A Deep Learning Framework Based on Denoising and 2D Image Encoding for Arrhythmia Classification.

Electrocardiogram (ECG) signals are essential for arrhythmia detection; however, they are frequently degraded by noise during acquisition, and their evaluation is vulnerable to data-leakage and patient-overlap issues that can compromise model assessment. Therefore, in this study, we propose an image-encoding-based arrhythmia classifier combined with a morphology-aware denoising autoencoder. We evaluate signals under a corrected, patient-independent protocol in which every model-selection decision was made on a separate validation partition. On a leakage-free test partition, the autoencoder achieved an SNR improvement of 5.63 dB and a correlation coefficient of 0.801, improving R-peak amplitude preservation and redetection accuracy under moderate-to-severe noise while introducing measurable morphology degradation when the input was already lightly contaminated. The proposed model encodes the denoised, beat-centered signal into images through an interleaved-grouping outer product with sorting and flipping, and classifies them with a multi-scale three-dimensional convolutional network. Under this protocol, the proposed model obtained the highest macro-F1 among five image-encoding architectures, but did not outperform four models operating directly on the denoised signal (macro-F1 32.5% versus 37.3-40.0%), indicating that the proposed encoding does not improve overall five-class classification under rigorous inter-patient evaluation. A controlled test showed that the encoding is exactly invariant to global signal-polarity inversion, unlike the sequential models. This targeted invariance, rather than a general accuracy advantage, is the contribution reported here.