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

Jiajun Gao; Yi Zhao; Chenyang Xu; Yuxi Zhou; Hao Wang

cs.LGJul 13, 2026

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

Authors: Jiajun Gao, Yi Zhao, Chenyang Xu, Yuxi Zhou, Hao Wang

Organizations: School of Economics and Management Xidian University Xi’an, China · School of Mechano-Electronic Engineering Xidian University Xi’an, China · School of Cyber Engineering Xidian University Xi’an, China · DCST, BNRist, RIIT, Institute of Internet Industry Tsinghua University Beijing, China

Abstract

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

Figures & tables

Fig. 1: Overview of the proposed ECGFounder-based neurological prognostication framework. Bedside monitoring ECG recordings within the first 72 hours after ROSC are divided into 10-second single-channel segments and encoded by the ECGFounder backbone. Segment-level deep embeddings are aggregated into a patient-level representation using quantile pooling and further compressed by PCA. The compact deep ECG representation is combined with static clinical covariates for patient-level CatBoost prediction.

Backbone	Deep-only (Raw)			Deep PCA-only
Backbone	test AUROC	test AUPRC	test CPC-MAE	test AUROC	test AUPRC	test CPC-MAE
ConvNeXt1D	$0.5714\pm 0.0200$	$0.7382\pm 0.0238$	$1.6865\pm 0.0105$	$0.4884\pm 0.0387$	$0.6734\pm 0.0378$	$1.7156\pm 0.0097$
SE-ResNet	$0.5664\pm 0.0228$	$0.7491\pm 0.0231$	$1.6804\pm 0.0177$	$0.5215\pm 0.0573$	$0.6987\pm 0.0469$	$1.6952\pm 0.0125$
ECG-JEPA †	$0.5828\pm 0.0426$	$0.7567\pm 0.0323$	$1.6787\pm 0.0125$	$0.6510\pm 0.0911$	$0.7857\pm 0.0637$	$1.6832\pm 0.0272$
ECG-FM †	$0.6649\pm 0.0365$	$0.8249\pm 0.0292$	$1.6512\pm 0.0213$	$0.6950\pm 0.0136$	$0.7976\pm 0.0123$	$1.6254\pm 0.0232$
ST-MEM †	$0.5659\pm 0.0763$	$0.7320\pm 0.0651$	$1.7045\pm 0.0097$	$0.6172\pm 0.0353$	$0.7666\pm 0.0299$	$1.6704\pm 0.0125$

TABLE I: Performance comparison of different ECG backbones under the same ECG-only and PCA-compressed evaluation settings.

\dagger

denotes backbones initialized with pretrained weights.

Backbone	Deep PCA + Static
Backbone	test AUROC	test AUPRC	test F1-score	test Recall	test CPC-MAE
SE-ResNet	$0.7134\pm 0.0220$	$0.8014\pm 0.0195$	$0.7989\pm 0.0242$	$0.8143\pm 0.0315$	$1.6204\pm 0.0381$
ConvNeXt1D	$0.7356\pm 0.0225$	$0.8464\pm 0.0147$	$0.7805\pm 0.0396$	$0.7810\pm 0.0259$	$1.6289\pm 0.0383$
ECG-FM †	$0.7655\pm 0.0199$	$0.8601\pm 0.0213$	$0.8037\pm 0.0153$	$0.8095\pm 0.0240$	$1.5508\pm 0.0553$
ECG-JEPA †	$0.7660\pm 0.0121$	$0.8619\pm 0.0073$	$0.7960\pm 0.0171$	$0.8190\pm 0.0358$	$1.5923\pm 0.0515$
ST-MEM †	$0.7680\pm 0.0139$	$0.8758\pm 0.0167$	$0.8175\pm 0.0238$	$0.8333\pm 0.0336$	$1.5537\pm 0.0552$

TABLE II: Downstream outcome classification (AUROC, AUPRC, F1-score, Recall) and cerebral performance category regression (CPC-MAE) results of different backbones under the proposed NeuroECG configuration (Deep PCA + Static).

