281

Papers

9/9

Roadblocks Active

0

Connections

Researchers systematically probed eight major proprietary language models against 16 DSM-5 mental health conditions using four adversarial attack styles, finding that safety guardrails reliably work only for suicide and self-harm. For major depression, eating disorders, and substance use disorder, failure rates reached 100%, meaning the models produced potentially harmful content without resistance. This is critical because these three conditions account for the majority of real-world mental health chatbot use, exposing a structural gap between where AI safety investment has gone and where clinical risk actually lives.

██████████ 0.9 digital-therapeutics Preprint

ReadSaveConnections

This review synthesizes 35 years of research to argue that major depression is not a single-pathway disease but the product of simultaneous dysregulation across immune activation (Th1/M1 macrophage), altered tryptophan and lipid metabolism, and oxidative stress — all interacting within a framework the authors call NIMETOX. The key mechanism proposed is an imbalance between the immune-inflammatory response system and the compensatory immunoregulatory system, which together explain symptoms from cognitive slowing to anhedonia. Readers should note this is a narrative review without a systematic search protocol, so the framework should be treated as hypothesis-generating rather than evidence-established.

██████████ 0.9 depression-biomarkers Peer-reviewed

Read

The authors built an AI agent (AG-PPO) where seven psychiatric conditions — anxiety, depression, mania, OCD, PTSD, impulsivity, addiction — can each be switched on and off by adjusting a single parameter that shapes how the agent appraises its situation, producing graded dose-response curves analogous to symptom severity. All seven disorders self-organize into a two-dimensional space where mania and anxiety sit at opposite poles, mirroring dimensional models of human psychopathology. This matters because it offers a controllable, reproducible simulation environment for testing computational theories of how psychiatric conditions arise from altered reward and threat appraisal, potentially accelerating hypothesis testing without patient data.

██████████ 0.9 computational-psychiatry Preprint

ReadSaveConnections

Using data from over 15,000 older adults in China, researchers compressed sleep timing, napping, physical activity, and social behavior into a single Circadian Rhythm Score (CRS) and showed it predicts depression risk almost as accurately as using all the raw behavioral features combined. The compression works because circadian disruption captures a shared underlying mechanism across these behaviors rather than independent risk factors. If validated in other populations, a single wearable-derivable score could replace multi-instrument depression screening batteries, lowering the barrier to large-scale passive monitoring.

██████████ 0.9 sleep-circadian-psychiatry Preprint

ReadSaveConnections

A fine-tuned 27-billion-parameter language model trained on 6,000+ AI therapy conversations predicts PHQ-9 depression scores with a Pearson correlation of 0.80 and detects clinically significant depression (PHQ-9 ≥ 10) with AUC of 0.91 — performance approaching clinician-rated assessments. The model learns passively from conversation text without asking users to complete questionnaires, which could enable continuous background monitoring during therapy chatbot use. Key caveats: the training data is proprietary and from a single platform, and PHQ-9 labels were collected before conversations occurred, meaning the model is predicting baseline severity from subsequent behavior rather than tracking change.

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

ReadSaveConnections

Across 140 prediction tasks using UK Biobank data, deep networks, gradient-boosted trees, and kernel methods were repeatedly matched or beaten by simple linear regression — not because nonlinear models are bad, but because measurement noise in biomedical data mathematically erases complex interaction patterns faster than it erases simple linear ones (formally, noise attenuates a degree-k interaction by reliability-to-the-k-th power). This result is machine-verified in Lean and has direct implications for mental health biomarker research: the premium on complex ML models may be largely wasted until measurement quality improves, and investment in better measurement instruments may outperform investment in better algorithms.

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

ReadSaveConnections

Analyzing 187,000 real ChatGPT conversation histories linked to PHQ-8 scores, researchers found that users with clinically significant depression symptoms used ChatGPT substantially more for mental health support, loneliness, and self-focused conversations, and showed pronounced late-night usage patterns and elevated first-person singular pronouns. This is among the largest naturalistic studies of how people with depression actually use AI chatbots, and it reveals organic help-seeking behavior that bypasses formal care. The behavioral signatures identified — timing, topic, pronoun use — could theoretically enable passive identification of at-risk users, raising both opportunity and significant ethical questions about consent and surveillance.

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

ReadSaveConnections

Instead of treating all users the same, WPG-MoE routes each person's social media posts to one of five specialized neural sub-models (experts) based on how they express distress — whether through direct self-disclosure, narrative episodes, sparse mentions, or mixed patterns. Training uses LLM-scored evidence from nearly 1 million posts as a teacher signal, while deployment relies only on PHQ-9 template similarity, avoiding expensive API calls at inference time. The approach outperforms single-model baselines on both Chinese and English social media datasets, suggesting that individual differences in how depression manifests in language are large enough to warrant personalized detection architectures.

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

ReadSaveConnections

By fine-tuning neural networks originally trained for action recognition on clinical video recordings, the researchers built a model that predicts depression severity from facial video and then used saliency maps to show which facial regions and temporal expression patterns drove each prediction. The approach improves on single-frame benchmarks, suggesting that the dynamics of facial movement over time — not just static expression — carry diagnostic information about depression. Interpretability via saliency maps is a step toward clinical trust, though the AVEC 2014 dataset used (150 participants) is small and the model would need validation in more diverse populations.

██████████ 0.8 depression-biomarkers 🔗 2 cited Preprint

ReadSaveConnections

This randomized sham-controlled pilot trial tested whether shining a 980-nanometer near-infrared laser on the prefrontal cortex for 10 minutes daily over one week could reduce insomnia in college students. The rationale is that insomnia involves prefrontal hypoactivity driving cognitive hyperarousal, and photobiomodulation may boost prefrontal function non-invasively. With only 37 participants this is too small to draw firm conclusions, but the multimodal assessment (sleep questionnaires, EEG, cognitive task) provides a richer signal than most pilot studies, and the approach is relevant for the large minority of students who cannot tolerate or access conventional treatments.

██████████ 0.7 sleep-circadian-psychiatry Preprint

ReadSaveConnections