Related Experiment Video
Updated: Sep 19, 2025
06:37
Author Spotlight: Addressing Technical and Subjective Challenges in Measuring Classroom Attention
Published on: December 15, 2023
4.2K
Large Language Model-informed ECG Dual Attention Network for Heart Failure Risk Prediction
Chen Chen1, Lei Li2, Marcel Beetz3
1Institute of Biomedical Engineering, Department of Engineering Science, University of Oxford, Oxford; Imperial College London; University of Sheffield, Sheffield.
IEEE Transactions on Big Data
|June 17, 2025
Summary
This study introduces a novel dual attention ECG network, enhanced by large language model pre-training, for predicting heart failure (HF) risk. The method shows improved accuracy in identifying patients at high risk for HF, particularly those with hypertension or myocardial infarction.
Area of Science:
- Cardiology
- Artificial Intelligence
- Biomedical Engineering
Background:
- Heart failure (HF) presents a growing global health concern with increasing mortality rates.
- Early detection and prevention strategies are crucial for mitigating the impact of HF.
- Predicting HF risk from electrocardiograms (ECGs) is challenging due to complex data and imbalanced risk groups.
Purpose of the Study:
- To develop and validate a novel deep learning model for predicting heart failure (HF) risk using 12-lead ECGs.
- To enhance the model's performance by incorporating large language model (LLM) pre-training and a dual attention mechanism.
- To assess the model's effectiveness in specific patient cohorts, including those with hypertension and myocardial infarction.
Main Methods:
- A lightweight dual attention ECG network was designed with cross-lead and lead-specific temporal attention modules.
- The network utilized LLM pre-training on an ECG-Report dataset for improved feature extraction and to mitigate overfitting.
- The model was fine-tuned and evaluated on two UK Biobank cohorts: patients with hypertension (UKB-HYP) and myocardial infarction (UKB-MI).
Main Results:
- LLM-informed pre-training significantly improved HF risk prediction accuracy in both UKB-HYP and UKB-MI cohorts.
- The dual attention network achieved superior predictive performance compared to existing methods, with C-index scores of 0.6349 (UKB-HYP) and 0.5805 (UKB-MI).
- The attention mechanisms provided enhanced interpretability of the ECG features contributing to risk prediction.
Conclusions:
- The proposed LLM-enhanced dual attention ECG network offers a promising approach for early and accurate heart failure risk prediction.
- This methodology demonstrates potential for advancing clinical risk assessment using complex ECG data, especially in high-risk patient populations.
- The findings highlight the value of integrating advanced AI techniques, like LLMs and attention networks, in cardiovascular disease management.
