Geographic Disparities and Demographic Reporting in the PhysioNet Open Access Database: Implications for Bias in Biosignal Algorithms

Introduction: Advances in artificial intelligence (AI) have significantly enhanced health monitoring through wearable devices that track various physiological signals. Despite these advancements, AI-based biosignal algorithms face biases that disproportionately affect marginalized groups based on demographic attributes like age and race. This bias often stems from a lack of diversity in datasets used to train these algorithms. Addressing this issue requires examining the sources of data used for training these models. The PhysioNet Open Access Database is one such key resource for biosignal research, providing an extensive collection of physiological signals. However, the extent of demographic representation within this database remains unclear.

Geographic distribution is important in the context of the PhysioNet database because the demographic characteristics of study populations can vary significantly across different regions. Diverse geographic representation ensures that AI models trained on these data are more likely to be generalizable and equitable. For instance, cardiovascular disease risk can differ across geographic regions (e.g., populations in North America versus Asia), and an algorithm trained predominantly on data from one region may not perform as well when deployed in another. Therefore, understanding the geographic distribution of the studies in the PhysioNet database is crucial for assessing the potential biases and limitations of the biosignal algorithms developed using this resource.

We specifically analyze the distribution of studies by continent and the reporting frequencies of demographic factors like race, ethnicity, sex, and age. Understanding these trends is crucial for developing equitable and generalizable biosignal algorithms that can be effectively applied across diverse populations.

Materials and

Methods: The PhysioNet Open Access Database contained 181 studies as of January 2024, of which 175 involved human participants. We collected data on the geographic location, race, ethnicity, sex, and age of participants, as well as the total number of subjects (N) and the field of study. This information was obtained directly from PhysioNet when available or from associated publications when not. In cases of discrepancies, data from PhysioNet was prioritized.

Data were categorized by continent based on the country of the study. Studies with undetermined locations were labeled as "N/A” and were removed from analyses. We used Python scripts to standardize the data and calculate descriptive statistics. The geographic distribution of studies was visualized using a world heat map. Chi-squared tests of independence were conducted to explore associations between demographic reporting and geographic location, with a significance level set at α = 0.05.

To compare the average age of participants across continents, we performed the Kruskal-Wallis test. We also analyzed the mean differences in male and female participation percentages across continents using the Kruskal-Wallis test, treating sex as a binary variable due to the lack of non-binary reporting. Additionally, studies that examined conditions specific to one sex (e.g., pregnancy) were excluded (N = 16). However, studies that included only one sex but addressed conditions applicable to both sexes (e.g., cardiac arrhythmias) were retained.

Results, Conclusions, and Discussions: Our analysis revealed notable geographic disparities in the distribution of PhysioNet studies. North America dominated with 91 studies (55%), while Europe had 56 studies (34%). Asia, Oceania, and South America each had fewer than ten studies, while Africa and Antarctica had none (chi-squared test; p=3.65e-67). This highlights a significant concentration of studies in North America and Europe, with underrepresentation from other continents, which could lead to biases in AI models trained predominantly on data from these regions.

Race reporting was limited, with 17 studies (74%) from North America, 4 from Europe (17%), and 2 from South America (9%) reporting race. Ethnicity reporting was even scarcer, predominantly from North America (13 studies, 93%) and only one from South America (7%) reporting ethnicity. The paucity of race and ethnicity data raises concerns about the generalizability of findings to more diverse populations.

Age was reported in 67 studies (50%) from North America, 50 (37%) from Europe, and fewer from other continents (chi-squared test; p=9.54e-48), suggesting disparities in age reporting practices. The average age of participants varied significantly by continent (Kruskal-Wallis test, p=0.00760), with North America having the highest average age (46.89 years), and South America having the lowest average age (21.43 years).

Sex was reported in 137 studies across all continents, with North America (66 studies, 48%) and Europe (53 studies, 39%) again leading in reporting frequency (chi-squared test; p=4.34e-48). The average difference in percentages of male and female participants also showed variability, with Asia having the highest mean difference (46.25%, indicating more males than females) and South America having nearly equal male and female participation (-0.31%).

These findings highlight significant gaps in demographic reporting and geographic diversity within the PhysioNet database. Increased transparency and standardized reporting are necessary to ensure that biosignal algorithms are equitable and generalizable across diverse populations. Moreover, efforts should be made to increase the representation of underrepresented regions to enhance the robustness and fairness of AI models trained on PhysioNet data.

Acknowledgements (Optional): This work was supported in part by NSF CAREER #2339669.