Wearable Data and Network Science Team Up to Predict Depression in College Freshmen

The first semester of college is one of the most psychologically precarious stretches in a person’s life. It coincides almost exactly with the peak age of onset for depression, roughly 19.5 years, and it arrives wrapped in upheaval: new social worlds, unfamiliar academic demands, disrupted sleep, and the sudden loss of routines that once structured daily life. Recent data suggest that more than 36 percent of United States college students have reported clinically significant depressive symptoms every year since 2018, and prevalence of mental disorders among college students may range from 12 to 50 percent worldwide. Because nearly two-thirds of American high school graduates go on to college, these figures carry real population-level weight. A newly published study protocol from Kansas State University and the University of Kansas describes an ambitious attempt to catch depression before it takes hold, not by asking students how they feel once a semester, but by tracking the minute-by-minute choreography of their days with a wrist-worn wearable and a battery of statistical tools borrowed from network science.

The project, known as the College Adjustment, Lifestyle and Mental health (CALM) Study, is a 16-week prospective cohort investigation of 144 first-year undergraduates, average age 18.2 years and 54.9 percent female, followed for 108 consecutive days from September 2 through December 19, 2025, during their first academic term. What distinguishes CALM from most prior research is its hybrid panel-burst design. Participants completed five monthly panel surveys measuring broader psychosocial factors such as loneliness, perceived stress, social support, sleep quality, college adjustment, and stressful life events. Nested within that longer arc were five intensive seven-day daily diary bursts, distributed across the semester using a randomized allocation scheme that balanced early, mid, and late timing within each study period. No bursts were scheduled during fall break to avoid contaminating the data with travel-related disruptions and atypical routines.

During each burst, students completed daily diary adaptations of the PHQ-8 depression scale and the GAD-7 anxiety scale each evening, reporting symptoms experienced since waking. The team deliberately decomposed some questionnaire items to sharpen conceptual precision: the psychomotor item was split into two, distinguishing slowed movement and speech from restlessness, and the sleep item was divided into trouble falling or staying asleep versus sleeping more than usual. The result was a 10-item daily depression measure designed specifically for network modeling, in which each symptom can be treated as a distinct node rather than a contribution to an undifferentiated sum. Anxiety was tracked with a six-item daily adaptation that dropped one GAD-7 item due to overlap with the depression measure’s restlessness item, reducing redundancy and participant burden.

The exposure side of the equation is equally novel. Each participant wore a Fitbit Charge 6 continuously, twenty-four hours a day, for the entire 108-day study. The device uses triaxial accelerometry and photoplethysmography-derived heart rate data to estimate daily time spent in sleep, sedentary behavior, light physical activity, and moderate-to-vigorous physical activity. Sleep, in particular, has strong validation support, with Fitbit Charge models showing strong agreement with polysomnography for total sleep time in young adults, and acceptable agreement with research-grade accelerometers for intensity-based activity categories in free-living conditions. Because the study emphasizes within-person, day-to-day variation rather than absolute energy expenditure, the investigators argue that proprietary classification algorithms are unlikely to meaningfully bias their core inferences. Alongside wearable data, the researchers captured daily weather from a Kansas Mesonet station and, remarkably, campus recreation center entry records, linked to participants through facility swipe data and student identification numbers.

The analytic architecture rests on two methodological pillars. The first is compositional data analysis, or CoDA, an approach from time-use epidemiology that treats sleep, sedentary behavior, light activity, and moderate-to-vigorous activity as interdependent components of a finite 24-hour day. Because time spent in one behavior necessarily displaces time in another, these components are perfectly collinear, and traditional regression models that treat them as independent exposures risk producing biased estimates. CALM’s analysts will normalize daily wake-to-wake wear times to 1,440 minutes, handle any zero values with a log-ratio expectation-maximization approach, and transform the compositions into isometric log-ratio coordinates using sequential binary partitioning. By rotating each behavior into the pivot coordinate position in turn, the team can isolate the relative importance of sleep, sedentary time, light activity, or moderate-to-vigorous activity while preserving the co-dependence among all four.

The second pillar is the network approach to psychopathology, a framework that reconceives mental disorders not as latent disease entities producing interchangeable symptoms, but as systems of causally interacting symptoms that influence one another directly and indirectly. Traditional approaches that model depression as a binary diagnosis or an aggregate severity score obscure the well-documented heterogeneity in symptom presentation, even within diagnostic categories. Network models, by contrast, can reveal self-reinforcing cascades such as insufficient sleep leading to fatigue, then to sadness, then to loss of interest, each link representing a potential intervention target. Network science has been established in clinical psychology for over a decade, but its integration with 24-hour movement behavior research has remained strikingly limited, a gap CALM is explicitly designed to close.

