Noise-Adaptive Multi-Scale CNN with Weighted Sampling for ECG Arrhythmia Classification

Abstract

Electrocardiogram (ECG) classification is used for the early detection of cardiac arrhythmias. However, classification performance is degraded by noise in wearable monitoring systems and the severe class imbalance inherent in ECG datasets, leading conventional deep learning models to be biased toward majority classes. To address these challenges, this paper proposes a Noise-Adaptive Multi-Scale Convolutional Neural Network (CNN) with Weighted Sampling for arrhythmia classification under noisy and class-imbalanced conditions. The proposed framework integrates three components: (1) a Multi-Scale CNN with parallel convolution kernels of 3, 5, and 11 samples to capture ECG morphological features at multiple temporal resolutions, (2) a weighted sampling strategy to improve minority heartbeat representation during mini-batch training, and (3) a noise-adaptive training strategy that exposes the network to baseline wander (BW), muscle artefact (MA), and electrode motion (EM) noise at five signal-to-noise ratio (SNR) levels from -6 dB to 6 dB. The framework was evaluated on the MIT-BIH Arrhythmia Database using a patient-independent evaluation protocol. Performance was evaluated under simulated noisy ECG conditions across three noise types and five SNR levels to assess classification robustness under progressively degraded signal conditions. Under noisy conditions, the proposed model consistently achieved the highest average Balanced Accuracy of 29.78% and average Macro F1-score of 26.51%, demonstrating robustness across different ECG noise types and SNR levels. These results indicate that the proposed framework achieves improved balanced heartbeat classification under noisy and class-imbalanced conditions and demonstrates its potential for ECG arrhythmia classification in noisy ambulatory monitoring environments.