Digital Phenotyping in Health Research: Scoping Review of Methods, Gaps, and Opportunities

Wearable and connected digital devices continuously generate large volumes of real-world behavioral, physiological, and environmental data, offering new opportunities for health monitoring and personalized care. Digital phenotyping has emerged as a promising paradigm; yet, there is limited consensus regarding how such data should be analyzed. Inconsistent analytical practices and insufficient methodological reporting may compromise reproducibility, comparability, and the validity of findings.

This scoping review examined current analytical practices in digital phenotyping in health research, identified methodological gaps, and highlighted the need for transparency and standardized approaches.

Literature searches were conducted in PubMed, Scopus, Embase, and Web of Science, with Google Scholar used as a complementary source. Eligible studies were published up to December 31, 2024, involved human populations, and used wearable digital devices in longitudinal health research. Studies were screened independently by multiple reviewers using predefined eligibility criteria. Data extraction focused on 6 methodological domains: sample size determination, variable selection and definition, data cleaning and preprocessing, digital phenotyping methods, predictive modeling, and statistical significance handling in large-scale data contexts.

A total of 162 studies were included. Most studies were published from 2018 onward (n=144, 89%) and were conducted primarily in North America (n=92, 57%). Activity trackers (n=81, 50%), smartphones (n=35, 22%), accelerometers (n=26, 16%), and smartwatches (n=24, 15%) were the most frequently used devices. The most common outcomes were activity level (n=67, 41%), sleep (n=65, 40%), step counts (n=63, 39%), and heart rate (n=46, 28%). Heterogeneity and limited reporting were observed across all methodological domains. Preprocessing was the most frequently reported component (n=88, 54%), although specific aspects such as missing-data handling remained inconsistently described (n=35, 22%). Sample size determination methods were reported in only 30% (n=48) of studies, and variable selection methods were described in 27% (n=43). Digital phenotyping approaches were identified in 30% (n=48) of studies and predominantly used regression-based models. Predictive modeling approaches were reported in 17% (n=28) of studies, with substantial diversity in algorithms and limited reporting of validation procedures. Only 4.3% (n=7) of studies explicitly discussed statistical challenges related to large or high-dimensional datasets. Practices varied across health fields, with no domain consistently demonstrating comprehensive reporting across all methodological components.

Digital phenotyping research is expanding rapidly, but methodological practices remain heterogeneous and insufficiently standardized. By providing a cross-domain overview of how wearable-derived longitudinal data are currently processed and analyzed in health research, this review highlights recurring gaps in the reporting and justification of analytical choices, particularly regarding sample size determination, preprocessing, variable definition, predictive modeling, and statistical inference. These findings emphasize the need for clearer analytical frameworks and more consistent reporting practices to improve transparency, reproducibility, and methodological rigor in wearable-based digital health research.

PROSPERO CRD420251234000; https://www.crd.york.ac.uk/PROSPERO/view/CRD420251234000.