280

Papers

10/10

Roadblocks Active

0

Connections

This narrative review proposes that gut microbiota does not directly cause psychiatric or neurological disease but instead raises or lowers the threshold at which the brain becomes vulnerable to breaking down. The distinction matters clinically: it suggests that microbiome interventions may be most useful as risk-reduction tools during sensitive developmental windows rather than as direct treatments. The framework is theoretical and lacks systematic review rigor, so its value is in organizing future longitudinal intervention trials rather than confirming a causal mechanism.

██████████ 0.9 gut-brain-axis Peer-reviewed

Read

This paper combines structural brain scans and resting-state fMRI in a single model that lets each imaging modality inform the other through a bidirectional attention mechanism, achieving 84.7% accuracy in classifying depression on a large multi-site dataset. The key advance over prior work is replacing simple feature stacking with explicit cross-modal interaction, which the authors show outperforms concatenation baselines. Results are promising but the absence of site-harmonization details and demographic breakdowns in the write-up means reproducibility and generalization need independent verification.

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

ReadSaveConnections

A persistent problem in clinical neuroimaging AI is that models trained at one hospital degrade when new sites are added, requiring retraining on all past data—which violates patient privacy. This paper addresses that by generating synthetic brain connectivity matrices that mimic real patient data, allowing the model to rehearse old knowledge without storing actual scans. The approach is tested across three disorders (depression, autism, schizophrenia) with publicly accessible datasets and code released on GitHub, making it one of the more reproducible papers in today's batch.

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

ReadSaveConnections

Most depression detection models output a severity score without any measure of how confident they should be—a significant problem if a clinician needs to know whether to trust the prediction. This paper adds a probabilistic layer that explicitly estimates both the uncertainty from noisy data (aleatoric) and from model limitations (epistemic), tested on four public audio-visual benchmarks including DAIC-WOZ. The method achieves competitive accuracy with better-calibrated uncertainty, which is a meaningful step toward systems clinicians could actually act on, though implementation details and code are not yet public.

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

ReadSaveConnections

Using brain scans from 222 university students, this study finds that using AI tools for task-focused work (writing, studying) is associated with larger prefrontal cortex volume and better grades, while using AI for emotional companionship shows a different and less positive profile. This is one of the first neuroimaging studies to examine how AI use patterns map onto brain structure, though the cross-sectional design means we cannot tell whether AI use shapes the brain or whether people with certain brain profiles gravitate toward certain uses. The findings are relevant to growing concerns about AI companionship apps and youth mental health.

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

ReadSaveConnections

This PhD thesis finds that EEG recorded while adults perform a cognitive task (stopping a response) classifies ADHD more accurately than resting-state EEG, with gamma-band activity over frontal and parietal regions doing most of the work. A separate analysis links white matter tract integrity to how much effort versus reward individuals are willing to trade off, measured through a computational model. The work is notable for directly comparing task-based and resting biomarkers rather than assuming one is sufficient, which has practical implications for which assessment context to use in clinical trials.

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

ReadSaveConnections

Using survival analysis on the large ABCD longitudinal cohort, this study identifies impulsivity, parental monitoring, and sleep disturbance as the most robust environmental predictors of when adolescents first try substances—and finds that genetic risk (polygenic scores) independently accelerates that timeline, especially for nicotine. The time-varying approach captures how risk factors shift across development rather than treating them as fixed, which is methodologically stronger than most prior work. These findings suggest that intervention windows differ by substance and that both behavioral and genetic risk should be considered in screening programs.

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

ReadSaveConnections

Researchers are increasingly using large language models to simulate patient populations for mental health research, but this paper shows those simulations are systematically distorted: LLMs compress the real-world range of symptom severity by 14–62% and one-third of simulated patients cross diagnostic thresholds between two runs of the same prompt. This matters because synthetic patient data is being used to test clinical tools and train other models, and these errors could systematically bias what gets built. The study evaluates four major LLMs against real epidemiological surveys, providing concrete metrics that developers and regulators could adopt as a quality standard.

██████████ 0.7 digital-therapeutics Preprint

ReadSaveConnections

This paper proposes detecting depression from voice recordings while simultaneously stripping out identifying information—separating the signal that indicates mood from the signal that reveals who is speaking. The mechanism uses adversarial training to force the model to disentangle these two types of features, plus a noise-injection layer to encrypt identity further. Privacy-preserving voice analysis could lower barriers to passive depression monitoring via phone apps, though the paper lacks a public code release and the full performance details are in truncated sections not yet available for review.

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

ReadSaveConnections

Rather than outputting just a depression score, this system uses a large language model to generate a step-by-step clinical narrative summarizing what it observed across text, audio, and video before arriving at a prediction. The interpretability angle addresses a real clinical bottleneck: practitioners need to know why a system flagged a patient, not just that it did. This is a design-forward paper rather than a methods-heavy one, and its impact will depend on whether the summaries are validated as actually useful by clinicians in prospective settings.

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

ReadSaveConnections