Wavelet-Assisted 1D Convolutional Autoencoder for Efficient ECG Compression with High Reconstruction Fidelity

Abstract

Reliable electrocardiogram compression is essential for reducing storage and transmission requirements in continuous monitoring and bandwidth-constrained healthcare applications while maintaining high waveform reconstruction fidelity. This study proposes a Wavelet-Assisted 1D Convolutional Autoencoder that integrates two-level Daubechies-4 discrete wavelet transform preprocessing with convolutional latent representation learning. Wavelet-based denoising is applied before segmentation to remove high-frequency disturbances. The reconstructed ECG is divided into non-overlapping 256-sample segments and standardized using zero-mean/unit-variance normalization. The 1D-CAE maps each segment to a compact 32-dimensional latent representation and reconstructs the waveform using a decoder with a linear output layer. Evaluation on MIT-BIH Arrhythmia Database records 100–104 shows that WT+1D-CAE achieves the best overall average reconstruction performance across the evaluated WT+AE and FT-AE configurations with MSE of 0.006554, MAE of 0.049382, RMSE of 0.079999, PSNR of 38.376 dB, PRD of 8.000%, and NPRD of 5.658%. The proposed model reduces each 256-sample input to 32 latent coefficients corresponding to an 8:1 representation compression ratio and 87.5% dimensionality reduction. Qualitative waveform comparisons show close agreement between original and reconstructed ECG signals with reconstruction errors mainly localized around rapid high-slope transitions. Statistical analysis indicates significant improvements over WT+AE across the evaluated metrics. The comparison with FT-AE shows significant improvements in MSE and MAE and favorable, although statistically non-significant, differences in RMSE, PSNR, PRD, and NPRD.