282

Papers

10/10

Roadblocks Active

0

Connections

A multi-agent AI pipeline autonomously analyzed wearable sensor data from over 9,000 participants and surfaced 41 candidate digital biomarkers for depression, with circadian instability features (sleep duration and onset variability) replicated across two independent cohorts. This matters because manual biomarker discovery is slow and hypothesis-driven; an automated system that can consistently identify candidates across independent datasets accelerates the path to passive, scalable mental health monitoring. The major caveat is that two of three datasets are proprietary to Google, limiting external validation.

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

ReadSaveConnections

Rather than measuring the average loudness or pitch of speech, this paper models the underlying dynamical structure of how vocal patterns evolve over time — capturing irregularities invisible to standard acoustic features — and achieves AUC 0.689 in depression detection, beating all static and other nonlinear baselines on the DAIC-WoZ corpus. The mechanism is that depression alters the trajectory of vocal state over time, not just mean acoustic properties, suggesting that the temporal fingerprint of speech carries clinically relevant information. Confidence is limited by a small depressed sample (n=42) and the absence of proper phase-space reconstruction, which is standard practice in nonlinear dynamics.

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

ReadSaveConnections

Adolescent alcohol exposure in rats produced persistent disruptions to rest/wake cycles that continued into adulthood, and blocking the orexin-1 receptor (OX1R) with SB-334867 dose-dependently reduced both alcohol consumption and motivation to drink. This is mechanistically significant because it implicates the orexin system — a key regulator of arousal and reward — as a potential pharmacological target linking adolescent alcohol use, sleep disruption, and adult addiction vulnerability. The findings are pre-clinical but provide a concrete neurobiological circuit to investigate in human adolescent substance use and co-occurring sleep disorders.

██████████ 0.9 sleep-circadian-psychiatry Peer-reviewed

Read

EviDep introduces a depression severity estimation model that not only predicts a score but also quantifies its own uncertainty using a statistical framework (Normal-Inverse-Gamma distributions), while separating redundant information across audio, video, and text modalities before fusing them. This matters clinically because a model that knows when it is uncertain is far safer to deploy than one that always outputs a confident prediction — uncertainty estimates could flag cases requiring human review. It achieves state-of-the-art accuracy on standard benchmarks while improving calibration, though the lack of shared code limits reproducibility.

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

ReadSaveConnections

When LLMs are asked to simulate mental health patient populations across 120 demographic groups, they produce individuals that look clinically plausible but compress the real-world distribution of symptom severity — erasing the extreme cases that matter most clinically — and 37% of simulated cases flip diagnostic status between two identical prompting runs. This is a direct threat to any research that uses LLM-generated synthetic patients for training classifiers or testing clinical protocols. The severity of variance compression varied by model (14% to 62%), meaning model choice in simulation studies has large, underappreciated consequences.

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

ReadSaveConnections

Two vision-language models applied to automated depression detection showed dramatically different accuracy (34% vs. 80%) across naturalistic versus laboratory settings, while also exhibiting systematic bias by gender and race that differed by model — one was more racially biased, the other more gender-biased. This highlights that before deploying AI depression screening, fairness and context-shift problems need to be solved simultaneously, not separately. The proposed explainability-based fairness interventions provide a concrete direction, but results are currently limited to zero-shot settings on small clinical datasets.

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

ReadSaveConnections

A randomized controlled trial found that a four-week online Acceptance and Commitment Therapy (ACT) program outperformed a placebo waitlist condition on insomnia severity, insomnia-related worry, psychological flexibility, and depression in college students — with gains maintained at one-month follow-up. This matters because college populations have high insomnia prevalence and poor access to in-person therapy, and demonstrating that a brief digital ACT program addresses both sleep and depression simultaneously could support scalable deployment. The three-timepoint RCT design is considerably stronger evidence than the observational studies dominating today's pipeline.

██████████ 0.8 sleep-circadian-psychiatry Peer-reviewed

Read

Electroacupuncture after stroke-like brain injury in rats reduced neuroinflammatory markers (NF-κB, NLRP3, caspase-1, IL-18), shrank infarct volume, and improved learning and memory performance. The mechanism identified — suppressing the NLRP3 inflammasome pathway — is directly relevant to mental health because this same pathway is implicated in treatment-resistant depression and cognitive symptoms in psychiatric disorders. While this is a rodent model and clinical translation is uncertain, it provides mechanistic support for investigating NLRP3 inhibition as an anti-inflammatory strategy in neuropsychiatric conditions.

██████████ 0.8 neuroinflammation Peer-reviewed

Read

PsyGAT structures a clinical interview as a graph where each spoken utterance is a node encoded with clinically meaningful psychological features, and the connections between nodes represent causal transitions between psychological states informed by Big Five personality traits. It achieves 89.99 Macro F1 on the DAIC-WoZ benchmark, surpassing both specialized graph models and large closed-source LLMs. The interpretability module — which generates causal graphs of psychological state transitions — is a meaningful step toward models that clinicians could actually interrogate, though the small depressed sample size (~42 cases) limits confidence in reported performance.

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

ReadSaveConnections

In 222 university students, using AI for functional tasks (writing, coding, research) was associated with larger gray matter volume in the dorsolateral prefrontal cortex and better hippocampal network organization, as well as higher GPA — while socio-emotional AI use showed distinct and less positive neural and wellbeing associations. This is an early empirical attempt to characterize how different AI usage patterns map onto brain structure and mental health in young people, a question with rapidly growing public health relevance. The cross-sectional design means causality cannot be established — brain differences could precede AI use patterns rather than result from them.

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

ReadSaveConnections