Systematic review: the integration and interpretation of Social Determinants of Health (SDH) in Digital Phenotyping research (DP)

Background: Digital phenotyping (DP) for mental health is an emerging field with the potential to become a powerful tool in clinical settings. However, given the risk of bias or incorrect variable interpretation in marginalized subgroups, for this technology to be used effectively, social determinants of health (SDH) must be integrated into DP research. However it is not known how commonly and in what manner SDH are integrated into DP research. Methods: We conducted a systematic review of studies applying DP in mental health populations to examine how SDH variables are collected, reported, and integrated. The primary objective was to inform the development of practical, standardized guidelines for incorporating SDH into DP research. Eligible studies used DP in clinical populations defined as diagnosed, symptom-based, or at risk for mental health conditions. We excluded studies focused solely on chronic physical illness, age as a risk indicator, neurodegenerative conditions, or tobacco-only substance use, as well as reviews, protocols, and conceptual papers. We searched PubMed, Embase (OVID), and APA PsycInfo (OVID) on october 4th 2024, Web of Science Core Collection on October 23rd 2024, and Scopus on October 24th, 2024. Studies were assessed for quality using the JBI Checklist and the Mixed Methods Appraisal Tool (MMAT). (PROSPERO ID: CRD42024626050).Results: A total of 825 unique papers were screened at the abstract level, 192 underwent full-text review, 74 were accepted for data extraction, and 63 articles were ultimately included in the review.. Considerable heterogeneity was observed in how SDH were reported and integrated into digital phenotyping research. Age, sex, gender, and race/ethnicity were the most frequently reported variables, yet definitions were often inconsistent. Moreover, SDH were most often used solely to describe the characteristics of the target population. Thirty-eight studies (60.3%) used advanced analytic methods, including machine learning. While some studies acknowledged potential demographic or representation biases, formal fairness or bias assessments were rarely conducted. Biases were most often related to age, sex/gender, race/ethnicity, and socioeconomic factors, and behavioral features sometimes had different interpretations across population subgroups. Overall, consideration of bias and the use of fairness or mitigation strategies in DP research remain limited.Conclusion: This review identified several current lacunes in the integration and interpretation of SDH in DP research. Based on our findings, we propose a set of guidelines for incorporating SDH into research projects, organized into three tiers.