MEDVSE began as collaborative research into estimating heart rate and blood oxygen saturation from smartphone fingertip video. The 2022 preprint is connected to our later journal article, Parameter-Efficient Deep Learning Models for Vital Sign Estimation From PPG, published in Healthcare Technology Letters in 2026.
From video to a vital signal
The pipeline extracts RGB signals from fingertip recordings, preprocesses the signal, and applies compact neural models to estimate vital signs. Reference measurements provide targets for training and evaluation.
The research asks a practical question: how much model capacity is needed to recover useful information from these signals when memory and compute are limited?
Designing for mobile constraints
The 2026 paper explores parameter-efficient models for heart rate, blood oxygen saturation and respiratory rate estimation. It compares compact fully convolutional and DCT-based architectures and demonstrates on-device deployment using TensorFlow Lite. See the paper for the evaluation settings and results for each task.
The MTHS dataset
The project includes the MTHS dataset of fingertip video recordings and reference vital-sign measurements. The original MEDVSE repository provides the earlier project’s code and dataset information; the journal article describes the later research.
Collaboration and publication history
This work was carried out with Mahdi Farvardin and Aboozar Ghaffari. Both the earlier preprint and the 2026 journal record are linked on the research page.