274

Papers

9/9

Roadblocks Active

0

Connections

A case-control study found that a variant in the DRD2 dopamine receptor gene (A2 allele) amplifies how family trauma and school bullying translate into suicidal behavior in adolescents. Depression acts as a mediating pathway between these environmental stressors and suicidal outcomes, not just a co-occurring condition. This matters because it identifies a gene-environment interaction that could guide targeted screening — adolescents carrying this allele in high-stress environments may warrant earlier mental health intervention.

██████████ 0.9 youth-mental-health-crisis Peer-reviewed

Read

This paper trains an unsupervised generative model exclusively on healthy brain EEG signals to learn what 'normal' looks like, then scores patient data by how far it deviates — bypassing the chronic shortage of labeled depression data. A separate module handles mismatched electrode configurations across clinical sites, which is a real-world barrier to multi-center deployment. The approach is architecturally clever, but reproducibility is poor due to absent code and incomplete hyperparameter reporting.

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

ReadSaveConnections

A pilot study combined two brain-scanning techniques — EEG (electrical activity) and fNIRS (blood oxygenation) — in a deep learning pipeline to classify depressive states, and demonstrated it can flag depression in individuals who haven't self-identified as depressed. This 'passive' detection capacity is clinically relevant for early intervention before symptoms escalate. The sample of 11 healthy university students is far too small for clinical conclusions, and the absence of a diagnosed depression cohort is a significant methodological gap.

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

ReadSaveConnections

A large language model fine-tuned on AI therapy chatbot transcripts predicts standardized depression scores (PHQ-9) with an AUC of 0.91 for clinically significant depression, meaning it can passively monitor users without requiring them to fill in questionnaires. The training pipeline expands a labeled dataset of 3,111 users to over 6,000 using AI-generated pseudo-labels, which is a practical scalability strategy. The core limitation is that the data comes from a single commercial platform and cannot be shared, making independent validation currently impossible.

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

ReadSaveConnections

InfoShield strips demographic signals (gender, age) from voice-based depression screening while preserving diagnostic accuracy — reducing gender inference from 92.6% to 55.5% with only a minor drop in depression detection F1. Standard privacy methods failed on speech because voice features evolve over time differently from static demographic labels, so the authors built a new cross-modal attention mechanism to align them. This is directly relevant for deploying voice-based mental health tools under strict privacy regulations like GDPR.

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

ReadSaveConnections

This study applies five different explainability techniques to an EEG-based depression classifier to check whether the model learns clinically plausible brain patterns or spurious correlations. Multiple methods converge on frontal, temporal, and right-hemisphere regions as most informative — consistent with established neuroscience of mood disorders — providing partial validation that the model is not simply overfitting to noise. Unexplainable AI predictions are a key barrier to clinical adoption, so this convergence evidence, even if preliminary, is meaningful for the field.

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

ReadSaveConnections

A mixed-methods study combining focus groups and a 435-person cross-national survey identified four distinct user profiles among mental health patients using digital self-monitoring apps, differing in technology comfort and motivation to track symptoms. The top barriers to sustained use were excessive measurement frequency and insufficient clinician involvement — not technology literacy. These findings give developers and clinicians concrete, actionable targets: reducing burden and integrating feedback loops are higher priorities than improving interfaces.

██████████ 0.8 digital-therapeutics Peer-reviewed

Read

A small RCT of 13 veterans found that combining a wearable-based digital intervention with endurance cycling produced more stable PTSD and anxiety symptom trajectories after the event than cycling alone, where gains partly reversed. The digital component used a machine learning model on continuous smartwatch data to detect hyperarousal and prompt self-management in real time. With only 7 participants in the intervention arm, no firm efficacy conclusions are possible, but the design and trajectory modeling approach are worth scaling.

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

ReadSaveConnections

Using multifractal analysis of sleep EEG from 290 older women, this study shows that cognitively healthy individuals display brain signal patterns closer to a theoretically optimal 'critical' state during non-REM sleep than those who later develop dementia-related impairment. The differences are statistically significant across all electrode locations (p ≤ 0.001) and most pronounced in deep sleep stages N2 and N3. This positions routine sleep EEG — already collected in sleep clinics — as a passive, non-invasive early warning signal for cognitive decline, which has direct implications for psychiatric monitoring in aging populations.

██████████ 0.8 sleep-circadian-psychiatry Preprint

ReadSaveConnections

This study extracts a single complexity measure (Hurst exponent via Detrended Fluctuation Analysis) from four EEG channels to automatically identify deep sleep stages in 290 older women, achieving 87% balanced accuracy with a simple Naive Bayes classifier. Complex linear models performed far worse (SVM: 51%), confirming the feature space is genuinely non-linear — a finding the authors visualize directly using manifold learning. Accurate automated deep sleep detection without manual scoring is a prerequisite for neurofeedback systems that aim to enhance sleep quality as a psychiatric intervention.

██████████ 0.8 sleep-circadian-psychiatry Preprint

ReadSaveConnections