280

Papers

12/12

Roadblocks Active

0

Connections

This paper treats the voice as a dynamical system — tracking how vocal patterns revisit the same acoustic states over time — and uses these 'recurrence' signatures to detect depression from conversational speech. The approach outperforms standard acoustic feature baselines (AUC 0.689), suggesting that the irregular, collapsed vocal dynamics of depression are better captured by nonlinear mathematics than by simple statistics like pitch or energy. If validated further, this could enable passive, non-intrusive depression screening from everyday phone calls or therapy sessions.

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

ReadSaveConnections

EviDep is an AI model that estimates depression severity from audio and video while also reporting how confident it is in each prediction — distinguishing uncertainty that comes from noisy data versus uncertainty about the model itself. By separating shared signals from audio and video from what each modality contributes uniquely, it achieves both better accuracy and better-calibrated confidence than prior methods across four benchmarks. In clinical practice, knowing when a model is uncertain is at least as important as its accuracy — this paper takes that requirement seriously.

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

ReadSaveConnections

This purely theoretical paper argues that some high-cognitive-potential and neurodivergent individuals naturally operate with brains that resemble the psychedelic state — with reduced top-down filtering and heightened sensitivity to raw sensory input — without any drugs involved. It extends the REBUS model (a leading theory of how psychedelics work) to propose that depression in these individuals represents a different mechanism than typical depression: a collapse of high-entropy neural processing rather than a deficit of arousal or serotonin. The paper proposes no data yet, but its implications for why standard antidepressants may fail in this population — and why psychedelic-assisted therapy might be especially relevant — are worth tracking as empirical work follows.

██████████ 0.9 psychedelic-mechanisms Peer-reviewed

Read

This paper combines structural brain scans (what the brain looks like) and functional brain scans (how brain regions talk to each other) using a bidirectional attention mechanism that lets each modality inform interpretation of the other, rather than simply stitching them together. Tested on a large multi-site dataset (REST-meta-MDD), this approach achieves 84.71% accuracy for detecting major depressive disorder — outperforming fusion methods that ignore cross-modal interactions. Multi-site performance matters because it is a prerequisite for any MRI biomarker to be usable across different hospitals.

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

ReadSaveConnections

PsyGAT models a clinical interview as a graph where nodes represent psychologically meaningful speech units (linked to DSM criteria) and edges track how a patient's emotional state evolves across the conversation — rather than treating the interview as a bag of features. It achieves 89.99 Macro F1 on the DAIC-WoZ benchmark, outperforming black-box deep learning and closed-source LLMs, while also being able to identify which utterances triggered which depressive symptoms. Interpretability of this kind is essential for clinician trust, which is the main barrier to AI adoption in psychiatric practice.

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

ReadSaveConnections

When predicting mental health outcomes from therapeutic writing, analyzing the high-level structure of a person's narrative — how they organize their story, whether it has coherent causality and resolution — outperforms counting words or measuring semantic similarity. This finding holds across 830 samples from six diverse Chinese intervention studies covering depression, anxiety, and PTSD, and suggests that LLMs capable of evaluating narrative structure add something that word-frequency tools fundamentally cannot capture. For digital mental health tools that use expressive writing, this implies that feedback should address story structure, not just emotional vocabulary.

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

ReadSaveConnections

When four major AI models (GPT-4o-mini, DeepSeek-V3, Gemini, GLM-4) are asked to simulate psychiatric patients across 28,800 synthetic profiles, they produce individual cases that look clinically plausible but systematically compress the realistic range of symptom severity — eliminating the extreme cases that matter most in clinical care, with variance suppressed by 14–62% depending on the model. Additionally, 37% of simulated cases flip their diagnostic classification between two identical runs, meaning these AI patient simulators are both systematically biased and unreliable. This is a direct warning for researchers using AI-generated patients to train or test clinical decision tools.

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

ReadSaveConnections

Two state-of-the-art vision-language models (Phi-3.5-Vision and Qwen2-VL) tested for depression detection showed not only dramatic performance swings across datasets (80% vs. 34% accuracy on different benchmarks) but also systematic racial and gender biases, with one model over-diagnosing depression in laboratory settings. The paper introduces fairness-aware prompting strategies that reduce false positives by roughly 40%, providing a concrete intervention to reduce algorithmic harm. Any AI tool intended for clinical depression screening must pass this kind of demographic parity test before deployment — this paper sets an early benchmark for what that evaluation should look like.

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

ReadSaveConnections

In 222 university students, using AI for task-oriented purposes (writing, studying) was associated with larger prefrontal and hippocampal brain regions and better academic performance, while using AI primarily for social-emotional needs showed a different neural and mental health profile. The brain differences were detectable via MRI, suggesting that how young people integrate AI into their lives — not just how much they use it — shapes neural development and psychological outcomes. This is one of the first neuroimaging studies of AI usage patterns and carries direct implications for how educational and mental health policies around AI use should be designed.

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

ReadSaveConnections

An online Acceptance and Commitment Therapy (ACT) program for college students with insomnia outperformed a placebo condition on insomnia severity, psychological flexibility, sleep-related worry, and depressive symptoms in a randomized controlled trial — making this one of the few rigorous tests of a fully digital sleep intervention in a high-risk population. College students have substantially higher insomnia rates than the general population, and insomnia is itself a major risk factor for depression and academic failure, so an effective scalable intervention has broad reach. The link between sleep improvement and reduced depressive symptoms in the same intervention also provides practical evidence that these two problems can be addressed together.

██████████ 0.8 sleep-circadian-psychiatry Peer-reviewed

Read