Identifying Hypertrophic or Dilated Cardiomyopathy: Development and Validation of a Fine-Tuned ResNet50 Model Based on Electrocardiogram Image

Jiayu Xu, Bo Chen, Weiyang Liu, Wei Dong, Yan Zhuang, Peifang Zhang, Kunlun HeBioengineering 2025 · DOI

TL;DR

This paper presents a deep learning model, a fine-tuned ResNet50, that can identify hypertrophic cardiomyopathy (HCM) and dilated cardiomyopathy (DCM) from standard 12-lead electrocardiogram (ECG) images. The model achieves high accuracy and shows promise as an inexpensive, accessible tool for early detection of these serious heart conditions.

What Problem Does This Paper Address?

Diagnosing hypertrophic cardiomyopathy (HCM) and dilated cardiomyopathy (DCM) can be challenging. While advanced imaging techniques like cardiovascular magnetic resonance (CMR) are accurate, they are often inaccessible, especially in low-resource settings. Electrocardiograms (ECGs) are widely available and inexpensive, but traditional ECG interpretation alone may not be sufficient for definitive diagnosis of these specific cardiomyopathies. The paper addresses the critical need for an accessible, accurate, and automated tool to help identify patients with HCM or DCM, bridging the gap between readily available ECG data and the need for specialized diagnostic capabilities.

Methods

The core of this study is a deep learning approach using a fine-tuned ResNet50 architecture. The researchers started with a pre-trained ResNet50 model and then adapted it for the specific task of classifying ECG images. Key methodological aspects include:

  • Image-based approach: Instead of analyzing raw ECG signal data, the model processes ECGs as images. This is advantageous because ECGs are commonly stored and reviewed in image formats, and it bypasses the need for complex signal preprocessing.
  • Pre-training on non-background ECGs: To focus the model on relevant ECG waveform features and minimize noise from the background grid, an initial pre-training phase was conducted using ECG images with the background removed.
  • Fine-tuning on standard ECGs: After pre-training, the model was fine-tuned on standard 12-lead ECG images. During this phase, most layers of the ResNet50 were frozen, allowing only the final layers to learn from the actual diagnostic images.
  • Data Augmentation: To improve the model's generalization and robustness to variations in ECG images, techniques like random rotation (up to 30 degrees) and random square cutouts were applied to the training data.
  • Interpretability with Grad-CAM: The Gradient-weighted Class Activation Mapping (Grad-CAM) technique was employed to visualize which regions of the ECG image the model found most important for its predictions. This helps understand why the model makes a certain diagnosis.

Datasets

The study utilized a dataset of 2849 ECG images from patients at a large academic hospital. This dataset was divided into training, validation, and testing sets.

  • Patient Cohorts:
    • HCM: 171 ECG images
    • DCM: 364 ECG images
    • Healthy Controls: 2314 ECG images (age- and sex-matched)
  • Temporal Validation Data: An additional set of 170 HCM and 174 DCM patients from the same center were used for prospective temporal validation, collected between January 2022 and December 2023.
  • Ground Truth: Diagnoses for HCM and DCM were confirmed by cardiovascular magnetic resonance (CMR) examinations performed within 30 days of the ECG.

Key Insights and Ideas

  • ECG Images as a Diagnostic Tool: The study strongly supports the idea that standard 12-lead ECG images, when analyzed by a sophisticated AI model, can be a powerful tool for identifying specific cardiomyopathies, complementing or even offering an initial screening for conditions traditionally requiring advanced imaging.
  • ResNet50's Efficacy: Fine-tuning a well-established architecture like ResNet50 proves highly effective for this complex diagnostic task, demonstrating its adaptability beyond general image recognition.
  • Importance of Visual Features: The success of an image-based approach highlights that subtle visual patterns within ECGs, which might be missed or require extensive training for human interpretation, can be effectively learned by deep learning models.
  • Localization of Diagnostic Clues: Grad-CAM analysis revealed that specific leads (V2 and V3, corresponding to anterior and anteroseptal regions) are particularly important for the model's predictions, aligning with known anatomical areas of interest in HCM/DCM diagnosis. This provides a degree of explainability and clinical relevance.