Features	Dim.	AUROC	AUPRC	CPC-MAE
Static only	6	$0.7689\pm 0.0034$	$0.8574\pm 0.0023$	$1.5762\pm 0.0274$
Deep only	1024	$0.7333\pm 0.0187$	$0.8540\pm 0.0163$	$1.5655\pm 0.0172$
Deep PCA-only	64	$0.7097\pm 0.0108$	$0.8011\pm 0.0231$	$1.5890\pm 0.0202$
Static + Deep	1030	$0.7931\pm 0.0040$	$0.8866\pm 0.0033$	$1.5494\pm 0.0026$
Static + Deep PCA	70	$\mathbf{0.8077\pm 0.0170}$	$\mathbf{0.8970\pm 0.0082}$	$\mathbf{1.5156\pm 0.0364}$

TABLE III: Feature ablation on the test set (mean

\pm

standard deviation).

Configuration	AUROC	AUPRC
Mean pooling	$0.7730\pm 0.0233$	$0.8615\pm 0.0285$
Max pooling	$0.7662\pm 0.0098$	$0.8607\pm 0.0099$
Quantile pooling ( $q=0.24$ )	$\mathbf{0.8077\pm 0.0170}$	$\mathbf{0.8970\pm 0.0082}$
Head-only fine-tuning	$0.6968\pm 0.0181$	$0.8499\pm 0.0125$
Immediate full unfreezing	$0.6213\pm 0.0146$	$0.7759\pm 0.0135$
Gradual unfreezing	$\mathbf{0.7333\pm 0.0187}$	$\mathbf{0.8540\pm 0.0163}$

TABLE IV: Key pooling and fine-tuning ablations on the test set. Pooling uses Static + Deep PCA; fine-tuning uses ECG-only features.

Fig. 2: Test-set SHAP summary for the fused NeuroECG model. The 15 leading features include static clinical variables and PCA-compressed ECG components. Positive SHAP values indicate a larger predicted probability of poor neurological outcome.

Explore similar work

Dynamical Predictive Modelling of Cardiovascular Disease Progression Post-Myocardial Infarction via ECG-Trained Artificial Intelligence Model
May 13, 2026Riccardo Cavarra, Lupo Lovatelli, Shaheim Ogbomo-Harmitt +4Electrocardiogram Foundation Models Cardiac
King’s College London, St Thomas’ Hospital, London, UK · North Bristol NHS Trust, Bristol, UK
ECGQuest: Benchmarking and Fine-Tuning Language Models for Electrocardiography
Aug 31, 2026Mohammadsina Hassannia, Matthew A. Reyna, Reza SameniElectrocardiogram Foundation Models Electrocardiography
Department of Biomedical Informatics Emory University Atlanta, GA 30322, USA · Department of Biomedical Engineering Georgia Institute of Technology and Department of Biomedical Informatics Emory University Atlanta, GA 30332, USA
Information-theoretic Multimodal Representation Learning for Electrocardiogram Signals
May 26, 2026Phu X. Nguyen, Konstantinos Kontras, Wei Dai +5Electrocardiogram Foundation Models Multimodal Representations
1KU Leuven · 3MIT · 4DFKI +1

May 13, 2026, cs.LGJ/K move · Enter open · S save

Dynamical Predictive Modelling of Cardiovascular Disease Progression Post-Myocardial Infarction via ECG-Trained Artificial Intelligence Model

Riccardo Cavarra, Lupo Lovatelli, Shaheim Ogbomo-Harmitt, Shahid Aziz+3

King’s College London, St Thomas’ Hospital, London, UK · North Bristol NHS Trust, Bristol, UK

Electrocardiogram Foundation Models Cardiac Prognostics Artificial Intelligence Models Contrastive Learning

Open paper AskPDF arXiv

Myocardial infarction (MI) is a leading cause of death, and its adverse outcomes are urgent to predict. Yet ECG-based prognostic models underperform because deep learning requires large, labelled datasets, which are scarce in medicine. Foundation models can learn from unlabelled ECGs via selfsupervision, but medically relevant training strategies remain underexplored. We propose a pretrained artificial intelligence model that combines patient-specific temporal information using contrastive learning with supervised multitask heads, then fine-tunes on post-MI outcome prediction. The proposed model outperformed a model trained from scratch (0.794 vs 0.608 AUC) showing that clinically structured ECG modelling improves classification in limited data regimes.

Abstract

Explore similar work

Dynamical Predictive Modelling of Cardiovascular Disease Progression Post-Myocardial Infarction via ECG-Trained Artificial Intelligence Model

ECGQuest: Benchmarking and Fine-Tuning Language Models for Electrocardiography

Information-theoretic Multimodal Representation Learning for Electrocardiogram Signals