Sparsity-constrained one-dimensional convolutional autoencoder for reliable photoplethysmography template quality verification

Photoplethysmography signal quality assessment is essential for ensuring reliable physiological analysis, particularly in real-world conditions where motion artifacts and pathological variations are common. This study proposes a lightweight one-dimensional convolutional autoencoder for automatic classification of photoplethysmography signal quality based on template-level reconstruction error. The method incorporates template-based segmentation and spline interpolation for baseline correction, enabling consistent signal representation prior to model inference. The proposed approach is evaluated on a dataset containing both normal and arrhythmic signals, providing a more realistic assessment compared to conventional studies. Experimental results demonstrate strong performance, achieving an accuracy of 96%, precision of 99%, recall of 96%, and F1-score of 98%. The model shows a high capability to preserve valid photoplethysmography signals while effectively rejecting corrupted segments, which is critical for clinical applications. In addition, the method outperforms a skewness-based signal quality index, particularly in handling arrhythmic cases. The model maintains a compact architecture with approximately 344,611 parameters (1.3 MB), making it suitable for deployment on resource-constrained edge devices.Analysis of the latent space reveals stable and structured feature representations, although partial overlap between normal and arrhythmic samples is observed due to the physiological nature of rhythm irregularities. Overall, the proposed method provides an efficient and clinically relevant solution for photoplethysmography signal quality assessment, with strong potential for real-world applications.