SLPM: a lightweight deep learning model for end-to-end paper ECG digitization

Objective: The digitization of paper electrocardiograms (ECGs) faces several challenges, including amplified errors during segmentation and signal extraction, severe noise interference, and poor generalization under complex conditions. To address these issues, we propose an end-to-end signal location prediction model (SLPM). Approach: SLPM employs a classification-regression joint learning framework to directly predict the presence and vertical coordinate of each signal point, achieving precise mapping from ECG images to time-series signals. A hierarchical squeeze-and-excitation bidirectional long short-term memory (SE-BiLSTM) feature enhancement mechanism is integrated, where SE attention strengthens waveform feature representation and BiLSTM captures lateral temporal dependencies, thereby improving the continuity and stability of signal prediction. Main Results: Experiments on the single-lead datasets PaperECG_Clean and PaperECG_Enhanced, derived from the PTB-XL dataset, demonstrate that SLPM achieves high-accuracy digitization performance even under distortion conditions, with a Pearson correlation coefficient of 0.97 and a signal-to-noise ratio (SNR) of approximately 13.64 dB. On the 12-lead dataset PaperECG_12 l, the model attains an SNR of 14.66 dB with only 0.31 million parameters. Significance: these results indicate that SLPM offers notable advantages in accuracy, efficiency, and generalization, representing a promising new approach for the high-fidelity digitization of paper ECGs.