276

Papers

10/10

Roadblocks Active

0

Connections

This study models the structure of how a person's voice state 'revisits' similar patterns during conversation — a property called recurrence — to detect depression. Using this approach on the DAIC-WOZ interview corpus, it achieved an AUC of 0.689, beating five alternative acoustic feature families including static descriptors and entropy measures. The result matters because it suggests depression distorts the rhythmic, self-similar structure of speech dynamics, offering a biologically interpretable signal beyond average vocal quality.

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

ReadSaveConnections

Rather than averaging acoustic features across a conversation, this paper tracks how unpredictable (entropic) vocal patterns are moment-to-moment, finding Shannon entropy of vocal trajectories outperforms recurrence, fractal, and coupling biomarkers on the same DAIC-WOZ dataset (AUC 0.646 vs. static baseline 0.593). The key finding is that depression-related signal lives in the variability of how voice changes over time, not in its average level. Paired with the recurrence paper above, this day produced two complementary, convergent results on temporal vocal dynamics — though neither has been validated beyond a single small corpus.

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

ReadSaveConnections

This review synthesizes evidence that obstructive sleep apnea (OSA) structurally and functionally damages the hippocampus — the brain region central to memory and emotional regulation — and impairs neuroplasticity. This matters for psychiatry because OSA is highly comorbid with depression and anxiety but frequently goes undiagnosed in mental health settings. Treating OSA could be an underutilized lever for improving psychiatric outcomes, particularly in treatment-resistant populations.

██████████ 0.9 sleep-circadian-psychiatry Peer-reviewed

Read

This population-level study maps how different mental health symptoms in adolescents drive each other forward over time, using a cross-lagged network approach that identifies which symptoms are likely causes versus consequences. This is more actionable than simple correlation studies because it suggests which symptoms to intervene on first to disrupt cascade effects. For the youth mental health crisis, understanding symptom sequencing is critical for timing and targeting early interventions.

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

Read

Deep learning models trained directly on raw speech signals — ignoring what is said and focusing only on how it sounds — detected depression and anxiety at 71% simultaneous sensitivity and specificity across roughly 5,000 unique U.S. participants. Combining these acoustic biomarkers with lexical features improved performance further, and the best model weights have been publicly released on HuggingFace. The scale of evaluation (34,000+ training subjects) and the public release make this a practically accessible tool, though the proprietary training data prevents independent audit of demographic fairness.

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

ReadSaveConnections

ADAPTS uses a team of LLM agents to automatically score depression and anxiety severity from clinical interview transcripts by breaking the task into symptom-by-symptom reasoning subtasks. On interviews where human raters disagreed most (high-discrepancy cases), the automated system's ratings were actually closer to expert benchmarks than the original human ratings were. This matters because inter-rater unreliability in clinical severity scoring is a persistent problem in both research and care, and automation that handles ambiguous cases well could improve consistency at scale.

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

ReadSaveConnections

Two vision-language models (VLMs) evaluated for depression detection from audio, video, and text show wildly inconsistent performance depending on dataset context — one model achieved 80.4% accuracy in one setting, while the other managed only 33.9% — and both exhibited measurable racial and gender biases. The paper proposes explainability-guided fairness interventions (chain-of-thought prompting and counterfactual loss) to partially correct these disparities. This is a concrete warning that multimodal AI tools for mental health screening are not deployment-ready without systematic fairness auditing.

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

ReadSaveConnections

A 27-billion-parameter open LLM (Gemma3:27b) running without any task-specific fine-tuning nearly matched fine-tuned models for classifying depression-related emotions in Reddit posts (micro-F1 0.75 vs. 0.80 for fine-tuned BART). The authors combined these emotion classifications into a weighted severity index applicable to population-level social media analysis. This matters because it lowers the barrier to passive mental health surveillance: capable zero-shot models now exist that don't require expensive labeled training datasets.

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

ReadSaveConnections

When LLMs simulate psychiatric patients, they generate individuals who sound clinically plausible but whose population-level distributions are systematically wrong: symptom variance is compressed by 14–62% depending on the model, and 37% of simulated cases flip diagnostic threshold between repeated runs of the same prompt. This matters because LLM-simulated patients are increasingly used in clinical training and research, and this hidden failure mode — individuals seem realistic but the population is a distortion — could silently bias everything built on top of such simulations.

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

ReadSaveConnections

Clustering 551 survey respondents by social media use and psychological wellbeing revealed six distinct user profiles rather than a single uniform relationship between screen time and mental health. The overall correlation between social media hours and anxiety was modest (r=0.28), suggesting that average effects mask important subgroup variation. For youth mental health policy, this supports moving away from blanket screen-time limits toward more targeted interventions matched to specific usage patterns and risk profiles.

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

ReadSaveConnections