Analysis of Photoplethysmographic Signals in Low-Dimensional Latent Spaces

Physiological signals are inherently complex yet serve as a critical source of clinical knowledge.Integrating this data into artificial intelligence systems has shown promising capabilities in supplementing usual clinical practices.However, the applicability of these systems may be constrained by the lack of transparency in the algorithms.The input space of photoplethysmography (PPG) signals is sourced from representative datasets with diverse purposes and extraction techniques.These signals are projected into a latent space using unsupervised (Autoencoders, AE), supervised (Fully Connected Neural Networks, FC-NNs), and semi-supervised methods (Uniform Manifold Approximation and Projection, UMAP).The optimal hyperparameters are found with a grid search using holdout cross-validation: reconstruction error for AE, classification loss for FCNN, and the Davies-Bouldin index for UMAP.UMAP demonstrates detailed geometry of the latent space accounting for the periodicity of signals.AE produces a compact representation, enabling reconstruction from hidden variables, while FCNN provides insight into the classification framework.Reduced-dimensionality latent spaces retain amplitude, frequency, periodicity, and morphological insights of PPG signals.