Feasibility and Evaluation of Effective Morphological Feature Representation for Contactless Blood Pressure Estimation Using Remote Photoplethysmography

This study evaluates the feasibility of morphological feature-based contactless blood pressure (BP) estimation using remote photoplethysmography (rPPG). A total of 52 morphological features are extracted from each cardiac cycle of rPPG and its first and second derivatives and arranges as temporal sequences for a long short-term memory (LSTM)-based model. In a calibration-based evaluation on a 16-subject public dataset, the 52-feature model achieved mean absolute errors (MAEs) of 11.45 mmHg for systolic BP (SBP) and 8.15 mmHg for diastolic BP (DBP), outperforming direct rPPG waveform input under the same training protocol. Model comparisons also showed that estimation performance varied with the architecture and model capacity. To identify an effective feature representation, recursive feature elimination (RFE) and principal component analysis (PCA) were compared across multiple dimensions. At 24 dimensions, PCA reduced MAE by 1.59 mmHg for SBP and 1.09 mmHg for DBP relative to RFE, while the 15-dimensional PCA representation was selected as the final compact representation considering both performance and feature dimension. The proposed approach was further evaluated against measurement-free and measurement-based reference baselines, while leave-one-subject-out (LOSO) validation was used to assess subject-independent generalization. LOSO showed substantial performance variation across held-out subjects, indicating limited generalizability to unseen subjects. Despite this limitation, the overall results support the feasibility of compact morphological rPPG feature representations for contactless BP estimation and highlight the importance of appropriate feature-space and model selection.