283

Papers

10/10

Roadblocks Active

0

Connections

This paper tests whether the pattern of how a person's voice repeatedly revisits similar acoustic states during conversation can signal depression, achieving an AUC of 0.689 on the DAIC-WOZ dataset — outperforming simpler acoustic summary statistics. The core idea is that depression may alter not just average vocal quality but the dynamic structure of how the voice system evolves over time. This matters because it points toward a class of voice features that are more theoretically grounded in the neuroscience of motor control, though the modest AUC and single-dataset validation mean clinical translation is still distant.

██████████ 0.9 depression-biomarkers Preprint

ReadSaveConnections

ADAPTS uses a team of specialized LLM agents that divide a long clinical interview into symptom-specific reasoning tasks, then assemble ratings for depression and anxiety severity — achieving inter-rater agreement (ICC=0.877) that rivals expert clinicians on a subset of high-disagreement cases. This is significant because clinical rating scales require expensive trained raters, and automating this process could scale structured outcome measurement to settings without specialist access. The main caveat is that evaluation relies on small, restricted-access datasets and no code has been released, making independent validation difficult.

██████████ 0.9 depression-biomarkers Preprint

ReadSaveConnections

EviDep introduces a depression severity estimation model that explicitly quantifies its own uncertainty using a statistical distribution (Normal-Inverse-Gamma) rather than outputting a single number, while using wavelet-based signal processing to combine audio, video, and text without redundant overlap. It reports state-of-the-art accuracy on four benchmark datasets including AVEC 2013/2014 and DAIC-WOZ. The uncertainty output is clinically meaningful: a model that knows when it doesn't know could flag cases for human review rather than guessing, which is a practical safety property for any deployment.

██████████ 0.9 depression-biomarkers Preprint

ReadSaveConnections

This paper tests two large vision-language models for depression detection and finds dramatic performance swings depending on whether the data comes from a controlled lab or naturalistic clinical setting — accuracy ranged from 33.9% to 80.4% across the same model family. Both models systematically over-predicted depression, a failure mode with real clinical consequences. The study then applies explainability-guided fairness interventions (including chain-of-thought prompting and counterfactual fairness constraints) that partially correct these biases, demonstrating that how you prompt and constrain a model matters as much as which model you choose.

██████████ 0.9 depression-biomarkers Preprint

ReadSaveConnections

This paper systematically compares different ways of summarizing vocal dynamics over time for depression detection, finding that entropy-based measures — capturing the unpredictability of acoustic trajectories — outperform simple static summaries and trajectory shape features, reaching AUC 0.646 with a statistically significant permutation test (p=0.017). The absolute performance is modest, but the finding that disorder and unpredictability in vocal patterns carries a depression signal is theoretically coherent with known psychomotor slowing in depression. This complements the recurrence-based vocal paper above and together they suggest temporal complexity of speech is worth systematic investigation.

██████████ 0.8 depression-biomarkers Preprint

ReadSaveConnections

PsychBench generates nearly 29,000 synthetic mental health patient profiles using four frontier LLMs and compares their statistical properties to real population surveys (NHANES, NESARC-III), revealing a consistent and alarming problem: LLMs produce individuals who seem plausible in isolation but, as a population, have compressed variance — the severe end of the distribution is dramatically underrepresented. On top of this, 37% of generated cases flip their clinical diagnosis category between two runs, despite high surface-level consistency. This matters because researchers increasingly use LLMs to generate synthetic training data or simulate patient populations, and this work shows those simulations are systematically biased toward the middle.

██████████ 0.8 digital-therapeutics Preprint

ReadSaveConnections

Using 830 Chinese therapeutic writing samples, this paper shows that how a person structures their narrative — the grammar of their story, how ideas connect, how coherent the whole text is — predicts depression, anxiety, and PTSD severity better than simply counting emotionally loaded words (the dominant approach in the field). The analysis uses LLMs to evaluate macro-level narrative structure drawn from linguistics (Labov's story grammar, rhetorical structure theory) and finds these features carry independent predictive signal. This suggests the field has been measuring the wrong thing: the organization of thought, not just its vocabulary, reflects mental state.

██████████ 0.8 depression-biomarkers Preprint

ReadSaveConnections

This narrative review synthesizes evidence that physical exercise improves adolescent mental health by triggering the release of myokines — signaling molecules secreted by contracting muscles that cross into the brain and influence mood, stress response, and neuroplasticity. The framing positions exercise as a non-pharmacological intervention with a specific molecular mechanism, not just a lifestyle recommendation. Caveat: this is an unreviewed working paper on Zenodo, so claims should be treated as a synthesis of existing literature rather than new evidence, and confidence in specific mechanistic claims is limited without access to the full text.

██████████ 0.8 youth-mental-health-crisis Peer-reviewed

Read

Applying unsupervised machine learning to 551 survey participants, this study identifies six distinct user clusters that differ in social media habits and self-reported anxiety, depression, loneliness, and sleep quality — finding a modest but consistent correlation (r=0.28) between social media hours and anxiety. The value is less in the specific numbers, which are exploratory, and more in demonstrating that social media's relationship to mental health is not uniform: different patterns of use (intensity, platform, context) map onto meaningfully different psychological profiles. The small sample and absence of validated clinical scales limit how far conclusions can be pushed.

██████████ 0.8 youth-mental-health-crisis Preprint

ReadSaveConnections

This paper designs a multi-agent LLM pipeline — with separate agents handling worker review, risk assessment, and legal/safety criteria — to screen text for self-harm risk, achieving a 40% reduction in false positives compared to single-agent baselines on the AEGIS 2.0 dataset (FPR 0.095 vs 0.159). The key innovation is a statistical framework that estimates each agent's confidence and uses a bandit algorithm to allocate more processing to ambiguous cases. Lower false positives matter in content moderation and crisis line contexts because over-triggering erodes trust and overwhelms human reviewers, while the maintained false negative rate means fewer at-risk individuals are missed.

██████████ 0.8 digital-therapeutics Preprint

ReadSaveConnections