features by their contributions to NeuroECG predictions on the test set. Static covariates held the largest individual rankings, while deep ECG importance was distributed across multiple components.

Feature-level SHAP analysis

Feature-group SHAP analysis

  • Static covariates accounted for 41.96% of total importance and deep ECG representations for 58.04%, with higher mean importance per static feature.The authors interpret the deep representation’s contribution as spread across multiple latent components.Source: D. Model Interpretability and SHAP… · p. 7, Figure 2 · p. 7

V. LIMITATIONS

The multimodal model achieved the study’s best overall performance, while the evidence is bounded by I-CARE waveform availability, weak supervision, and latent ECG features.

V. LIMITATIONS

  • External validation was not performed, so the findings have not been tested on independent institutional cohorts.Source: §V · p. 7
  • Because 195 of 607 I-CARE patients lacked usable ECG, the final cohort was restricted to patients with usable recordings.Source: §V · p. 7
  • Patient-level outcome labels used for segment-level fine-tuning create weak supervision, and the evaluated MIL strategy performed less well in this cohort.Source: §V · p. 7
  • PCA-compressed ECG features are latent representations; their relation to clinically interpretable ECG patterns remains a subject for further study.Source: §V · p. 7

VI. CONCLUSION

  • The multimodal model combining deep ECG representations with static clinical covariates improved over clinical covariates alone and achieved the best overall performance in the experiments.Source: §VI · p. 7
Loading 2609.18891v1…