Parameter-Efficient Deep Learning Models for Vital Sign Estimation From PPG

Photoplethysmography (PPG) enables non-invasive estimation of vital signs across a range of sensors, from wrist wearables and clinical monitors to smartphone cameras. We present parameter-efficient, end-to-end deep models for heart rate (HR), blood oxygen saturation ( ${\mathrm{SpO}}_2$ ) and respiratory rate (RR) estimation. These models operate directly on PPG signals and integrate preprocessing as a layer. Our family of fully convolutional architectures includes a baseline FCN, a residual FCN, a ConvNeXt-inspired FCN and a compact DCT-based variant. These models achieve strong cross-dataset performance while using only 3.3K-53K parameters. On PPG-DaLiA, the best performing model achieves 6.07 $\pm$ 2.70 bpm MAE for HR. In HR, ${\mathrm{SpO}}_2$ and RR tasks on the BIDMC dataset, the best-performing method achieves error rates of 2.2 bpm, 3.14% and 1.49 breaths per minute, respectively. We also release MTHS, a smartphone-based vital-sign dataset comprising 62 fingertip-PPG recordings (35 men and 27 women) with 1-Hz HR and ${\mathrm{SpO}}_2$ annotations. Additionally, we demonstrate practical feasibility via on-device deployment using TFLite on smartphones. These results show that compact, fully convolutional and DCT-based designs provide accurate, generalizable, and deployment-ready PPG vital-sign estimators across sensors and datasets.