290

Papers

10/10

Roadblocks Active

0

Connections

In 222 university students, using AI tools for practical tasks (writing, coding) was linked to larger prefrontal and visual cortex gray matter volumes and higher grades, while socio-emotional AI use showed no such benefits and trended toward worse mental health outcomes. This matters because it suggests that how people use AI — not just how much — may differentially shape brain development and psychological wellbeing in young adults. The cross-sectional design limits causal claims, but the neural specificity of the finding adds weight beyond self-report studies.

██████████ 0.9 youth-mental-health-crisis Preprint

ReadSaveConnections

When four major LLMs were asked to simulate 28,800 psychiatric patient profiles, they produced individuals that looked clinically plausible but populations that were statistically compressed — eliminating the extreme cases (severe, treatment-resistant) that matter most in mental health research. Variance in symptom severity was shrunk by 14–62% depending on the model, and 37% of cases crossed a diagnostic threshold between two identical prompting runs. This is a direct warning against using LLM-generated data to train or evaluate digital mental health tools without population-level validation.

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

ReadSaveConnections

This model fuses structural brain scans (gray matter shape) with functional MRI connectivity patterns using a bidirectional attention mechanism, achieving 84.71% accuracy for classifying major depressive disorder — outperforming simpler feature-concatenation approaches. The key advance is explicitly modeling how structure and function interact rather than treating them as independent signals, which is biologically more accurate. The study uses a public multi-site dataset with 10-fold cross-validation, making it more credible than single-site work, though no code is released.

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

ReadSaveConnections

A major practical problem in clinical neuroimaging AI is that models degrade when new hospital sites are added (different scanners, protocols). FORGE solves this by generating synthetic brain connectivity matrices from old sites to rehearse alongside new data, rather than storing patient records. Tested across autism, depression, and schizophrenia datasets, this approach preserves prior performance while learning new sites — a step toward deployable, multi-institution brain disorder classifiers. Code is publicly available.

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

ReadSaveConnections

Task-based EEG during a stop-signal test outperformed resting-state EEG for classifying adult ADHD, with gamma-band power over frontal and parietal regions as the strongest discriminating features — suggesting that the brain's response to cognitive demand is more informative than its baseline state. A separate analysis linked white matter tract integrity to computational estimates of how much effort individuals are willing to exert for reward, connecting brain structure to a core motivational deficit in ADHD and depression. This dual approach (classification plus mechanistic modeling) is more informative than purely diagnostic work.

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

ReadSaveConnections

In nearly 12,000 children followed from age 10 across four years, impulsivity, low parental monitoring, and nicotine-related genetic risk were the most robust predictors of early substance use — surviving adjustment for confounders in causal models, unlike many weaker environmental associations. The use of marginal structural models (a causal inference method) moves beyond correlation to estimate what would happen if modifiable factors like parental monitoring were changed. This is directly actionable for prevention programs targeting the pre-adolescent window.

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

ReadSaveConnections

This system uses an LLM to generate a running clinical narrative at each stage of a three-step pipeline — screening, severity grading, and continuous scoring — which then guides how audio, video, and text signals are fused to estimate depression. The design addresses a key weakness of black-box multimodal models: the LLM summaries provide a human-readable rationale alongside each prediction. Performance improvements over prior state-of-the-art are reported on two interview datasets, though reliance on the proprietary GPT-o3 API limits independent replication.

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

ReadSaveConnections

EviDep estimates not just depression severity from audio-visual data but also how confident the model is in each prediction — using a statistical framework (Normal-Inverse-Gamma distribution) that distinguishes uncertainty from genuine ambiguity versus data noise. The model also actively separates overlapping information across modalities, reducing redundancy that typically degrades fusion systems. Tested on four public benchmarks, it claims both accuracy and calibration improvements, though full implementation details are not yet public.

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

ReadSaveConnections

DRIFT addresses a real clinical problem: antidepressant trials measure dozens of symptoms across mood, sleep, and cognition, but standard analysis collapses these into a single summary score, discarding information about which domains actually improve for which patient. The method uses factor analysis to extract latent symptom dimensions and then finds the treatment assignment rule that performs best in the worst-case weighting of those dimensions — producing robust individualized treatment recommendations. Applied to the EMBARC sertraline trial, this could help identify which depressed patients genuinely benefit from SSRIs versus placebo.

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

ReadSaveConnections

A 27-billion-parameter open-source LLM (Gemma3) applied zero-shot to Reddit posts classified eight depression-linked emotions with macro-F1 of 0.70, closing most of the gap to fine-tuned supervised models — suggesting large open models may be usable for passive mental health monitoring without labeled training data. Applied to nearly 470,000 Reddit comments across mental health communities from 2024–2025, the model found stable depression risk profiles over time. The in-the-wild data has no ground-truth labels, limiting interpretation, but the methodology is transparent and the benchmark dataset is public.

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

ReadSaveConnections