Deep learning models can effectively reduce the impact of movement artifacts in photoplethysmography (PPG) signals from wearable devices.However, they require large datasets, since they are often trained in a purely datadriven way.We previously proposed an unfolded neural network that integrates prior knowledge of the PPG signal's sparse structure by encoding it into a sparse representation with a learned dictionary of kernels.In this study, we extend this framework by analysing how hyperparameters affect the extracted representations and the model's performance.We train the model using a synthetic dataset based on the PulseDB dataset and test with daily life signals from PPG-DaLiA.Our analysis suggests that the sparsity parameter and the kernel size influence the recruitment of kernels and the sparsity of their temporal activations.Kernel size and dictionary size also affect the denoising performance.On synthetic data, the method improved the signal-to-noise ratio from -7.06 dB to 13.63 dB and reduced the heart rate mean absolute error (MAE) by 74%.On the PPG-DaLiA, it lowered the MAE by 34%.The proposed method outperformed the reference methods.Our findings show that the proposed method effectively improves the quality of PPG signals from wearable devices.