Deep Learning-Based Risk Assessment and Prediction of Cardiac Outcomes Using Single-Lead 24-Hour Holter-ECG in Patients with Heart Failure or Myocardial Infarction

Ju Youn Kim, Kyung Geun Kim, Sunghoon Joo, Mineok Chang, Kyoung‐Min Park, Young Keun On, Young Soo Lee, Seung‐Jung ParkJournal of Clinical Medicine 2025 · DOI

TL;DR

This paper introduces a deep learning (DL) model that uses single-lead 24-hour Holter-ECG data to predict cardiac death and ventricular arrhythmias in patients with heart failure or myocardial infarction. The DL model significantly outperforms traditional markers like heart rate turbulence and T-wave alternans, and its predictive power is further enhanced when combined with ejection fraction (EF).

What Problem Does This Paper Address?

Sudden cardiac death (SCD) remains a major concern for patients with heart failure (HF) and myocardial infarction (MI), even those with preserved or mildly reduced ejection fraction (EF). Current risk stratification methods, often relying on EF, are insufficient for identifying all high-risk individuals, leading to a need for improved, non-invasive tools. While deep learning has shown promise in healthcare, applying it to lengthy Holter-ECG data presents challenges due to data volume and complexity. This study aims to bridge this gap by developing and validating a DL model for risk stratification using readily available Holter-ECG data.

Methods

The researchers developed a deep learning model based on a pretrained Transformer architecture. To handle the long duration of Holter-ECG recordings, they used a Masked Autoencoder (MAE) for pretraining on standard 12-lead ECGs, followed by fine-tuning using EF as a surrogate label. The model was designed to process single-lead ECGs by randomly selecting leads during training to improve generalization. The 24-hour Holter recordings were segmented into 9-second intervals, each encoded into a latent vector. A downstream classifier, a compact 1D ResNet, was then trained on these latent vectors to predict the composite outcome of cardiac death and ventricular arrhythmias. The model's performance was evaluated using five-fold cross-validation and compared against conventional markers like heart rate turbulence (HRT), T-wave alternans (TWA), and ejection fraction (EF).

Datasets

The study utilized data from the K-REDEFINE study cohort, a prospective, multi-center registry of 1108 patients hospitalized for acute MI or acute HF. These patients underwent 24-hour ambulatory ECG monitoring. The raw Holter-ECG data was converted into MIT format for DL-based analysis. The study did not introduce a new benchmark dataset but rather leveraged existing data from a well-characterized cohort with adjudicated clinical outcomes.

Key Insights and Ideas

  • Holter-ECG as a Rich Source for DL: The study demonstrates that even single-lead, 24-hour Holter-ECG data, when processed with advanced DL techniques, contains sufficient information for robust cardiac risk prediction.
  • Transformer Architecture's Suitability: The use of a pretrained Transformer model proved effective in encoding lengthy ECG signals into manageable latent vectors, overcoming a key challenge in applying DL to Holter data.
  • EF's Complementary Role: While the DL model alone showed strong performance, combining its predictions with EF significantly improved accuracy, suggesting a synergistic relationship between ECG-derived features and structural/functional cardiac information.
  • Targeting the "Unseen" High-Risk Population: The DL model showed particular promise in identifying high-risk individuals with preserved or mid-range EF, a group often underserved by traditional EF-based risk stratification.
  • Non-Invasive and Scalable Potential: The non-invasive nature of Holter-ECG and the potential for large-scale screening make this DL approach highly attractive for widespread clinical application.

Key Results

  • The DL model achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.74 for the composite outcome of cardiac death and ventricular arrhythmias.
  • Combining the DL model with EF improved the AUROC to 0.77.
  • For cardiac death alone, the DL model achieved an AUROC of 0.79, further increasing to 0.82 when combined with EF.
  • Traditional markers performed less effectively: HRT (AUROC 0.62) and TWA (AUROC 0.55). EF alone achieved an AUROC of 0.75.
  • Model-derived risk stratification identified a seven-fold increase in cardiac death risk in the high-risk group compared to the low-risk group (HR 7.47).
  • This risk stratification was particularly effective in patients with EF > 40%.

