Advanced search
Start date
Betweenand


Hemoglobin Estimation from Smartphone-Based Photoplethysmography with Small Data

Full text
Author(s):
Show less -
Silva, Diego F. ; Junior, Jose G. B. de M. ; Domingues, Lucas V. ; Mazru-Nascimento, Thiago ; Almeida, JR ; Spiliopoulou, M ; Andrades, JAB ; Placidi, G ; Gonzalez, AR ; Sicilia, R ; Kane, B
Total Authors: 11
Document type: Journal article
Source: 2023 IEEE 36TH INTERNATIONAL SYMPOSIUM ON COMPUTER-BASED MEDICAL SYSTEMS, CBMS; v. N/A, p. 4-pg., 2023-01-01.
Abstract

Photoplethysmography (PPG) is a well-known technique to estimate blood pressure, oxygen saturation, and heart frequency. Recent efforts aim to obtain PPG from wearable and mobile devices, allowing more democratic access. This paper explores the potential of using a smartphone camera as a PPG sensor, getting a time series based on the RGB values of video recordings of patients' fingertips. Through this PPG, we apply machine learning for the non-invasive estimation of hemoglobin levels. We assume a realistic scenario where the data has a low volume and potentially a low quality. The generalization capacity of the models built on these scenarios usually achieves undesirable performance. This paper presents a novel dataset that comprises real-world mobile phone-based PPG and a comprehensive experiment on how different techniques may improve hemoglobin estimation using deep neural architectures. In general, cleaning, augmentation, and ensemble positively affect the results. In some cases, these techniques reduced the mean absolute error by more than thirty percent. (AU)

FAPESP's process: 23/02680-0 - Transfer of Learning to Deal with Devices Heterogeneity
Grantee:José Gilberto Barbosa de Medeiros Júnior
Support Opportunities: Scholarships in Brazil - Master
FAPESP's process: 22/03176-1 - Machine learning for time series obtained in mHealth applications
Grantee:Diego Furtado Silva
Support Opportunities: Research Grants - Initial Project