282

Papers

10/10

Roadblocks Active

0

Connections

A multi-agent AI system autonomously analyzed wearable sensor data from over 9,000 participants and identified 41 candidate digital biomarkers for depression, with sleep timing irregularity surviving validation in two independent datasets. The key finding — that how much sleep duration varies night-to-night correlates with depression severity — points to circadian rhythm disruption as a measurable, passive signal rather than something requiring clinical interviews. This is important because it opens a path to low-cost, continuous depression monitoring from consumer devices, though the reliance on a proprietary unpublished dataset limits full independent verification.

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

ReadSaveConnections

This scoping review systematically maps the mental health consequences of violent conflict exposure in adolescents, finding that violence triggers a wide spectrum of conditions — from anxiety and PTSD to psychotic disorders and substance abuse — affecting both internal emotional states and outward behavior. The breadth of impact is significant because it challenges single-disorder intervention models: adolescents exposed to violence rarely present with just one diagnosable condition, suggesting that treatment programs targeting only PTSD or only depression will miss a large share of pathology. For researchers and policymakers, this establishes a clearer evidence base for why multi-component, trauma-informed mental health programs in conflict-affected youth populations are necessary.

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

Read

This paper proposes a deep learning architecture that simultaneously processes two types of brain imaging — structural MRI (brain anatomy) and functional MRI (brain activity patterns) — and learns how they interact bidirectionally rather than simply combining them. The method outperforms standard approaches that just concatenate features, suggesting that the relationship between brain structure and function carries diagnostic information that neither modality alone captures. For depression research, this matters because it moves toward imaging-based biomarkers that reflect the actual biology of the disorder rather than proxies, though the paper's generalizability beyond the specific datasets used remains to be tested.

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

ReadSaveConnections

When leading AI language models are asked to simulate psychiatric patients for training or research purposes, they produce individuals who sound clinically plausible but whose collective statistics are badly wrong — variance in symptom severity is compressed by 14–62% compared to real populations, and 37% of simulated patients flip their diagnosis on a second run. This is a critical finding for anyone developing AI-powered therapy tools or using synthetic patients to test clinical algorithms, because training on these simulations would teach systems to expect a more homogeneous, less severe range of mental illness than actually exists. The work establishes a concrete benchmark (PsychBench) for measuring this distortion.

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

ReadSaveConnections

In a sample of 222 university students, those who used AI tools for task-oriented purposes (writing, studying) showed larger gray matter volume in prefrontal and visual cortex regions and better academic performance, while socio-emotional AI use (for companionship or emotional support) showed a different profile. The structural brain differences are correlational and cannot establish causation — heavier cognitive engagement with AI may reflect or reinforce pre-existing neural patterns rather than cause them. This study is early but relevant to the youth mental health debate because it suggests that how young people use AI, not merely how much, may shape psychological and neurological outcomes differently.

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

ReadSaveConnections

This paper proposes a mechanism by which nighttime blue light exposure disrupts the brain's waste-clearance system (the glymphatic system) by misaligning the water channels (AQP4) in support cells, preventing the brain from flushing out proteins like beta-amyloid and tau during sleep. The authors suggest that 40 Hz light or sound stimulation could counteract this by re-entraining the neural circuits that control this clearance process. The hypothesis is speculative and not empirically tested in this paper, but it connects circadian disruption to neurotoxic protein accumulation in a way that is directly relevant to both sleep psychiatry and longer-term neurodegeneration risk.

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

Read

TAAC addresses a practical barrier to deploying voice-based depression screening: people are reluctant to submit recordings if it means giving up their identity. The system separates depression-relevant features from speaker identity features in audio using an adversarial training approach, allowing diagnosis without preserving a recognizable voice print. If validated robustly, this privacy-preserving architecture could enable passive, continuous depression monitoring in real-world settings where privacy is a dealbreaker — though key performance details are not fully visible in the available preprint text.

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

ReadSaveConnections

This paper reveals a serious methodological flaw in automated depression detection research: AI models trained on clinical interview transcripts can achieve high accuracy by learning the interviewer's fixed script rather than anything the patient actually says, since the interviewer's questions occur at predictable points and differ subtly by interview type. This means many published depression detection results may be inflated artifacts of interview structure rather than genuine signals of depression. The finding applies across three independent datasets and two different model architectures, making it a broadly relevant warning for the field.

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

ReadSaveConnections

DRIFT is a statistical method for estimating which depressed patients are most likely to benefit from a specific treatment (sertraline vs. placebo), applied here to the EMBARC randomized controlled trial. Unlike prior approaches that optimize for a single outcome measure, DRIFT is designed to be robust across multiple symptom domains simultaneously, including outcomes not used during training — addressing the common problem that a treatment strategy optimized for one scale performs poorly on others. The application to real RCT data for major depressive disorder makes this one of the more clinically grounded precision psychiatry papers in today's set.

██████████ 0.8 treatment-resistant-depression Preprint

ReadSaveConnections

This system uses a large language model to generate progressively richer clinical summaries — first flagging likely depression, then estimating severity, then identifying probable causes — and uses those summaries to guide how video, audio, and text features are fused together for final prediction. The approach outperforms baseline models on two benchmark datasets and, importantly, produces human-readable rationales alongside its predictions, addressing a key barrier to clinical adoption of automated tools. The reliance on a proprietary LLM (GPT-o3) for the summary generation step limits reproducibility and long-term stability of the pipeline.

██████████ 0.7 depression-biomarkers Preprint

ReadSaveConnections