Limitations and Future Work

The authors acknowledge several limitations:

  • Low Incidence of Ventricular Arrhythmias: The composite outcome analysis was necessary due to the low incidence of ventricular arrhythmias, potentially obscuring specific predictive differences. Separate analysis for cardiac death alone showed robust performance.
  • Single-Center Cohort: The model was developed and validated on a single national cohort, necessitating external validation to confirm generalizability across different populations and healthcare systems.
  • "Black Box" Nature: The DL model's lack of interpretability is a barrier to clinical adoption. Future work will focus on incorporating explainability techniques (e.g., attention maps, SHAP) to identify key ECG features driving predictions.
  • Three-Year Follow-up: The study focused on early risk stratification. Future longitudinal studies are needed to assess model performance stability over time and determine optimal reassessment frequency.
  • Coarse EF Categorization: The integration of EF into the final risk score used categorical bins, which may limit precision. More sophisticated continuous EF modeling or ensemble methods could be explored.
  • Implementation Considerations: Practical aspects like computational infrastructure, workflow integration, cost-effectiveness, and regulatory pathways for real-world implementation were not addressed.

Future work will involve multicenter validation studies, incorporating explainability techniques, and exploring more sophisticated methods for integrating continuous EF data. The authors also suggest that this DL approach could complement existing clinical workflows and aid in developing personalized management strategies.

Summary updated

Abstract

<b>Background:</b> Deep learning (DL) models using Holter-ECG may enhance risk stratification after heart failure (HF) or myocardial infarction (MI). <b>Objective:</b> To evaluate the prognostic performance of a Holter-based DL model for predicting major adverse cardiac events (MACE), compared with conventional noninvasive markers. <b>Methods:</b> In the K-REDEFINE study, 1108 patients with acute MI or HF underwent 24 h Holter monitoring. A DL model was trained using raw Holter-ECG data and tested for predicting a composite of cardiac death and ventricular arrhythmias. Its performance was compared with heart rate turbulence (HRT), T-wave alternans (TWA), and ejection fraction (EF). <b>Results:</b> During follow-up, 56 adjudicated cardiac deaths (1.18%/yr) and 21 ventricular arrhythmias (0.44%/yr) occurred. The DL model showed an area under the receiver operating characteristic curve (AUROC) of 0.74 (95% CI, 0.70-0.77) for the composite outcome, improving to 0.77 (0.74-0.81) when combined with EF. In comparison, HRT and TWA showed lower AUROCs of 0.62 and 0.55, respectively. For cardiac death alone, the AUROC reached 0.79, further improving to 0.82 with EF. Model-derived risk stratification revealed a seven-fold increase in cardiac death risk in the high-risk group compared to the low-risk group (HR 7.47, 95% CI 2.24-24.96, <i>p</i> < 0.001). This stratification remained particularly effective in patients with EF > 40%. <b>Conclusions:</b> A DL algorithm trained on single-lead Holter-ECG data effectively predicted cardiac death and ventricular arrhythmia. Its performance surpassed conventional markers and was further enhanced when integrated with EF, supporting its potential for noninvasive, scalable risk stratification.
BibTeX
@article{kim2025deep,
  author = {Ju Youn Kim and Kyung Geun Kim and Sunghoon Joo and Mineok Chang and Juwon Kim and Kyoung‐Min Park and Young Keun On and June Soo Kim and Young Soo Lee and Seung‐Jung Park},
  title = {{Deep Learning-Based Risk Assessment and Prediction of Cardiac Outcomes Using Single-Lead 24-Hour Holter-ECG in Patients with Heart Failure or Myocardial Infarction}},
  journal = {Journal of Clinical Medicine},
  year = {2025},
  doi = {10.3390/jcm14207209},
}
Figure