291

Papers

10/10

Roadblocks Active

0

Connections

A multi-agent AI system analyzed passive wearable data from over 9,000 participants and identified 41 candidate depression biomarkers, with sleep duration variability and sleep onset variability replicating across two independent cohorts. This matters because most digital biomarker findings fail to replicate even once, so cross-cohort consistency elevates these circadian disruption signals above typical noise. The system also used adversarial validation and human expert review, making its methodology more rigorous than typical automated discovery pipelines.

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

ReadSaveConnections

In rats with co-occurring chronic pain and depression, electroacupuncture reduced brain inflammation and neuronal damage in the anterior cingulate cortex by dampening microglial activation through a specific anti-inflammatory signaling cascade (α7nAChR-TLR4/MyD88/NF-κB). This matters because pain-depression comorbidity affects tens of millions of people and responds poorly to standard antidepressants, so identifying a modifiable neuroinflammatory mechanism opens a distinct therapeutic angle. The anterior cingulate cortex is a convergence zone for both pain and mood processing, making it a plausible target for intervention.

██████████ 0.9 neuroinflammation Peer-reviewed

Read

EviDep adds principled uncertainty quantification to multimodal depression severity estimation: rather than just predicting a score, it tells you how confident that prediction is and whether uncertainty comes from noisy data or from the model itself. It combines wavelet-based signal processing with a Bayesian framework across voice, text, and video inputs, achieving state-of-the-art accuracy on four public benchmarks. Knowing when a model is uncertain — not just what it predicts — is a prerequisite for safe clinical deployment of AI depression assessment tools.

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

ReadSaveConnections

By letting structural MRI and resting-state functional MRI inform each other bidirectionally through dual cross-attention, this model reached 84.71% accuracy for classifying major depressive disorder on a large multi-site dataset — outperforming approaches that treat each imaging type separately. The key insight is that brain structure shapes function, and function in turn reflects structural integrity, so capturing that interplay extracts more signal than either modality alone. There is currently no biological test for MDD, so improvements in imaging-based classification directly advance the case for objective diagnosis.

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

ReadSaveConnections

Tracking nearly 12,000 children over four years, this study found that impulsivity traits, poor parental monitoring, and sleep disturbance were the strongest modifiable predictors of early substance use, while nicotine-related polygenic risk scores provided additional signal that environmental measures cannot explain. Causal modeling methods help distinguish predictors that are plausible intervention targets from those that are merely correlated. Early substance initiation is one of the clearest pathways to adult mental health disorders, making these findings directly actionable for prevention programs.

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

ReadSaveConnections

When four leading AI models were asked to simulate psychiatric patients, they compressed symptom variation by 14–62% compared to real population data, systematically erasing the clinical extremes that matter most for diagnosis and treatment planning. Worse, 37% of simulated patients crossed diagnostic thresholds between two identical test runs, revealing unreliability that would be dangerous in clinical use. This matters because AI-generated patient simulations are increasingly used to develop and test mental health tools — if the simulations are biased, the tools trained on them will be too.

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

ReadSaveConnections

EEG recorded during a cognitive task (stop signal) classifies ADHD more accurately than resting-state EEG, with gamma-band activity over frontal brain regions being the key distinguishing signal — suggesting that the brain under task demand reveals disorder-related differences that are invisible at rest. Separately, white matter tract integrity in motor-planning circuits predicted individual differences in effort-versus-reward sensitivity, a core dimension of motivation relevant to depression as well as ADHD. The task-versus-rest finding has practical implications: diagnostic biomarker studies may need to move away from passive resting protocols.

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

ReadSaveConnections

The DRIFT framework addresses a core problem in precision psychiatry: predicting which antidepressant will work for a specific patient when symptom measures are high-dimensional and partially incomplete. Applied to the EMBARC antidepressant trial, the method's treatment effect predictions generalized to outcomes not used during training — including side effects and self-reported symptoms — which is a meaningful test of real-world utility. Given that roughly half of patients with depression fail their first medication trial, tools that improve treatment matching could have substantial population-level impact.

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

ReadSaveConnections

This system uses a large language model to progressively summarize voice, text, and video data into increasingly detailed clinical reports, then feeds those reports back into a depression severity classifier — mimicking how a clinician builds a picture over successive assessments. The three-stage pipeline (binary screen → severity class → continuous score) outperformed existing baselines on standard benchmarks. Generating human-readable clinical summaries alongside predictions addresses a key adoption barrier: clinicians need to understand why a system flagged a patient, not just that it did.

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

ReadSaveConnections

Testing nine AI language models on Reddit posts, the study found that the 27-billion-parameter Gemma3 model achieved 75% accuracy in classifying depression risk without any task-specific training — competitive with fine-tuned models. Applied to nearly 470,000 real Reddit comments, it identified temporally stable risk profiles across mental health communities. This demonstrates that off-the-shelf large language models can support scalable, low-cost passive surveillance of depression-related distress in online spaces, though the lack of clinical ground-truth labels limits confidence in the in-the-wild findings.

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

ReadSaveConnections