Technical Summary: Utilizing Machine Learning to Classify Type of Heartbeat in Athletes from Electrocardiogram Data

Problem Statement

Sudden cardiac events in athletes, while rare, carry severe consequences, necessitating accurate and scalable screening methods. Electrocardiogram (ECG) analysis is the standard for cardiac evaluation; however, its widespread implementation in athletic screening is hindered by clinician workload and the difficulty of distinguishing clinically meaningful arrhythmias from normal variations. Existing deep learning models for ECG analysis often face limitations in inter-patient generalization and interpretability, and they frequently fail to capture sport-specific physiological adaptations. Furthermore, standard datasets often lack the granularity required to map complex clinical annotations to standardized evaluation frameworks. This study addresses the need for an automated, interpretable machine learning framework capable of accurately identifying rare but clinically significant arrhythmias—specifically supraventricular ectopic beats (SVEB) and ventricular ectopic beats (VEB)—to support cardiac screening in athletic populations.

Methodology

Data Source and Preprocessing

The study utilized the PhysioNet ECG Arrhythmia Database, a large-scale repository containing 45,152 patient ECG recordings. Each recording is a 10-second, 12-lead signal sampled at 500 Hz.

• Lead Selection: Three clinically informative leads (I, II, and V1) were extracted to balance diagnostic information with screening efficiency.
• Preprocessing: Signals were resampled to 256 Hz. Preprocessing steps included baseline wander removal (second-order high-pass Butterworth filter, 0.5 Hz cutoff), linear detrending, z-score normalization, and amplitude clipping to ±8.0 mV. Records with flatline segments or missing values were excluded, resulting in a standardized input shape of (2560, 3).

Label Mapping: SNOMED-CT to AAMI

To ensure clinical interpretability and alignment with standard evaluation metrics, fine-grained SNOMED-CT annotations were mapped to the Association for the Advancement of Medical Instrumentation (AAMI) standard.

• Classes Retained: Normal (N), Supraventricular Ectopic Beat (SVEB), and Ventricular Ectopic Beat (VEB).
• Classes Excluded: Fusion (F) and Unknown/Paced (Q) classes were excluded per standard evaluation criteria.
• Distribution: The resulting dataset comprised 69.3% Normal, 22.0% SVEB, and 7.1% VEB, highlighting a significant class imbalance.

Data Splitting and Augmentation

The dataset was partitioned via stratified sampling into 70% training (30,411 records), 10% validation (4,345 records), and 20% test (8,690 records).

• Augmentation: A balanced data generator sampled equal numbers of records per class. Real-time domain-aware augmentation included random temporal shifts (±12 samples), amplitude scaling (0.98–1.02), additive Gaussian noise (σ 0.01–0.03), and low-amplitude baseline modulation.

Model Architectures

Two deep learning models were developed and trained using focal loss (γ=2.0) with class weights (0.15, 0.35, 0.50) to address class imbalance.

1. Residual CNN with Lead Attention:

   • Architecture: A 1D Convolutional layer (64 filters) followed by a custom lead attention mechanism (inspired by squeeze-and-excitation) applying global average pooling and dense layers. The core consisted of three residual blocks (128, 256, 512 filters) with 1×1 projections.
   • Output: Three fully connected layers (256, 128, 64 units) and a softmax output.

2. Bidirectional Gated Recurrent Unit (BiGRU):

   • Architecture: A 1D Convolutional layer (32 filters) and max-pooling layer preceded two stacked bidirectional GRU layers (128 and 64 hidden units).
   • Regularization: Dropout (0.3) within recurrent layers and dense layers (0.4 and 0.3).
   • Output: Two dense layers and a softmax output.

Ensemble Strategy

A weighted ensemble combined the outputs of the CNN and BiGRU. Fusion weights were derived from validation performance using a score formula: 0.3 × macro-F1 + 0.7 × VEB-F1. This weighting prioritized VEB classification due to its association with sudden cardiac death risk. Threshold tuning was applied to the ensemble's VEB probability on the validation set.

Key Results

Models were evaluated on the held-out test set using precision, recall, F1 score, accuracy, and macro-averaged F1.

• Residual CNN: Achieved an overall accuracy of 90.5% and a macro-averaged F1 of 0.820. It performed strongly on Normal (F1=0.954) and SVEB (F1=0.848) classes but showed lower VEB performance (F1=0.659) with a tendency to over-predict VEB events (334 false positives).
• BiGRU: Outperformed the CNN across all classes, achieving 93.0% accuracy and a macro-averaged F1 of 0.844. It improved VEB precision (0.637 vs. 0.585) while maintaining high recall, reducing total false positives to 252.
• Weighted Ensemble: The ensemble achieved the best performance with 93.5% overall accuracy and a macro-averaged F1 of 0.855.
  • VEB Performance: The ensemble reached an F1 score of 0.696 for the critical VEB class, correctly classifying 462 instances while reducing false positives to 241.
  • Normal/SVEB: F1 scores improved to 0.975 (Normal) and 0.894 (SVEB).

Significance and Claims

The paper claims that the proposed machine learning framework demonstrates that automated ECG analysis can provide accurate, interpretable, and scalable screening support for athletic populations.

• Standardization: By mapping SNOMED-CT to AAMI standards, the study enables clinically interpretable classification that aligns with established evaluation protocols.
• Ensemble Efficacy: The weighted ensemble strategy successfully leverages the strengths of both spatial (CNN) and temporal (BiGRU) feature extraction, outperforming individual models and specifically improving the detection of high-risk VEBs.
• Clinical Utility: The authors propose integrating these models as decision-support tools in pre-season physical exams. The system is designed to highlight ECG segments warranting cardiologist review, thereby improving the efficiency and safety of athlete screening without replacing clinical judgment.
• Limitations: The study modestly acknowledges that VEB detection remains challenging, with false positives persisting. It notes that the models were trained on a research dataset and have not yet been validated against prospectively collected clinical ECGs from specific athletic populations, limiting immediate claims regarding real-world generalizability.

The work concludes that combining ECG testing with machine learning-based heartbeat classification offers a pathway to more accurate identification of abnormal cardiac activity, ensuring athletes receive timely and appropriate clinical evaluation.