Key Results

The fine-tuned ResNet50 model demonstrated strong performance:

  • DCM Detection:
    • AUROC: 0.996
    • AUPRC: 0.944
  • HCM Detection:
    • AUROC: 0.980
    • AUPRC: 0.953
  • Temporal Validation: The model maintained stability and reliability, showing consistent performance on new, prospectively collected data.
  • Interpretability: Grad-CAM analysis indicated that the V2 and V3 leads were most crucial for predicting both DCM and HCM, with specific features like abnormal QRS waves (for DCM) and high voltage (for HCM) being key discriminators.
  • Distinguishing between conditions: The model effectively distinguished between DCM and healthy controls, and between HCM and healthy controls, with some misclassification between HCM and DCM (13% of HCM patients predicted as DCM).

Limitations and Future Work

The authors acknowledge several limitations:

  • Limited Scope: The model was primarily developed for DCM and HCM, and its generalizability to rarer cardiomyopathy subtypes or conditions with significant comorbidities (like CAD or secondary LV hypertrophy) needs further investigation.
  • Single-Center Study: The study was conducted at a single academic hospital, and the population comprised a significant proportion of referral patients. This may limit the direct applicability to broader, more diverse clinical settings.
  • Comorbidities: Validation in real-world cohorts might involve patients with overlapping comorbidities, which could affect model performance.

Future work suggested includes:

  • Broader Validation: Validating the model in diverse, multi-center cohorts is crucial for establishing its generalizability.
  • Expansion to Other Conditions: Extending the model's capabilities to detect other types of cardiomyopathies, such as Restrictive Cardiomyopathy (RCM) or Arrhythmogenic Right Ventricular Cardiomyopathy (ARVC).
  • Integration into Clinical Workflow: Exploring how this AI-powered tool can be seamlessly integrated into existing clinical workflows to aid cardiologists in early diagnosis and patient management.

Summary updated

Abstract

There is no established detecting tool for hypertrophic cardiomyopathy (HCM) and dilated cardiomyopathy (DCM). This study aimed to develop a deep-learning-based model for identifying HCM and DCM using standard 12-lead electrocardiogram (ECG) images. We obtained a cohort of patients with HCM (171 ECG images) or DCM (364 ECG images), confirmed by cardiovascular magnetic resonance (CMR) examinations, who underwent both ECG and CMR within 30 days at our institution. Age- and sex-matched healthy controls (2314 ECG images) were selected from our Health Check Center. A total of 2849 ECG images were processed via a fine-tuned ResNet50 architecture, with stratified five-fold cross-validation for model training, validation, and testing. The proposed model demonstrated strong performance in distinguishing DCM, achieving an area under the receiver operating curve (AUROC) of 0.996 and an area under the precision-recall curve (AUPRC) of 0.940. For the detection of HCM, the model also achieved an AUROC of 0.980 and an AUPRC of 0.953, respectively. The model prospectively exhibited stability in temporal validation. Furthermore, representative images of the Gradient-weighted Class Activation Mapping (Grad-CAM) technique analysis showed the regions corresponding to the anterior and anteroseptal leads were the most important areas for the prediction of HCM or DCM. This temporally validated fine-tuned ResNet50 model shows promise to inexpensively detect individuals with HCM or DCM.
BibTeX
@inproceedings{xu2025identifying,
  author = {Jiayu Xu and Bo Chen and Weiyang Liu and Wei Dong and Yan Zhuang and Peifang Zhang and Kunlun He},
  title = {{Identifying Hypertrophic or Dilated Cardiomyopathy: Development and Validation of a Fine-Tuned ResNet50 Model Based on Electrocardiogram Image}},
  booktitle = {Bioengineering},
  year = {2025},
  doi = {10.3390/bioengineering12030250},
}
Figure