Hemoglobin Estimation from Smartphone-Based Photoplethysmography with Small Data

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.