JMIR Publications

ABSTRACT

Background:

Despite the rapid expansion of digital phenotyping research, findings remain heterogeneous, with digital phenotyping features showing only modest and inconsistent associations with clinical outcomes. One important reason is that commonly used aggregation-based descriptive and correlational approaches often overlook the temporal dynamics embedded in digital phenotyping data. This has motivated growing interest in network analysis to pair with digital phenotyping, as it enables the modeling of dynamic, within-person interplay among relevant symptoms, behaviors, and contextual factors. However, no scoping review, to our knowledge, has examined the application of network analysis within digital phenotyping research for mental health, despite its theoretical and clinical potential.

Objective:

This scoping review aims to synthesize empirical network-based digital phenotyping studies in the mental health field. More specifically, it aims to identify methodological patterns, limitations, and gaps across four domains: 1) study characteristics, 2) digital phenotyping design, 3) network method characteristics, and 4) analytic applications.

Methods:

Following the PRISMA Extension for Scoping Reviews, this scoping review searched PubMed/MEDLINE, Scopus, Web of Science Core Collection, and PsycINFO (EBSCO) databases. After independent title-and-abstract and full-text screening by two reviewers, 25 data fields were independently extracted from each included study by two reviewers. Finally, descriptive synthesis summarized each extracted field using counts and percentages.

Results:

Among the 1,851 records from database searches, 123 studies published between 2016 and 2026 were finally included. Most samples were young and middle-aged adults from a small set of high-income countries. Active ecological momentary data were included as nodes in all included studies, while passive sensing data were incorporated in only 7.3% of studies. Dynamic multivariate vector autoregressive (VAR) approaches stood out as the dominant method family (74.8%), concentrating on multi-level model fitting and time-invariant parameter estimation. Finally, 91.1% of the studies used network models to describe symptom dynamics or structure, whereas prediction, causal discovery, and clinical application remain rare.

Conclusions:

The integration of digital phenotyping and network analysis is growing in mental health research, but its broader potential is yet to be realized. Current evidence is concentrated in younger and middle-aged populations in high-income countries, and the predominant approach is to estimate time-invariant VAR models on active self-report data. Future research should therefore broaden population representation and integrate more passive behavioral, physiological, and contextual data to comprehensively map the symptom networks. Also, while group-level and multilevel VAR models have more lenient data requirements, person-specific and time-varying models may better capture individual heterogeneity and evolving symptom dynamics, although their application requires sufficient sampling density and longer observation periods to achieve reliable estimates. Finally, future research should move beyond descriptive analyses toward predictive and causal approaches that better support clinical translation, including risk forecasting, personalized intervention, and clinical decision-making.