283

Papers

11/11

Roadblocks Active

0

Connections

A multi-agent AI system autonomously generated and statistically validated 41 candidate digital biomarkers for depression from wearable data spanning nearly 10,000 participants across three cohorts, identifying sleep duration and onset variability as the most consistent signals. The system chains hypothesis generation, statistical analysis, adversarial self-criticism, and literature grounding into a single pipeline that mimics a data science team. This matters because it demonstrates that AI can compress months of exploratory biomarker research into a single automated run — though reproducibility is limited by proprietary model infrastructure and one unpublished dataset.

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

ReadSaveConnections

A study of 830 Chinese therapeutic writing samples found that high-level narrative structure — how a person organizes their story, maintains coherence, and builds to a resolution — predicts depression, anxiety, and PTSD severity far better than word-choice patterns or sentence-level meaning. This challenges the dominant approach of using word frequency tools like LIWC for mental health text analysis, suggesting clinicians and researchers are measuring the wrong thing. The implication is that future digital screening tools should evaluate story arc and narrative logic, not just emotional vocabulary.

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

ReadSaveConnections

A three-stage AI pipeline uses a large language model to generate progressively detailed clinical summaries that guide the fusion of text, audio, and video signals for depression severity assessment, achieving state-of-the-art performance on two benchmark datasets. The key innovation is that each stage produces a human-readable clinical rationale, making the model's reasoning inspectable rather than opaque. However, the core step depends on proprietary GPT-o3, which means the system cannot be fully reproduced or deployed independently without API access and significant cost.

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

ReadSaveConnections

Two leading vision-language AI models were tested on depression detection tasks and found to exhibit systematic racial and gender biases — one tended to over-predict depression across all groups, while the other showed disparate accuracy across demographic subgroups, with performance differing dramatically between laboratory and naturalistic settings. The study introduces fairness-aware prompting and a counterfactual loss function as partial remedies. This is an important caution for anyone deploying AI depression screening tools: high average accuracy can mask serious inequity in who gets screened accurately.

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

ReadSaveConnections

A multi-agent AI pipeline structured as a decision tree — with specialized nodes for risk assessment and legal/clinical escalation — reduced false positive rates in self-harm screening by 40% compared to a single-model baseline while maintaining similar sensitivity. The system provides statistical guarantees that errors grow only logarithmically as deployment scale increases. Reducing false positives matters enormously in crisis screening because unnecessary interventions erode trust and overwhelm services, though the evaluation datasets are small (161 and 250 samples) and independent validation is needed.

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

ReadSaveConnections

Across 28,800 synthetic patient profiles generated by four frontier AI models, every model produced clinically plausible-sounding individuals but severely distorted population-level patterns — compressing the spread of symptom severity by 14–62% and making severe cases nearly invisible. Worse, 37% of simulated cases crossed diagnostic thresholds between repeated runs with identical inputs, revealing fundamental instability. This means AI-generated patient data, increasingly used to train clinical systems and run therapy simulations, may embed systematic blind spots toward the most severely ill patients.

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

ReadSaveConnections

A PhD thesis using EEG and brain imaging found that measuring brain activity during an active cognitive task (stop-signal) outperforms resting-state brain scans for classifying adult ADHD, with gamma-band activity over frontal regions as the strongest signal. A separate analysis linked white matter tract integrity — particularly tracts connected to the supplementary motor area — to how individuals weigh effort against reward in decision-making. These findings matter because they suggest that passively recorded resting brain states may be too variable to reliably identify ADHD, and that motivation deficits in ADHD have a specific structural brain correlate.

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

ReadSaveConnections

Tracking nearly 12,000 children over four years, this study found that impulsivity, poor parental monitoring, and sleep disturbance were the strongest modifiable predictors of earlier alcohol, nicotine, and cannabis initiation, while genetic risk scores showed the largest effect for nicotine specifically. Crucially, the causal modeling framework separates factors that merely correlate with early use from those that appear to drive it, pointing toward parental monitoring as a leverage point for intervention. This is one of the more methodologically rigorous youth substance use studies to date, using a large publicly accessible dataset with time-varying risk factor modeling.

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

ReadSaveConnections

Among 222 university students, those who used AI tools primarily for functional tasks (writing, research, problem-solving) showed larger volumes in brain regions linked to working memory and visual processing, and better-organized hippocampal networks, compared to those who used AI primarily for emotional companionship. Functional AI use also correlated with higher grades, while socio-emotional AI use showed a divergent neural and academic profile. This is an early but concrete signal that how — not just how much — young people use AI tools may have measurable consequences for brain structure, warranting longitudinal follow-up.

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

ReadSaveConnections

In a rat model of stroke-induced brain injury, electroacupuncture reduced brain infarct volume and restored memory performance on a maze task by suppressing the NLRP3 inflammasome pathway — a molecular cascade that triggers destructive inflammation in neural tissue. The same pathway is increasingly implicated in depression and treatment-resistant mood disorders, making these mechanistic findings relevant beyond stroke neurology. While an animal study cannot directly translate to psychiatric practice, identifying electroacupuncture's anti-inflammatory mechanism provides a testable hypothesis for its reported effects on depressive symptoms.

██████████ 0.7 neuroinflammation Peer-reviewed

Read