[Paper Review] ECG-DelNet: Delineation of Ambulatory Electrocardiograms with Mixed Quality Labeling Using Neural Networks

Guillermo Jiménez-Pérez, Alejandro Alcaine+1 more|arXiv (Cornell University)|May 11, 2020

ECG Monitoring and Analysis4 citations

TL;DR

ECG-DelNet proposes a U-Net-based deep learning framework for delineating P, QRS, and T waves in ambulatory ECGs using mixed-quality labels. Trained on the PhysioNet QT database with data augmentation and regularization, it achieves 98.73% recall and 99.14% precision for QRS waves, matching state-of-the-art DSP methods despite limited data.

ABSTRACT

Electrocardiogram (ECG) detection and delineation are key steps for numerous tasks in clinical practice, as ECG is the most performed non-invasive test for assessing cardiac condition. State-of-the-art algorithms employ digital signal processing (DSP), which require laborious rule adaptation to new morphologies. In contrast, deep learning (DL) algorithms, especially for classification, are gaining weight in academic and industrial settings. However, the lack of model explainability and small databases hinder their applicability. We demonstrate DL can be successfully applied to low interpretative tasks by embedding ECG detection and delineation onto a segmentation framework. For this purpose, we adapted and validated the most used neural network architecture for image segmentation, the U-Net, to one-dimensional data. The model was trained using PhysioNet's QT database, comprised of 105 ambulatory ECG recordings, for single- and multi-lead scenarios. To alleviate data scarcity, data regularization techniques such as pre-training with low-quality data labels, performing ECG-based data augmentation and applying strong model regularizers to the architecture were attempted. Other variations in the model's capacity (U-Net's depth and width), alongside the application of state-of-the-art additions, were evaluated. These variations were exhaustively validated in a 5-fold cross-validation manner. The best performing configuration reached precisions of 90.12%, 99.14% and 98.25% and recalls of 98.73%, 99.94% and 99.88% for the P, QRS and T waves, respectively, on par with DSP-based approaches. Despite being a data-hungry technique trained on a small dataset, DL-based approaches demonstrate to be a viable alternative to traditional DSP-based ECG processing techniques.

Motivation & Objective

• To develop a deep learning-based method for ECG wave delineation that generalizes across diverse morphologies and data quality levels.
• To overcome data scarcity in ECG annotation by leveraging data regularization and augmentation techniques.
• To adapt the 2D U-Net architecture for 1D ECG signals to enable end-to-end segmentation of P, QRS, and T waves.
• To evaluate performance in both single-lead and multi-lead scenarios, assessing generalization across input configurations.
• To demonstrate that deep learning can match or exceed traditional DSP-based delineation methods despite smaller datasets.

Proposed method

• Adapted the U-Net architecture for 1D ECG signals by applying 1D convolutions and transposed convolutions for sequence-to-sequence segmentation.
• Applied strong L2 and dropout regularization to prevent overfitting on the small QT database (105 recordings).
• Used pre-training on low-quality labels to improve model generalization and initial feature learning.
• Implemented ECG-based data augmentation, including heart rate variation, SNR adjustment, and wave shape modulation.
• Evaluated model variants with different depths (3–7 levels) and widths (1–4 blocks per level) to assess capacity impact.
• Trained and validated models using 5-fold cross-validation with standard segmentation metrics (precision, recall, F1-score).

Experimental results

Research questions

• RQ1 Can a U-Net-based deep learning model achieve high-precision ECG wave delineation on small, mixed-quality ECG databases?
• RQ2 How do data augmentation and regularization strategies affect model performance when training on limited annotated data?
• RQ3 Does multi-lead input improve delineation performance compared to single-lead input, despite increased input dimensionality?
• RQ4 To what extent does architectural capacity (depth and width) influence model generalization on small datasets?
• RQ5 Can deep learning-based delineation match or exceed traditional DSP-based methods in terms of precision and recall?

Key findings

• The best-performing model achieved 90.12% precision and 98.73% recall for P waves, 99.14% precision and 99.94% recall for QRS waves, and 98.25% precision and 99.88% recall for T waves.
• The model outperformed or matched state-of-the-art DSP-based methods in both single- and multi-lead configurations.
• Model performance degraded significantly with increased capacity when no regularization was applied, highlighting the importance of regularization over architectural scaling.
• Pre-training on low-quality labels and data augmentation significantly improved generalization, especially in data-scarce settings.
• The use of strong regularization (L2, dropout) and architectural priors (SDo) was more effective than increasing model depth or width.
• Multi-lead input did not yield higher performance than single-lead input, suggesting that lead fusion strategies may mask underlying errors or introduce noise.