288

Papers

10/10

Roadblocks Active

0

Connections

Published in PNAS, this paper proposes a synaptic and circuit plasticity hypothesis to explain why ketamine rapidly reduces depression and suicidality in patients who have failed other treatments. The mechanism of ketamine has remained contested despite its clinical use, and a credible, testable hypothesis published in a high-visibility journal could accelerate drug design for next-generation antidepressants. This is significant because treatment-resistant depression affects roughly 30% of depressed patients and has very few evidence-based options.

██████████ 0.9 treatment-resistant-depression Peer-reviewed

Read

CoDaS is a multi-agent AI pipeline that autonomously mines wearable sensor data to discover candidate biomarkers, identifying 41 depression-related digital signals across two independent cohorts totaling over 8,000 participant-observations. Crucially, circadian instability features — specifically variability in sleep duration and sleep onset timing — replicated across both cohorts, suggesting these are robust signals rather than dataset artifacts. The system's ability to replicate findings across independent datasets is a meaningful step toward trustworthy AI-assisted biomarker discovery, though reproducibility is limited by proprietary datasets.

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

ReadSaveConnections

This paper treats the voice as a nonlinear dynamical system and extracts features based on how vocal state trajectories recur over time — a more sophisticated characterization than traditional pitch or energy statistics. Recurrence-based features achieved a mean cross-validated AUC of 0.689 for depression detection, outperforming entropy dynamics, Hurst exponent, and conventional acoustic baselines. The finding matters because it suggests depression leaves a measurable imprint on the temporal structure of speech, not just its average properties, which opens a path toward passive voice-based screening.

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

ReadSaveConnections

Using Shannon entropy applied to acoustic feature trajectories across an utterance, this paper shows that how unpredictably a voice changes over time is more informative for depression detection than standard pooled summaries of the same features (AUC 0.646 vs. 0.593). The entropy approach reached statistical significance under permutation testing, adding confidence that the signal is not due to chance. Alongside the companion recurrence paper, this constitutes a small but coherent evidence cluster supporting temporal vocal complexity as a depression marker.

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

ReadSaveConnections

This paper combines structural MRI and resting-state functional MRI using a dual cross-attention mechanism that lets each imaging modality refine its representation based on the other, rather than simply concatenating features. Evaluated on the public REST-meta-MDD dataset, the approach outperforms conventional fusion strategies across multiple brain atlas configurations, suggesting the interaction between brain structure and function carries diagnostic information that single-modality or naive multimodal approaches miss. The public dataset enables some external validation, though the absence of a held-out test set is a limitation.

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

ReadSaveConnections

PsychBench tested four major LLMs (including GPT-4o-mini and DeepSeek-V3) by having them generate 28,800 synthetic psychiatric patient profiles and comparing the results against real population data from NHANES and NESARC-III. The key finding is a coherence-fidelity dissociation: LLMs produce individual profiles that sound clinically plausible but collectively misrepresent the true population distribution, compressing variance by 14–62% and erasing the clinical extremes that matter most for risk assessment. This is a direct warning for researchers using LLM-simulated patients to train or test mental health AI tools.

██████████ 0.9 digital-therapeutics Preprint

ReadSaveConnections

This study stress-tests two vision-language models (Phi3.5-Vision and Qwen2-VL) on depression detection using both controlled laboratory video and naturalistic clinical interview footage, finding dramatic performance swings — one model reached 80.4% accuracy in naturalistic settings while the other achieved only 33.9%. Both models showed systematic bias toward over-predicting depression in the laboratory dataset, and the paper applies fairness auditing across gender and race. The results underscore that multimodal AI for mental health diagnosis is still highly environment-dependent and demographically inconsistent.

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

ReadSaveConnections

Analyzing 830 Chinese therapeutic writing samples, this paper shows that macro-level narrative structure — how a person organizes and connects the events in their story — predicts mental health outcomes better than word-level sentiment features or sentence-level semantic embeddings. The mechanism is that psychological distress alters not just what people say but how they structure meaning across an entire narrative, which requires LLM-based analysis to detect. This matters because therapeutic writing is used in many digital mental health interventions, and richer NLP evaluation could make these tools more clinically informative.

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

ReadSaveConnections

K-SENSE improves automated detection of depression and stress in social media text by injecting commonsense knowledge about mental states (drawn from the COMET model) alongside standard language representations, then using contrastive learning to make similar mental states cluster together. It achieves F1 of 86.1% on a stress detection benchmark and 94.3% on a depression dataset, improving over the best prior baselines by 1.5–2.6 percentage points. The practical value is in screening large volumes of social media text where ground truth labels are sparse, though the proprietary Depression_Mixed dataset limits reproducibility.

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

ReadSaveConnections

PsyGAT encodes clinical interview transcripts as graphs where nodes represent utterances characterized by psychological features — capturing not just what is said but clinically meaningful patterns in how people respond — and achieves 89.99 Macro F1 on the DAIC-WoZ benchmark, outperforming both prior graph-based models and GPT-4. A persona-based data augmentation strategy addresses the common problem of class imbalance in depression datasets, where depressed individuals are underrepresented. The interpretable graph structure is notable because it could in principle surface which conversational patterns drove a prediction, supporting clinician trust.

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

ReadSaveConnections