Skip to main content
Back to timeline
arXivSource publication:

OAK-Mind infers depression and anxiety change without training, beating supervised baselines on two public sensing datasets

Synopsis

Inferring mental health from passive sensing has relied on supervised models trained on labeled data, which are hard to obtain. OAK-Mind routes smartphone and wearable sleep, activity, mobility and phone-use data through several agents grounded in literature-derived knowledge, then a synthesis agent and a symptom-rating agent score the PHQ-4 items, judging whether symptoms improved, stayed stable or worsened relative to a person's own baseline without task-specific fine-tuning or other participants' labels. On GLOBEM and College Experience, depression-change accuracy reached 0.56 and 0.52, above the strongest supervised baselines compared.

Source-provided article image: OAK-Mind: Zero-Training Mental Health Change Inference from Passive Sensing via Knowledge-Grounded Multi-Agent LLMs
arXiv

Interpretation

Mental health change can now be inferred from passive sensing without task-specific training, using only a person's baseline questionnaire and a later sensing window to judge improvement, stability or worsening. Earlier work trained supervised models mapping sensing features to self-reported mental health outcomes, which depends on sufficiently large, high-quality labeled sensing datasets that are difficult to obtain. On GLOBEM, depression-change accuracy rose from 0.47 for the strongest supervised baseline to 0.56, and balanced accuracy from 0.46 to 0.57; on College Experience, accuracy rose from 0.48 to 0.52 and balanced accuracy from 0.44 to 0.52.

The system splits inference into three layers: four behavioral agents separately compare sleep, activity, mobility and phone use between baseline and follow-up windows, a synthesis agent combines cross-domain evidence, and a symptom-rating agent estimates follow-up PHQ-4 item scores and compares them with baseline. Each behavioral agent's prompt carries domain knowledge retrieved from a mental health knowledge graph, whose relations were extracted by an LLM from literature and clinical reference documents and filtered by behavioral-domain and outcome keywords. Evaluation used 124 test pairs from GLOBEM and 40 from College Experience; in the main comparison all LLM methods used GPT-5.4 at temperature 1.2, with each method run five times and means reported.

Perspective

Evaluation covers three-class depression and anxiety change on two public datasets, GLOBEM and College Experience; participants are college students, College Experience is restricted to iOS users, and only assessments with complete 14-day sensing windows are retained.

A rise or fall in self-reported PHQ-2 and GAD-2 scores does not necessarily equal a clinically meaningful change, and the evaluation included no clinical assessments or clinician judgments; whether the intermediate narratives accurately describe the sensing data or help people review predictions was not tested.