Representation biases in remote photoplethysmography datasets – a narrative review

Abstract

Purpose – Cardiovascular disease is a global public health challenge, with its rising prevalence severely impacting quality of life and leading to chronic conditions and fatalities. Monitoring heart rate (HR) is a crucial indicator of cardiac health. Remote photoplethysmography (rPPG) and HR monitoring through videos have shown promise for remote healthcare applications. Various deep learning approaches have reported state-of-the-art results on different datasets. While these results are encouraging, their applicability to real-world deployment is limited due to the presence of different types of biases present in these datasets. Design/methodology/approach – We conducted a thorough search on various scholarly databases (e.g., Google Scholar and IEEE Xplore) with keywords “rPPG” or “remote photoplethysmography”, datasets “heart rate” and similar variants. This led us to identify 32 public and private commonly used rPPG datasets. Findings – By doing an extensive literature review on 32 commonly used rPPG datasets, we showed that most of the available datasets do not consider important demographic factors along the dimensions of age, sex, ethnicity and including patients during data collection. The most underrepresented category was patients, as only one dataset included them in the study. This could be related to the ease of data collection; however, it could lead to models that may not be generalizable to other groups, demographics or subsets of populations. Overall, our findings reveal that representation bias is present across various rPPG datasets. Research limitations/implications – Our findings are limited to these 32 rPPG datasets, and further investigation of more datasets could reveal insights into the representation bias problem. Practical implications – Biased rPPG datasets could lead to biased deep learning models that may not generalize in real-world conditions, leading to misdiagnosis and under- or over-estimation of HR estimation. Social implications – Our in-depth analysis provides key insights to researchers to collect inclusive and diverse datasets and develop algorithms to mitigate biases in estimating HR and rPPG among different demographics. Originality/value – This is one of the first literature reviews that delves deeper into common rPPG datasets and evaluates them for the presence of representation bias along the dimensions of age, sex, ethnicity and including patients.