OAK-Mind: Zero-Training Mental Health Change Inference from Passive Sensing via Knowledge-Grounded Multi-Agent LLMs
Authors: Yishu Ji, Zhihan Jiang, Will Ke Wang, Yuang Fan, Anind K. Dey, Xuhai "Orson" Xu
Organizations: Paul G. Allen School of Computer Science and Engineering, University of Washington, Seattle, United States · Biomedical Informatics, Columbia University, New York, New York, United States · Columbia University, New York, New York, United States · Information School, University of Washington, Seattle, Washington, United States
Wearable and smartphone sensors capture behavioral signals associated with mental health, yet inferring mental health outcomes from these signals has traditionally relied on supervised models requiring labeled data and offering limited insight into their predictions. Large language models (LLMs) offer new opportunities to reason over behavioral evidence, but existing LLM-based approaches to passive sensing have largely continued to rely on task-specific fine-tuning or sensor-specific pretraining. In this paper, we introduce OAK-Mind, a zerO-training, multi-Agent, Knowledge-grounded LLM pipeline for inferring mental health changes from passive sensing data. The pipeline organizes sensing data into behavioral dimensions including sleep, physical activity, location mobility, and smartphone use, and orchestrates specialized agents grounded in mental health literature to interpret longitudinal behavioral changes. It synthesizes their assessments and applies questionnaire-based symptom rating to infer changes relative to an individual's baseline assessment with human-readable rationale grounded in observed behavioral patterns and domain knowledge. Across datasets and prediction tasks, while without task-specific training, OAK-Mind achieves comparable or better performance than the most predictive supervised baselines trained on additional labeled samples. It also demonstrate consistent improvement upon direct prompting across four tested LLM backbones. Ablation analyses support the benefits of combining behavioral interpretation with symptom rating, distributing analysis across domain-specific agents, and incorporating domain knowledge. Our findings highlight the potential of knowledge-grounded, multi-agent LLMs for training-free inference of mental health changes from passive sensing data.
Figures & tables
Subscale
Item "In past 2 weeks, how often have you been bothered by:"
Response options
Depression (PHQ-2)
- Feeling down, depressed, or hopeless
0 = Not at all 1 = Several days 2 = More than half the days 3 = Nearly every day
- Little interest or pleasure in doing things
Anxiety (GAD-2)
- Feeling nervous, anxious, or on edge
- Not being able to stop or control worrying
Table 1. PHQ-4 items, subscales, and response options. Item scores are summed within each subscale (range: 0–6).
Figure 1. The overall design of the OAK-Mind pipeline
Domain
Task
Domain Knowledge Summary
Sleep
Depression
Improvement in depression-related sleep often appears as easing of disruptions, reduced variability, fewer extreme nights, or more predictable timing.
Anxiety
Higher anxiety often appears alongside shorter sleep duration, and might coincide with reduced sleep efficiency, more time spent awake after sleep onset, poorer overall sleep quality and more disturbed sleep.
Activity
Depression
Improvement in anhedonia-related activity patterns often appears as reduced inactivity, more consistent daily movement, or partial re-engagement rather than high activity levels.
Anxiety
More activity and higher step counts generally linked to lower anxiety. More sedentary time and lower fulfillment of activity guidelines are associated with higher anxiety.
Mobility
Depression
Depressive symptom improvement may involve reduced withdrawal, slightly increased outings, or more consistent mobility patterns, even if overall movement remains limited.
Anxiety
Lower diversity of places visited (more repetitive, less varied locations) and more time spent at home may be associated with higher anxiety
Table 2. Domain knowledge embedded in each Behavioral Change Agent for the depression and anxiety tasks.
