Most attempts to detect depression from phone or social-media behavior answer a yes-or-no question: is this person depressed? Clinicians need more than that. They need to know which symptoms are present, such as loss of pleasure, poor sleep or fatigue, because those drive treatment. Separately, a line of network research has shown that the shape of a person's contact network carries psychological signal that message counts miss. That shape includes how many distinct people they talk to, how fast those people change, and whether their closest contacts stay stable. This proposal asks whether that shape predicts specific depression symptoms, and whether any such signal is truly symptom-specific rather than a proxy for overall severity.
The study uses passively logged calls, texts and messaging with hashed contact identifiers, paired with repeated item-level questionnaire responses. It builds per-person dynamic contact signatures from these logs. It then tests whether those signatures predict the part of each symptom that the overall severity score does not explain, and how many weeks ahead they do so. A positive result would give a privacy-light, text-free route to symptom profiles. A negative result would still be useful: it would show that "fine-grained" passive sensing collapses to a single severity dial, and it would say so cleanly.
Core Claim
The claim is that dynamic ego-network features carry item-specific information about depression symptoms, beyond what communication volume, mobility and the total symptom score already provide. An ego-network is the set of people one person communicates with. The dynamic features are alter turnover (how much the set of contacts changes between windows), persistence of the communication distribution, rank volatility, erosion of the inner circle, and reciprocity. The strongest pre-registered guess is that shrinkage or erosion of the stable inner circle tracks the "lack of interest" and "feeling down" items. A falling ratio of outgoing to incoming contact is expected to track the fatigue items, and the sleep and appetite items should show little topological signal. What would be new is not "networks predict depression." It is a test of which symptoms networks predict, at which lead time, and whether that prediction is anything more than a disguised severity estimate.
What Exists and Where It Stops
Three strands motivate the gap.
The first uses network structure to predict mental state, but always a coarse one. Predicting Depression via Social Media found that people heading into depression contract into dense, insular ego-networks. It predicted only a binary at-risk label. Heterogeneous Subgraph Network with Prompt Learning for Interpretable Depression Detection on Social Media maps posts onto the 20 items of the Zung depression scale using prompted language models, combines this with interaction subgraphs, and reports 94.9% accuracy. It then collapses everything back to depressed versus not, and it degrades on low-activity users. Social network structure is predictive of health and wellness shows that centrality and clustering add large relative F1 gains over wearables for four wellness surveys. Those surveys are single-item scales, not symptom inventories.
The second strand supplies the method. Personality traits and ego-network dynamics defines turnover as the Jaccard overlap of active contacts across windows, persistence as the Jensen-Shannon divergence of communication distributions, and rank dynamics. It shows these separate personality traits. For example, the median Jaccard overlap is 0.210 for high-Openness people versus 0.259 for low-Openness people, and the inner circle stays stable overall. We reuse these measures, but point them at symptoms rather than traits. Network analysis of the NetHealth data shows that 74% of participants' local centralities shift over time, which is enough within-person variation to model. It treated depression as a static baseline.
The third strand constrains the design. Differential temporal utility of passively sensed smartphone features predicts PHQ-8 and GAD-7 (standard depression and anxiety questionnaires) 0, 1 and 2 weeks ahead in 1,013 people. It found that home duration is a distal warning sign (β = 0.219) and that app messaging, not calls or texts, tracks depression. Its communication features were raw counts, and its target was the total score. Our lead-time design copies its structure so the two are directly comparable.
On the diagnostic-granularity side, Towards Identifying Fine-Grained Depression Symptoms from Memes and Your Model Is Not Predicting Depression Well And That Is Why establish why symptom-level labels matter. They also show how fragile crowd-sourced symptom labels are: crowd-versus-clinician agreement on anhedonia reached a Cohen's kappa of only 3–9%. That fragility is why we use self-reported questionnaire items as targets rather than annotated text.
One honest limitation shapes everything. A single person's phone log records ties between that person and their contacts, but not ties among the contacts. Clustering coefficients, the central feature in De Choudhury's work, therefore cannot be computed from individual logs. Our main feature set is restricted to measures that need only the person's own log. Clustering is tested only in the NetHealth sub-study, where many contacts are themselves enrolled participants.
Research Program
Features. For each participant and each rolling window, we compute a dynamic signature from hashed contact identifiers. It includes:
- number of active contacts;
- entropy of the communication distribution;
- turnover and persistence relative to the previous window;
- rank volatility;
- inner-circle stability, measured as the retention of the top-10 contacts;
- per-contact reciprocity, meaning the ratio of initiated to received contact;
- the share of communication going to newly appearing contacts.
Features are computed separately per channel (calls, SMS, app messaging). The LifeSense result suggests the channels carry different signals.
Targets. Targets are the eight item scores of the PHQ-8, each on a 0–3 scale, at each assessment.
Separating symptom-specific signal from severity. The central analytic move is to split each item into a shared part and an item-specific part:
$$y_{it,k} = \lambda_k\, g_{it} + u_{it,k},$$
Here $y_{it,k}$ is person $i$'s score on item $k$ at assessment $t$. The term $g_{it}$ is a latent general severity factor estimated from all eight items, $\lambda_k$ is item $k$'s loading on it, and $u_{it,k}$ is the item-specific remainder. "Granular" prediction is only credible if network features predict $u_{it,k}$. Predicting $y_{it,k}$ alone could be achieved entirely through $g_{it}$.