The two frameworks converge in multilevel vector autoregressive, or mlVAR, models. These models simultaneously estimate temporal, or lagged, associations, contemporaneous same-day associations, and stable between-person differences, while cleanly separating within-person dynamics from between-person variation. In practical terms, the models will test whether a night of shortened sleep predicts a spike in fatigue the following day, whether an unusually sedentary day precedes low mood, and whether the direction of influence runs the other way, with symptoms reshaping behavior. Centrality indices such as strength, expected influence, betweenness, and closeness will identify which behaviors and symptoms exert the greatest pull on the network. The investigators hypothesize, drawing on prior evidence, that insufficient sleep may be more proximally tied to fatigue and concentration problems, whereas prolonged sedentary time may relate more strongly to anhedonia and low mood.

Two secondary aims extend the design further. First, person-centered clustering techniques, including latent profile analysis and model-based clustering, will identify distinct 24-hour movement behavior profiles, and separate network models will be estimated within each subgroup to test whether vulnerability processes differ across behavioral phenotypes. Second, rolling seven-day calendar windows spanning the full 108-day period will allow time-varying network analyses, tracking how global connectivity, network density, and node centrality evolve across the semester. Network theory holds that systems often become more densely interconnected before transitioning into more severe states, a phenomenon known as critical slowing down. If symptom-behavior networks tighten ahead of symptom escalation, as projects like WARN-D have sought to exploit for early warning signals, CALM’s time-varying networks could flag precisely when risk intensifies, overlaying academic events such as midterm periods and finals week onto the temporal estimates to contextualize the patterns.

The sample was powered for the daily diary component rather than the monthly surveys. Assuming a small within-person effect size of d equal to 0.20 and roughly 30 to 35 daily observations per participant, simulation-based estimates indicated that about 107 participants would provide 80 percent power, and allowing for 30 percent attrition the team targeted 139 enrollees, ultimately recruiting 144. The cohort was predominantly White at 78.5 percent, with 16.7 percent first-generation college students, drawn from Health and Human Sciences, Engineering, and Arts and Sciences among other colleges. At baseline, 80.6 percent reported no prior mental health diagnosis. Students were excluded if they had mobility impairments that would alter gait patterns or if they participated in varsity or club athletics, reducing heterogeneity in structured physical activity exposure. The protocol was approved by the Kansas State University Institutional Review Board, and participants could earn up to $155 in tiered incentives tied to diary completion and valid wearable wear days, keeping the device at study’s end.

The limitations are candidly acknowledged. A single-university, predominantly White sample constrains generalizability, consumer wearables rely on proprietary classification algorithms, lagged associations do not establish causality, and the burst design means brief symptom fluctuations outside assessment windows may have gone unmeasured. Daily suicidality assessment was ethically excluded given the absence of real-time clinical monitoring. Yet the payoff could be substantial. By embedding rigorously modeled, wearable-derived behavioral exposures within time-varying symptom networks, CALM moves the field beyond aggregate depression scores toward identifying which modifiable behaviors destabilize or stabilize symptom networks, in whom, and when. The hybrid panel-burst design also lays a foundation for future predictive modeling of individual risk trajectories and, ultimately, just-in-time adaptive interventions that use passive sensing to deliver personalized behavioral support precisely when emerging adults are most vulnerable. As wearables become ubiquitous among young people, that convergence of precision behavioral modeling and scalable digital intervention may reshape how depression is detected and prevented during one of life’s most consequential transitions.

Subject of Research: Using network science and wearable devices to study temporal relationships between 24-hour movement behaviors and depression during the transition to college.

Article Title: Using network science to examine temporal relationships between 24-hour movement behaviors and depression during the transition to college: protocol for a prospective cohort study

Article References: Using network science to examine temporal relationships between 24-hour movement behaviors and depression during the transition to college: protocol for a prospective cohort study. (n.d.). https://doi.org/10.1186/s44167-026-00102-5

Image Credits: AI Generated

DOI: 10.1186/s44167-026-00102-5

Keywords: 24-hour movement behaviors, depression, college students, network analysis, compositional data analysis, wearable devices, sleep, sedentary behavior, physical activity, intensive longitudinal methods, precision psychiatry, multilevel vector autoregression