264

Papers

10/10

Roadblocks Active

1

Connections

This study used untargeted plasma metabolomics in 85 depressed adolescents and 46 healthy controls, then applied three machine learning algorithms (LASSO, random forest, SVM) to identify a three-metabolite consensus signature that classifies adolescent depression with reasonable accuracy. The diagnostic model was validated via ROC-AUC, calibration curves, and a held-out validation set. If the metabolite panel replicates, it could offer a blood-based screen at a life stage where depression is often undiagnosed and where early intervention has the highest payoff.

██████████ 1.0 depression-biomarkers Peer-reviewed

Read

CoDaS is a multi-agent LLM system that autonomously cycles through hypothesis generation, statistical testing, adversarial validation, and literature grounding to discover digital biomarkers from wearable sensor data. Across nearly 10,000 participant-observations, it flagged 41 candidate mental health biomarkers, with sleep timing variability emerging independently in two separate depression cohorts — pointing to circadian instability as a robust signal. The system's reproducibility is constrained by proprietary LLMs and non-public datasets, but the consistency across cohorts makes the circadian findings worth tracking.

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

ReadSaveConnections

This paper builds a three-stage system where a large language model generates progressively detailed clinical summaries from an interview, which then guide a fusion of text, audio, and video features for depression detection. The approach outperforms prior state-of-the-art on two clinical interview datasets and, critically, produces human-readable rationales for its predictions — a step toward explainability that clinicians actually need. Reproducibility is limited by reliance on proprietary GPT-o3 outputs, meaning exact replication is not currently feasible by outside researchers.

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

ReadSaveConnections

This study extracts enhanced entropy measures from EEG signals — quantifying the complexity or irregularity of brain electrical activity — and uses these as features to both detect depression and classify its severity. The approach is attractive because EEG is relatively low-cost and non-invasive compared to neuroimaging. Detailed methodology and dataset information are limited in available text, so independent assessment of effect sizes and generalizability requires access to the full paper.

██████████ 0.9 depression-biomarkers Peer-reviewed

Read

Using EEG from 20 healthy adults, this study models the brain as an energy landscape — where lower energy means the network is locked into a more stable, constrained state — and finds that processing sad faces drives the brain into significantly more stable (lower-energy) configurations than happy faces, especially in the alpha band (Cohen's d = 0.83). This offers a quantitative, physics-inspired framework for understanding why negative emotional states may be 'stickier' than positive ones, with potential relevance to understanding rumination in depression. The small sample limits conclusions but the methodology is clearly specified and the effect sizes are noteworthy.

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

ReadSaveConnections

DRIFT is a statistical framework that estimates which individual patients will respond to a treatment — here applied to sertraline vs. placebo in the EMBARC major depression RCT — even when outcomes are measured across many symptom domains simultaneously. It uses adversarial optimization to find treatment rules that are robust to uncertainty about which outcome dimensions matter most, addressing a key weakness of standard treatment effect methods that collapse symptom complexity into single scores. A closed-form solution and publicly available trial data make this methodology relatively accessible for replication.

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

ReadSaveConnections

This paper tests a state-of-the-art deep learning sleep staging model — trained on healthy subjects — against a new dataset of 100 ischemic stroke patients, finding that performance degrades sharply and that gradient-based visualizations show the model attending to physiologically uninformative EEG regions in patients. The result is a concrete demonstration that AI sleep tools validated on healthy populations cannot be safely assumed to work in clinical groups, which is a significant warning for deployment in psychiatric patients who often have disrupted sleep. The iSLEEPS clinical dataset is not yet public, limiting immediate replication of the key finding.

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

ReadSaveConnections

This theoretical paper proposes a formal definition of plasticity as a ratio of system size to total connectivity strength, arguing that this ratio predicts whether a system — including a brain — can flexibly adapt or is locked into rigid patterns. The framework suggests there is an optimal plasticity window at intermediate connectivity, which may explain why both over-connected (e.g., manic) and under-connected (e.g., severely depressed) brain states are maladaptive. No empirical data are analyzed; the framework's value depends entirely on future experimental validation.

██████████ 0.8 neuroplasticity-interventions Preprint

ReadSaveConnections

This paper demonstrates that depression detection models trained on clinical interview data can achieve high accuracy by learning patterns in the interviewer's scripted prompts — not the patient's actual speech — across three independent datasets. This is a fundamental validity problem: if an AI 'detects' depression by recognizing where in a script it is rather than what the patient says, the tool is useless clinically and misleadingly good on benchmarks. The finding should prompt re-evaluation of reported accuracy figures for many existing depression detection systems.

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

ReadSaveConnections

In 222 university students, this MRI study finds that using AI for functional tasks (writing, coding) correlates with larger prefrontal and hippocampal brain regions and better academic performance, while using AI for social-emotional needs (companionship, emotional support) correlates with smaller amygdala and temporal regions and worse depression and anxiety scores. The cross-sectional design means causality cannot be determined — it is unknown whether emotionally vulnerable students seek AI companionship, or whether such use worsens mental health — but the neuroimaging data add biological depth to the policy conversation about AI companion apps.

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

ReadSaveConnections