Towards Low-resource Deep Learning for Biomedical Signal Classification: An ECG and EEG Based Approach

Abstract

This study presents two lightweight deep learning pipelines for ECG arrhythmia and EEG emotion classification. These pipelines are designed for resource-constrained environments and can be executed on standard CPU hardware. The proposed framework combines self-supervised representation learning, hybrid neural architectures and GAN-based data augmentation into a single workflow to address limited labelled data and class imbalance. Class-wise GANs are used to balance ECG beats from the MIT-BIH Arrhythmia Database. A ViT–TFN classifier receives the robust latent representations that the SimCLR and MoCo models learn from the enhanced data. This model covers long-range temporal dependencies as well as fine-grained morphological characteristics. The emotion dataset follows the same augmentation and self-supervised learning pipeline before being processed by a GCN–GAT classifier that makes use of attention-based weighting and channel-wise spatial correlations. The EEG model achieves 94.15% accuracy across negative, neutral and positive emotions, the ECG model obtains 92.14% accuracy, with good performance for normal and supraventricular beats. The observed class separability is supported by Principal Component Analysis and tDistributed Stochastic Neighbour Embedding visualizations, which show distinct clusters in the learned feature space. Overall, the findings show that hybrid architectures, augmentation and self-supervised learning can provide competitive performance in real world low-resource settings.