Rare Arrhythmia Diagnosis Using Synthetic ECG Data

Abstract

In cardiovascular healthcare, an accurate diagnosis of cardiac arrhythmias is needed to reduce mortality due to heart diseases and enhance clinical results. Although deep learning-based automated electrocardiogram (ECG) classification has demonstrated encouraging outcomes, the detection of uncommon and potentially fatal arrhythmias is still very difficult to classify because of the scarcity of data and the obvious class imbalance among classes. Due to these constraints, machine learning models frequently perform poorly when it comes to identifying minority classes. For example—third-degree atrioventricular block, ventricular tachycardia, and ventricular fibrillation disorders are difficult to classify because of their rarity and necessitate immediate medical attention (Rajpurkar et al., 2017; Liu et al., 2023). To overcome the lack of data in rare arrhythmia classes, this study suggests a unique architecture that makes use of synthetic data production via a Guided Evolutionary Synthesiser (GES). GES can generate high-fidelity ECG signals with little annotation needed, maintaining class-specific characteristics crucial for clinical differentiation, in contrast to traditional GAN-based methods that require large labelled datasets and are computationally demanding (Abhishek et al., 2018). To improve class representation, the artificial ECG signals are added to the training set after being verified by statistical and expert-driven methods. The synthetic signals are validated and integrated into the training set to increase the availability of data for three distinct rare arrhythmia groups. This research uses a deep neural network adapted to ECG time-series data that is based on ResNet-50 for ECG classification. To guarantee resilience against class imbalance, the model integrates sophisticated training techniques such as class-weighted loss functions and hyperparameter optimisation (Romdhane & Pr, 2020). Each class measures its performance with accuracy, sensitivity, specificity and F1-score with a particular emphasis on improving the diagnosis of under-represented arrhythmias. Results show that the suggested GES-augmented training pipeline significantly improves performance across all parameters for rare arrhythmia classes. The efficacy of the synthetic data augmentation approach is validated by the GES- enhanced model’s better F1-scores and sensitivity for minority classes when compared to the baseline model trained without synthetic data.