Edge-Aware EEG/ECG Compression Using Dual VQ-Autoencoders with Cross-Modal Attention

Abstract

Continuous remote monitoring of electroencephalogram (EEG) and electrocardiogram (ECG) signals using wearable Internet of Medical Things (IoMT) devices is challenged by the limited bandwidth of low-power wireless communication protocols and stringent power requirements of edge devices. We propose a dual-stream vector-quantized variational autoencoder (VQ-VAE) framework for joint compression of simultaneously collected EEG and ECG signals into compact discrete codebook indices for low-bandwidth transmission, and accurate reconstruction of diagnostic information using a modality-specific medical feature loss function. Separate lightweight one-dimensional convolutional encoders with residual blocks are optimized for edge devices, while separate transposed convolutional decoders are implemented on a cloud server. Each encoder maps 256sample signal windows into 32 codebook indices of a 1024-entry codebook, achieving a 16:1 compression ratio. A bidirectional 8-head cross-modal attention mechanism exploits brain-heart axis correlations for improved reconstruction of both modalities. A composite loss function with modality-specific weighting addresses morphological accuracy for ECG and spectral accuracy for EEG. Evaluated on the CAP Sleep Database, the framework achieves ECG PRD of 8.7% with 99.3% post-reconstruction QRS detection accuracy, and EEG PRD of 22.4% with 0.87 spectral coherence. Ablation experiments confirm that crossmodal attention reduces PRD by 3–5% for both modalities, and modality-specific loss weighting reduces EEG PRD from 84% to 22.4%.