Can Passive Sensing Replace Questionnaires for Mental Health Prediction?

Status: Manuscript under review at IEEE Journal of Biomedical and Health Informatics (JBHI).

Overview

A three-study head-to-head comparison across three universities (N=1,559) testing whether weeks of continuous passive smartphone and wearable sensing can replace brief personality questionnaires for predicting mental health and academic performance. The central finding: personality beats sensing in all 11 primary comparisons, and sensing adds almost nothing beyond personality. The result holds across traditional ML, deep learning (MLP, 1D-CNN), and time-series foundation-model embeddings (MOMENT).

Research Question

Despite a decade of investment in passive sensing for mental health, the premise that behavioral signals from phones and wearables can replace or augment self-report has rarely been tested against a brief questionnaire in the same people. This project asks the head-to-head question: under what conditions, if any, does passive sensing add value beyond a brief personality questionnaire?

Studies

Together: 3 universities, 3 time periods (2013–2021), 15 study–outcome combinations, 4 ML algorithms plus deep learning, and 44 robustness analyses.

Key Results

Questionnaires dominate at the population level

• Personality wins all 11 primary comparisons (14/15 including the exploratory 28-person sample; the one exception, Study 1 PHQ-9, is attributed to small-sample overfitting). Personality-only R² = .08–.54; sensing-only R² ≤ .01 in every primary comparison
• Two BFI Neuroticism items, selected inside nested CV, reach R² = .36 for Study 2 depression (CES-D) vs. −.12 for 28 raw sensing features. One of the items overlaps almost completely with depression-scale content, so this marks the ceiling of personality’s construct-proximity advantage, not a recommended screening protocol
• Neuroticism is the top SHAP predictor in all 8 mental-health study–outcome–model combinations of Studies 2–3
• Conscientiousness is the top SHAP predictor of GPA across all four model types in the two samples with registrar GPA (Studies 1–2)
• Deep learning does not rescue sensing: MLP, 1D-CNN and MOMENT never beat Ridge; MOMENT foundation-model embeddings R² = −1.0 to −1.7
• Sensing features are highly reliable (ICC(3,k) = 0.73–0.98) — the problem is construct relevance, not measurement quality

Sensing adds little beyond personality

• Sensing added significant value in 1/8 incremental-validity tests (Study 3 BDI-II, ΔR² = .024) and 0/5 DeLong AUC tests after Benjamini–Hochberg FDR
• None of the 44 robustness analyses changed the result; results stayed flat across 7–92-day windows

Within-person signal

In a time-aligned panel (weekly PHQ-4 matched to 7-day sensing windows; 3,149 person-weeks), sensing reached R² = .011 vs. personality .072. Within person, however, 17% of participants (N = 392) had a person-specific sensing–mood R² > .30, so sensing may still be useful for individually calibrated monitoring.

Methodology

Pipeline

• Feature extraction across 13 modalities for Study 1 (87 features); Fitbit + communication for Study 2 (28); Fitbit + phone + GPS for Study 3 (19 curated + 2,597 RAPIDS)
• 4 ML algorithms in parallel: Elastic Net, Ridge, Random Forest, SVR; plus MLP, 1D-CNN and MOMENT foundation-model embeddings
• Repeated 5-fold cross-validation (5 × 5; leave-one-out for Study 1) with permutation tests; nested F-tests for incremental validity and DeLong tests for AUC, each with Benjamini–Hochberg FDR
• SHAP analysis for cross-model feature importance; LPA, mediation, and PLS-SEM for Study 1 supplementary

Robustness

44 supplementary checks covering reliability, ablation, RAPIDS comparison, idiographic prediction, missingness-as-signal, dose–response (7–92 days), within-person tracking, prospective change, cross-study transfer, residualized prediction, reverse prediction, demographic controls, COVID exclusion, and more.

Reproducibility

Three universities, four cohorts (GLOBEM INS-W_1 through INS-W_4), two BFI variants, seven mental health instruments.

Technical Stack

Project Structure

Five-layer script organization: shared utilities → data preparation → core analyses → robustness checks (44 supplementary analyses) → publication materials. Seven-chapter report covering each study individually, cross-study synthesis, clinical utility, and grand synthesis.

Limitations

• All three datasets are college-age cohorts; generalization to clinical or older populations is untested
• “Personality wins” reflects between-person prediction at the population level; idiographic deployment remains an open opportunity
• Sensing modalities are heterogeneous across studies (different devices, sampling rates, derived features)
• 27-person Study 1 result for PHQ-9 (sensing R² = 0.20) does not replicate at scale — small-N overfitting cautionary tale

Authors

Yanuo Zhou (first and co-corresponding author) and Cyrus Su Hui Ho (co-corresponding author), Department of Psychological Medicine, Yong Loo Lin School of Medicine, National University of Singapore.

GitHub Repository

sensing-vs-questionnaires-analysis

Timeline

Duration: November 2025 – April 2026 (Independent research project). Manuscript submitted to IEEE JBHI; currently under review.