284

Papers

11/11

Roadblocks Active

0

Connections

Researchers built a reinforcement learning agent with six cognitively grounded 'appraisal signals' and showed that tweaking a single dial per disorder reliably induced recognizable behavioral analogs of anxiety, depression, PTSD, mania, OCD, addiction, and impulsivity. The seven disorder-states self-organized into a two-dimensional space — mirroring findings from human psychopathology research — and recovery required different strategies depending on whether the disorder was reward-based (remove the distortion) or avoidance-based (gradual exposure). This matters because it offers a fast, dose-controlled testbed for studying disorder mechanisms and potential interventions before any human or animal study.

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

ReadSaveConnections

Using data from over 15,000 older Chinese adults, this study compressed multiple behavioral streams — sleep, physical activity, social engagement — into a single Circadian Rhythm Score that retained nearly all the predictive power of the raw measures for identifying depression. A gradient-boosted tree model using this compressed score was interpretable and revealed nonlinear, saturation-like relationships between circadian disruption and depression risk. The practical payoff is a simpler, more deployable screening tool that does not sacrifice accuracy, which matters for large-scale population health monitoring.

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

ReadSaveConnections

An adversarial evaluation of eight widely deployed proprietary LLMs across 16 DSM-5 clinical conditions found that safety guardrails are only consistently reliable for suicide and self-harm — failure rates for eating disorders, substance use disorder, and major depressive disorder reached 100% under multi-turn adversarial prompting using journalism framing or fiction wrappers. The finding is alarming because these same models are already embedded in consumer-facing mental health applications. The study uses an automated LLM judge scoring each model on 8–9 harm dimensions per condition, making the methodology reproducible in structure even if exact replication is constrained by proprietary model access.

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

ReadSaveConnections

A randomized, sham-controlled pilot trial tested seven daily 10-minute sessions of near-infrared light (980 nm) delivered to the prefrontal scalp in 37 college students with insomnia, targeting the prefrontal hypoactivity thought to underlie dysfunctional sleep cognitions. Multimodal outcomes including sleep questionnaires, EEG, and an ERP cognitive task were tracked across three weeks. While the sample size is small and reproducibility is limited by unreported preprocessing details, this is one of the few controlled trials of a non-pharmacological brain stimulation approach for insomnia in a young adult population.

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

ReadSaveConnections

The paper identifies specific units within vision-language models that function analogously to the brain's nucleus accumbens reward circuitry, then uses targeted activation patching to causally induce an anhedonia-like state — the AI begins preferring low-effort, low-reward options without losing general task ability. This is notable not because AI systems have mental illness, but because it demonstrates that reward-anticipation circuits can be functionally isolated and perturbed in ways that mirror validated clinical measures of anhedonia. The approach could provide a rapid, controllable testbed for screening candidate antidepressant mechanisms before animal or human trials.

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

ReadSaveConnections

Analyzing donated ChatGPT conversation histories from 766 participants stratified by depression symptom severity (PHQ-8 threshold of 10), this study found that people with moderate-to-severe depression use the chatbot significantly more for emotional support, loneliness, and interpersonal topics, with pronounced late-night usage and more first-person and absolutist language. The scale — 187,000 conversations — makes this one of the largest naturalistic studies of how depressed individuals interact with AI. A key limitation is that the donated data cannot be shared, so the specific patterns identified here function as hypotheses for future replication rather than settled findings.

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

ReadSaveConnections

By testing eight candidate ways to model chronic stress in recurrent neural networks trained on a working memory task, this study found that strengthening inhibitory-to-excitatory synapses best reproduced three experimentally observed signatures seen in stressed animals: inhibitory dominance, excitatory hypofunction, and impaired performance. Networks trained to be 'resilient' maintained performance under stress but lost flexibility, generalizing poorly to longer memory demands — a potential computational explanation for why stress-resilience can come with cognitive trade-offs. This provides a mechanistic, testable framework linking synaptic changes to behavioral deficits in stress-related disorders.

██████████ 0.8 computational-psychiatry Preprint

ReadSaveConnections

Using PRISMA methodology, this review screened 543 papers down to 45 that describe free-text mental health datasets sourced outside of social media — a category that exists but is severely underutilized relative to Twitter and Reddit corpora. The review finds strong bias toward English-language depression datasets and substantial inconsistency in demographics, annotation methods, and data types across the field. This matters because over-reliance on social media data likely introduces selection bias that limits how well AI screening tools generalize to clinical populations.

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

ReadSaveConnections

This paper argues that psychiatric AI systems commit a form of epistemic injustice by reducing the complex, first-person experience of mental illness to algorithmic labels, effectively dismissing patients as unreliable knowers of their own condition. The analysis uses interpretative phenomenological analysis to show how this reductionism is not merely a philosophical concern but a practical one that could undermine patient trust and therapeutic alliance. For researchers building AI screening or diagnostic tools, this raises a concrete design question: how much of the subjective texture of mental illness can be captured in training labels, and what is lost when it cannot?

██████████ 0.8 digital-therapeutics Peer-reviewed

Read

A four-stage pipeline — perception, knowledge retrieval, chain-of-thought reasoning, and cross-agent auditing — outperformed single-agent LLM baselines on PHQ-8 depression severity prediction from clinical interview transcripts, reducing mean absolute error from 5.35 to 5.02. The cross-agent validation step is the key mechanism: one agent checks another's reasoning trace, reducing the drift and hallucination that plague single-model clinical use. Confidence in the results is limited by the absence of stated model names and statistical significance tests in the available text, but the code is open-sourced, which allows independent scrutiny.

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

ReadSaveConnections