TL;DR
Depression is notoriously heterogeneous, yet many digital phenotyping studies attempt to find "universal" behavioral markers. This study demonstrates that the relationship between smartphone-sensed behavior and depression severity is highly individualized and dynamic, with 72% of symptom variance occurring between individuals rather than within them over time.
What Problem Does This Paper Address?
Digital phenotyping—using smartphone sensors to passively track behavior—promises a future of objective, continuous mental health monitoring. However, the field is plagued by inconsistent results: a behavioral marker (like "time spent at home") might correlate with depression in one study but show the opposite effect in another.
The authors argue this inconsistency stems from heterogeneity. Depression isn't a single "state"; it manifests differently across diagnostic subgroups (e.g., Major Depressive Disorder vs. Bipolar Disorder) and varies significantly within a single person over time. This paper seeks to disentangle "between-participant" differences (how Person A differs from Person B) from "within-participant" fluctuations (how Person A’s behavior changes as their mood shifts).
Methods
The researchers analyzed data from the Mobile Monitoring of Mood (MoMo-Mood) study, focusing on 62 outpatients experiencing major depressive episodes.
- Passive Sensing: They used the AWARE app to collect accelerometer data (activity), GPS (mobility), call/SMS logs (sociability), and screen usage (a proxy for sleep and phone interaction).
- Active Sensing: Participants completed the 9-item Patient Health Questionnaire (PHQ-9) biweekly for up to a year.
- Statistical Approach:
- Kendall Rank Correlation: Used to compare population-level associations vs. individual-level variability.
- Rolling-Window Analysis: Examined how correlations between behavior and mood changed over different time scales (28 days to a year).
- Multilevel Modeling (MLM): A stepwise approach to separate fixed effects (demographics) from random effects (individual baselines) and identify which behavioral markers actually predict symptom changes.
Datasets
The study utilized the MoMo-Mood dataset, featuring 62 clinical outpatients categorized into three subgroups:
- Major Depressive Disorder (MDD): n=41
- Borderline Personality Disorder (BPD): n=12
- Bipolar Disorder (BD): n=9
The data included 671 biweekly observations covering 224 behavioral features.
Key Insights and Ideas
- The 70% Rule: The baseline multilevel model revealed that roughly 72% of the variance in depression scores was attributable to differences between individuals. This suggests that a person's "average" depression level is a much stronger signal than the biweekly fluctuations captured by sensors.
- Correlation "Washout": When pooling all participants, behavioral markers showed weak or non-significant correlations. However, looking at individuals revealed strong positive and strong negative correlations. In a global model, these individual signals cancel each other out, leading to the false conclusion that the behavior doesn't matter.
- The SMS Paradox: In their most complex model, "outgoing SMS count" showed a positive effect at the between-participant level (people who text more on average tended to have higher PHQ-9 scores) but a negative effect at the within-participant level (when a specific person texted more than their own average, their depression was lower).
- Symptom Thresholds: Certain symptoms like "psychomotor retardation" and "suicidal ideation" only emerged at high severity levels, whereas "sleep trouble" and "low energy" were constant contributors across the spectrum.
Key Results
- Demographic Predictors: Employment status was the strongest predictor of lower depression severity ($\beta = -4.79, P = .004$), followed by older age ($\beta = -0.12, P = .050$).
- Nighttime Movement: Lower nighttime movement (measured via accelerometer) was a significant marker for reduced symptom severity ($\beta = -0.79, P = .024$ in the full model), likely reflecting more stable sleep patterns.
- Model Performance: Adding time-varying behavioral features to a model that already included demographics only improved the marginal $R^2$ (the variance explained by fixed effects) by about 2.1%. This highlights the difficulty of using passive data to predict subtle mood shifts compared to stable traits.
Limitations and Future Work
The authors acknowledge several hurdles:
- Sample Size: While the longitudinal data is rich, the total number of participants (62) and the small sizes of the BPD and BD subgroups limit generalizability.
- Linearity: The study relied on linear multilevel models. The authors suggest that future work should explore non-linear relationships using techniques like gradient-boosting trees.
- Personalization: The results strongly advocate for idiographic (person-specific) modeling. Instead of asking "What does the average depressed person do?", researchers should ask "What does this person do when they are becoming depressed?"
- Data Sparsity: Biweekly PHQ-9 surveys may be too infrequent to catch rapid clinical changes. Moving toward more frequent "active" sampling could improve model sensitivity.
Summary updated