Model. We fit mixed-effects ordinal models with person-mean-centered features, so that effects reflect within-person change rather than stable between-person differences. This partly addresses the homophily-versus-influence confound noted in the network literature. A small gradient-boosted model is fit alongside as a nonlinear check.
Lead times. Lead times follow the LifeSense scheme: concurrent, 1 week ahead and 2 weeks ahead.
Data access. This must be verified first. The study needs LifeSense logs with hashed contact identifiers and item-level PHQ-8 responses. If identifiers are unavailable, NetHealth becomes the primary cohort, provided its semestral surveys include item-level depression responses across waves. If neither holds, the project cannot run as designed. We treat this as a go/no-go check before any modeling.
Baselines
The naive reference point has two parts. The first predicts each item from the person's own previous item score. This carry-forward baseline is hard to beat in repeated-measures data. The second predicts each item as its population loading times the person's predicted total score. This "severity-only" baseline embodies the hypothesis that granularity adds nothing.
The strongest existing competitor is the LifeSense feature set from Differential temporal utility of passively sensed smartphone features: home duration, circadian movement, app-category use, and call, SMS and messaging counts. It is the strongest because it was built on the same cohort, targets the same questionnaire, and already captures the withdrawal that topology might merely restate. That paper reports regression coefficients, not predictive accuracy. There is no reported result found on Lacuna for item-level PHQ-8 prediction from passive sensing, so we re-implement this feature set ourselves. The re-implemented model uses the same participants, windows, model family, cross-validation folds (split by person and by time) and hyperparameter budget as our model.
For context only, not direct comparison, the closest reported multi-label symptom result is a Hamming loss of 0.388 for PHQ-4 items from interview text on DAIC-WOZ, from Passive Diagnosis Incorporating the PHQ-4 for Depression and Anxiety. Hamming loss is the fraction of item labels predicted wrongly.
The key ablation is an identity-permutation control. Within each person, contact identifiers are randomly relabeled between consecutive windows. This keeps every per-window volume, size and entropy value fixed but destroys turnover, persistence, rank dynamics and inner-circle stability. Any gain that survives this shuffle came from counts, not from the structure of who is contacted.
Beating the baselines means improving held-out prediction of the item-specific remainders $u_{it,k}$ over LifeSense features plus the severity-only baseline. The improvement must be significant under person-level bootstrap and must vanish under identity permutation. We report per-item mean absolute error and within-person correlation of predicted versus observed change.
Why not just…?
Why not just count calls and messages? This is the most serious threat. LifeSense shows that counts already carry signal. People who call less may simply also call fewer distinct people, making topology redundant. The identity-permutation ablation settles this directly because it holds counts fixed. If our gains disappear under it, we will report that ego-network dynamics add nothing beyond volume for symptom prediction. The MTL finding that turnover separates personality traits is suggestive, but it involved traits, not symptoms, and it does not settle the question.
Why not just predict the total score and read items off it? If depression symptoms move together within a person, item-level prediction is just severity prediction relabeled. Several of the "fine-grained" results in the file could be explained this way, since none separate shared from item-specific variance. The decomposition above is designed to catch this. If topology predicts only $g_{it}$, the honest conclusion is that passive network sensing supports staging but not symptom profiling. That is still a useful, publishable boundary on the granularity agenda.
Why not just use text, as the social-media symptom models do? Phone logs contain no text, and collecting message content is far more invasive. Text-based symptom labels are also unreliable: the PRIMATE re-annotation found crowd-versus-clinician kappa of 3–9% on anhedonia. Clinical-text models also transferred badly to Reddit, with accuracy 18 points below the majority-class baseline in Passive Diagnosis Incorporating the PHQ-4 for Depression and Anxiety. Text is not a like-for-like alternative in this setting. We do not claim topology beats text where both exist; the file contains no dataset with both network structure and item-level labels to test that.
Why not just add GPS mobility? Home duration may capture the same withdrawal. We include it in the strongest baseline rather than argue it away, so any reported gain is incremental over mobility.
Decisive Checks
The first check is data feasibility: are contact-level hashed logs and item-level PHQ-8 responses available together?
The second is the headline test: do dynamic ego-network features improve held-out prediction of item-specific remainders over LifeSense features, and does that improvement disappear under identity permutation? A pattern where gains concentrate on the interest and mood items and not on sleep or appetite would support the symptom-specificity hypothesis. Uniform small gains across all items would point back to severity.
The third is lead time: does inner-circle erosion appear two weeks before the change in the interest item, the way home duration does for total severity?
The fourth runs only in NetHealth, where contacts are co-enrolled. It tests whether true clustering adds anything beyond the ego-only signature. This tells future passive-sensing studies whether they need network data from contacts at all.
Every analysis reports results for low-activity users separately, because the subgraph literature shows graph signals degrade there.
References
- Predicting Depression via Social Media
- Heterogeneous subgraph network with prompt learning for interpretable depression detection on social media
- Social network structure is predictive of health and wellness
- Personality traits and ego-network dynamics
- Network analysis of the NetHealth data: exploring co-evolution of individuals' social network positions and physical activities
- Differential temporal utility of passively sensed smartphone features for depression and anxiety symptom prediction: a longitudinal cohort study
- Towards Identifying Fine-Grained Depression Symptoms from Memes
- Your Model Is Not Predicting Depression Well And That Is Why: A Case Study of PRIMATE Dataset
- Passive Diagnosis Incorporating the PHQ-4 for Depression and Anxiety