Agent & Role
Input Prompt Structure
Output
①Behavioral Change Agents : Compare a participant’s baseline and follow-up behavioral data
Role: "You are a <domain> behavior expert analyzing wearable-derived <domain> features and their association with depression." Domain knowledge: <Domain knowledge summary> , Data: <Baseline features X1,i > ; <Follow-up features X2,i > . Task: "Compare baseline vs. follow-up <domain> data and describe how symptom-related <domain> patterns evolve…"
< Domain change narrative> × 4
②Cross-Domain Synthesis Agent : Integrate evidence across domains
Role: "You are synthesizing behavioral changes across multiple domains to assess depression state change." Data: < Domain change narrative> × 4; <baseline subscale score s1,i > . Task: "Assess whether symptom-related signals have eased, persisted, or worsened…"
< Cross-domain synthesis narrative>
③Symptom Rating Agent : Map the synthesis onto questionnaire and infer the direction of change
Role: "You are a mental health assessment expert that rates the individual’s mental state changes." Questionnaire: <PHQ-4 items and response scale> Data: < Cross-domain synthesis narrative> ; <baseline subscale score s1,i > Task: "Analyze each symptom dimension based on the behavioral changes. Then make an overall judgment…"
< Predicted change y^i and Rationale ri >
Table 3. Role, prompt input, and output of each agent in OAK-Mind (shown for Depression Task). Quoted text shows the instructions given to each agent; <> denotes the information an agent receives or produces. A set of full prompts is provided in Appendix B .
Figure 2. An illustrative OAK-Mind prediction for one participant in the GLOBEM depression task. The figure shows snippets from the sensing input and agent outputs.
Category
Method
Sensing data representation
ML Methods
Xu et al. ( Xu et al., 2021 )
Daily values categorized as low, medium, or high using the 33rd and 67th percentiles of the corresponding window.
Saeb et al. ( Saeb et al., 2015 )
Mean of each sensing variable over the corresponding window.
Chikersal et al. ( Chikersal et al., 2021 )
Mean and linear slope of each sensing variable over the full 14 days, the earliest seven days, and the latest seven days in the corresponding window.
Borelli et al. ( Borelli et al., 2025 )
Mean, standard deviation, minimum, maximum, linear slope, and proportion of missing days for each sensing variable in the corresponding window.
Ahmed et al. ( Ahmed et al., 2025 )
Mean, standard deviation, interquartile range (IQR), skewness, excess kurtosis, and lag-one autocorrelation for each sensing variable in the corresponding window.
Prior Zero-shot LLM Methods
Kim et al. ( Kim et al., 2024 )
Daily average sensing sequences in each corresponding window.
Table 4. Sensing data representations used in our adaptations of existing ML and LLM methods. All representations are derived separately for the 14-day windows preceding T1 and T2 .
Category
Model
GLOBEM
College Experience
Accuracy ↑
Bal. Acc. ↑
Macro F1 ↑
Wt. F1 ↑
Accuracy ↑
Bal. Acc. ↑
Macro F1 ↑
Wt. F1 ↑
Task: Depression Change ( Improved , Stable or Worsened )
Naive Baseline
Majority-Class Baseline
0.45
0.33
0.21
0.29
0.40
0.33
0.19
0.23
ML Methods
Xu et al. ( Xu et al., 2021 )
0.32
0.28
0.29
0.33
0.31
0.30
0.29
0.29
Saeb et al. ( Saeb et al., 2015 )
0.39
0.39
0.37
0.38
0.38
0.35
0.33
0.35
Chikersal et al. ( Chikersal et al., 2021 )
0.44
0.43
0.40
0.40
0.48
0.44
0.41
0.44
Table 5. Model performance comparison for depression and anxiety change classification on two datasets across traditional machine learning baselines, majority voting, and LLM-based methods. The best overall results are in bold , and the second best results are underlined
Figure 3. Relative performance gains of OAK-Mind over direct prompting on GLOBEM (%) on (a) depression change and (b) anxiety change prediction. For each backbone, OAK-Mind and the plain LLM baseline use the same pretrained model.
Variant
Behavioral Analysis & Synthesis
Domain Knowledge
Domain -Specific Agents
PHQ-4 Filling
GLOBEM
College Experience
Accuracy ↑
Bal. Acc. ↑
Macro F1 ↑
Wt. F1 ↑
Accuracy ↑
Bal. Acc. ↑
Macro F1 ↑
Wt. F1 ↑
Task: Depression Change
Full pipeline
✓
✓
✓
✓
0.56
0.57
0.46
0.48
0.52
0.52
0.51
0.52
Without Domain Knowledge
✓
✓
✓
0.53
0.54
0.48
0.48
0.49
0.51
0.49
0.50
Unified Behavioral Agent
✓
✓
✓
0.46
0.49
0.40
0.41
0.50
0.51
0.50
0.51
Behavioral and Synthesize Only
✓
✓
✓
0.37
0.43
0.32
0.33
0.43
0.44
0.42
0.41
Table 6. Ablation study for depression and anxiety change classification on two datasets. The best results for each dataset and task are in bold .
