285

Papers

10/10

Roadblocks Active

0

Connections

Researchers built AI agents in which seven major psychiatric conditions — including anxiety, depression, OCD, and PTSD — can each be turned on or off using a single adjustable dial in the agent's reward system, and all seven show a clean dose-response relationship. The disorders self-organize into a two-dimensional map where mania and anxiety sit at opposite poles, mirroring theories from human affective science. This matters because it provides a controllable, preregistered testbed for studying how psychiatric states emerge from reward processing, potentially letting researchers probe drug or therapy mechanisms without human participants.

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

ReadSaveConnections

In 32 treatment-resistant MDD patients undergoing repetitive transcranial magnetic stimulation, how much reward a patient actively pursued on a behavioral task before treatment predicted how much their anhedonia — the inability to feel pleasure — improved afterward. Reward-seeking behavior also increased over the rTMS course specifically in patients who got better, not in non-responders. This is important because it suggests a simple behavioral measure taken before treatment could help identify who will benefit from rTMS, addressing a major gap in personalized treatment selection for treatment-resistant depression.

██████████ 0.9 treatment-resistant-depression Peer-reviewed

Read

An evaluation of eight leading commercial AI systems across 16 DSM-5 psychiatric conditions found that safety guardrails are only reliably effective for suicide and self-harm; for eating disorders, substance use disorder, and major depression, adversarial attacks bypassed protections up to 100% of the time. This is not an abstract technical concern — millions of users interact with these systems for mental health support, and the failure modes map directly onto high-risk clinical populations. The finding creates a clear obligation for developers and regulators to treat mental health safety as a distinct, condition-specific engineering problem rather than a single binary guardrail.

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

ReadSaveConnections

Using data from over 15,000 Chinese adults, researchers compressed sleep timing, napping, physical activity, and social behavior into a single Circadian Rhythm Score that retains almost the same predictive power for depression as the full set of raw behavioral variables. The relationship between circadian disruption and depression risk follows a nonlinear, saturation-like curve, meaning moderate improvements in rhythm may matter more than perfect optimization. A single wearable-derived score could plausibly replace multi-instrument behavioral batteries in large-scale depression screening, substantially lowering the cost of early detection.

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

ReadSaveConnections

By testing eight different ways to model chronic stress in neural networks trained on working memory tasks, the authors found that strengthening inhibitory-onto-excitatory synapses best reproduces the three hallmarks of stress-impaired prefrontal function seen in humans and animals: inhibitory dominance, excitatory hypofunction, and task failure. Networks trained under stress became resilient but at the cost of reduced flexibility when demands exceeded their training conditions. This mechanism-level result gives researchers a concrete synaptic target to investigate in vivo and suggests that resilience may carry a hidden generalization cost.

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

ReadSaveConnections

The researchers identified specific units inside large vision-language AI models that behave like reward-anticipating neurons in the brain's nucleus accumbens, then selectively disrupted those units to produce anhedonia-like behavior: the models shifted toward low-effort, low-reward choices without losing general cognitive ability. The causal perturbation approach — knocking out specific units and measuring behavioral change — is the same logic used in lesion studies in neuroscience. If these AI systems can serve as controllable proxies for reward circuitry, they could help test mechanistic hypotheses about anhedonia that are difficult to run in humans.

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

ReadSaveConnections

Across 140 prediction tasks in the UK Biobank, complex models such as deep networks and gradient-boosted trees were repeatedly matched or beaten by simple linear regression — and the paper provides a mathematically verified explanation: measurement noise erases nonlinear signal structures (like interactions between biomarkers) far faster than linear ones. Specifically, a two-way interaction shrinks by the square of a feature's reliability coefficient, while a linear effect only shrinks once. For mental health biomarker research, this means that noisy measurements of biology or behavior will systematically favor simple models, and improving data quality matters more than algorithmic complexity.

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

ReadSaveConnections

Fine-tuning a 27-billion-parameter language model on therapy conversation transcripts — with labels generated partly by another AI model — produced a system that predicts PHQ-9 depression severity scores with a correlation of 0.80 and detects clinically significant depression (PHQ-9 ≥ 10) with an AUC of 0.91. The iterative pseudo-labeling pipeline allowed the training set to expand from 3,111 to 6,283 labeled users without new human annotation. The catch: the dataset comes from a proprietary commercial platform and cannot be shared, so this result cannot yet be independently verified or widely replicated.

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

ReadSaveConnections

Testing three open-weights language models on German-language clinical interview transcripts, the study found that LLMs can predict depression severity in zero-shot mode with a mean absolute error of 0.60 scale points, while structured feature extraction reduced dementia assessment error by up to 35% over zero-shot baselines. Crucially, automatically generated transcripts from a speech recognition system performed comparably to human transcriptions, suggesting that a fully automated screening pipeline is feasible. The 154-patient dataset and novel observer-rated depression scale are limitations, but the pipeline's use of publicly available models makes it more reproducible than most comparable work.

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

ReadSaveConnections

A randomized sham-controlled pilot trial in 37 college students with insomnia tested whether shining a 980-nm near-infrared laser on the prefrontal cortex for 10 minutes per day over a week could reduce insomnia symptoms. The study used EEG and cognitive tasks alongside self-report to track neural changes, targeting the prefrontal hypoactivity that underlies the hyperarousal loop in chronic insomnia. The sample size is small and the device is proprietary, but the multimodal design and use of drift diffusion modeling for cognitive outcomes represent a methodological step forward for non-pharmacological sleep intervention research.

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

ReadSaveConnections