Variability in Self-reported Depression Symptomology and Associated Mobile-Sensed Behavioral Patterns in Digital Phenotyping: An Observational Study
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Digital phenotyping studies using smartphone-sensed data have identified several behavioral markers associated with depression. However, the generalizability of these markers is limited by the variability in depressive symptoms and associated behaviors, both between and within individuals. While existing research acknowledges this variability, its in-depth analysis remains underexplored.
Objective
This study investigates variability from two angles: (1) variability in self-reported depression symptoms and (2) variability in associated smartphone-sensed behavioral markers. To this end, we examine depression at the symptom level, across diagnostic groups, and by severity level. Associations between depression and behavioral markers are analyzed at the study population, diagnostic group, and individual levels.
Methods
We analyzed a dataset of smartphone-sensed data from 64 patients diagnosed with major depressive episodes belonging to three different subgroups: major depressive disorder (MDD, n=42), borderline personality disorder (BPD, n=12), and bipolar disorder (BD, n=10). Depression severity was assessed by self-reported 9-item Patient Health Questionnaire (PHQ-9) scores. Differences in depression were analyzed at the item and group levels using non-parametric distributional testing, and at the severity level through exploratory data analysis. Variability in behavior–depression associations were analyzed using correlation analysis, and multilevel modeling was used to identify behavioral markers of depression at both between-person and within-person levels, while controlling for demographics and contextual factors.
Results
Our analysis identified group-level differences in five PHQ-9 items: appetite changes, low self-worth, difficulty concentrating, psychomotor disturbances, and suicidal thoughts. Sample level analysis revealed weak associations between 3.1% (3/98) of the features and depression severity. At the individual level, we found mixed correlations for 98% of the features (96/98) across participants, being positive for some and negative for others. Multilevel modeling indicated that between-person differences explained 58.5% of depression severity variance. Adding behavioral features, demographic background, and contextual factors increased the explained variance to 63.7%. Significant within-person behavioral predictors included nighttime communication application usage (β = 0.31, 95% CI = [0.01-0.61], P = .042), morning battery level variability (β = 0.39, 95% CI = [0.06-0.72], P = .019), morning screen usage (β = 0.35, 95% CI = [0.04-0.65], P = .027), average sleep duration (β = 0.39, 95% CI = [0.11-0.67], P = .006), and morning incoming SMS count (β = 0.36, 95% CI = [0.05-0.66], P = .022), all associated with higher depression severity. Daily outgoing SMS count was associated with lower depression severity (β = −0.36, 95% CI = [-0.-0.66], P = .022).
Conclusions
We found differences in self-reported depression severity and associations between behavioral markers and depression severity. The results point out the importance of accounting for the differences in depression symptoms and associated behaviors in analysis; ignoring these differences can lead to misleading conclusions and poor generalization.