Figure 4. An illustrative comparison of domain-specific and unified behavioral analysis for a GLOBEM depression instance. OAK-Mind produces separate domain-level assessments before cross-domain synthesis, whereas the unified variant interprets all domains within one narrative. Highlight colors identify the behavioral domains discussed, red text marks summary judgments, and wave underlines identify statements that prioritize sleep changes in the unified interpretation. With synthesis and symptom rating agents, OAK-Mind predicts Stable correctly for this case while the unified variant predicts Worsened .
Mental health assessment relies on episodic self-report scales, which convert subjective states such as stress into numerical scores but provide only sparse snapshots of wellbeing. Wearable devices offer longitudinal behavioral and physiological signals for continuous, low-burden monitoring. Recent LLM-driven personal-health agents enable natural language queries over wearable signals, but mainly handle short-term, retrieval-based lookups (e.g., highest step count over a week). They do not evaluate whether agents can reason over long-term signals to predict wellbeing scores paired with evidence-grounded rationales. To address this gap, we introduce BALMS, the first systematic benchmark of LLM-based agentic systems for longitudinal mental health sensing. BALMS spans 3 real-world longitudinal datasets, 2 task families (closed-form wellbeing-score prediction and rationale generation auto-graded by an LLM-as-Judge), 3 agentic paradigms evaluated across 5 open- and closed-source LLM backbones. We find that zero-shot agents rarely outperform a simple mean baseline, except with stronger backbones or compact, semantically meaningful features. Chain-of-thought prompting improves reasoning-oriented backbones, but does not guarantee temporal grounding or numerical correctness. Together with more analysis on efficiency and temporal scaling, BALMS highlights the need for longitudinal mental health agents that selectively retrieve history, ground temporal evidence, and reason over interpretable behavioral features.
Longitudinal passive sensing enables continuous health prediction, yet models often fail under cross-dataset distribution shifts. Traditional ML overfits cohort-specific artifacts, while Large Language Models (LLMs) struggle to reason reliably over long, heterogeneous time-series. We introduce TimeSRL, a two-stage LLM framework that routes predictions through an explicit semantic bottleneck. The model first abstracts raw signals into high-level natural language, then predicts behavioral outcomes from these abstractions alone. This forces the model to reason over semantic concepts that we argue generalize better than raw numbers. We optimize this process end-to-end using Group Relative Policy Optimization (GRPO) with Reinforcement Learning from Verifiable Rewards (RLVR), learning outcome-aligned abstractions without gold intermediate annotations. Instantiated on mental-health prediction, TimeSRL achieves state-of-the-art performance on a benchmark designed to stress-test cross-cohort generalization under a rigorous leave-one-dataset-out (LOSO) protocol, reducing mean absolute error (MAE) over strong non-LLM ML and LLM baselines by 3.1--10.1% and 9.5--44.1% for anxiety, and 3.2--9.6% and 27.4--57.6% for depression (all ps<0.05). TimeSRL significantly outperforms prior methods in cross-benchmark transfer across different sensing pipelines, rivaling its own within-domain performance without target-domain fine-tuning. These results demonstrate that semantic abstractions are reusable and point to a new direction for generalizable behavior modeling via RL-tuned LLMs.
Yuang Fan, Lilin Xu, Millie Wu +8
Columbia University, USA · University of North Carolina at Chapel Hill, USA · Yale University, USA +3
Large Language Models (LLMs) have shown remarkable capabilities for complex tasks, yet adaptation in medical domain, specifically mental health, poses specific challenges. Mental health is a rising concern globally with LLMs having large potential to help address the same. We highlight three primary challenges for LLMs in mental health - lack of high quality interpretable and knowledge grounded training data; training paradigms restricted to core capabilities, and evaluation of multi turn dialogue settings. Addressing it, we present oMind framework which includes training and aligning LLM agents for diverse capabilities including conversations; high quality ~164k multi-task SFT dataset, as a result of our generation pipeline based on Structured Knowledge retrieval, LLM based pruning, and review actions. We also introduce oMind-Chat - a novel multi turn benchmark dataset with expert annotated turn level and conversation level rubrics. Our diverse experiments on both core capabilities and conversations shows oMind LLMs consistently outperform baselines. oMind-LLM also shows significantly better reasoning with up to 80% win rate.