Title: Enhancing Robustness in Biomedical Time-Series Classification via Innovative Data Augmentation for Deep Learning: Insights from ECG and EEG Studies

Abstract

Comprehensive patient evaluation, particularly within the realm of synchronous monitoring, hinges critically on the accurate and unified analysis of varied biological signals, including Electrocardiograms (ECG) and Electroencephalograms (EEG). While multi-sensor fusion has seen significant progress, a substantial void persists in the creation of unified architectural frameworks capable of effectively processing and extracting features from physiologically distinct signals. Furthermore, traditional methodologies frequently suffer from biased performance outcomes due to the intrinsic class imbalance prevalent in numerous biomedical datasets. To tackle these challenges, this research introduces a novel, unified deep learning framework that delivers state-of-the-art performance across heterogeneous signal categories.

The proposed approach combines a Convolutional Neural Network (CNN) built on ResNet with an attention mechanism. This structure is bolstered by an innovative data augmentation strategy involving the time-domain concatenation of multiple augmented signal variants, thereby generating more enriched data representations. Distinct from previous studies, our methodology scientifically elevates signal complexity to secure superior predictive capabilities, outperforming current state-of-the-art benchmarks. The preprocessing pipeline encompassed wavelet denoising, baseline removal, and standardization. To mitigate class imbalance, the study employed a synergistic approach utilizing the aforementioned advanced data augmentation techniques alongside the Focal Loss function. Additionally, regularization methods were implemented during the training phase to guarantee robust generalization.

The efficacy of the proposed architecture was rigorously tested across three benchmark datasets: the UCI Seizure EEG, MIT-BIH Arrhythmia, and PTB Diagnostic ECG. The model achieved remarkable accuracy rates of 99.96%, 99.78%, and 100%, respectively, underscoring its resilience across varied clinical contexts and signal types. Moreover, the architecture demonstrates high efficiency, requiring approximately 130 MB of memory and processing each sample in roughly 10 milliseconds. These characteristics indicate its strong potential for deployment on resource-constrained wearable devices and low-end hardware.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC