290

Papers

10/10

Roadblocks Active

0

Connections

A deep learning model trained on raw 30-second speech clips — using a fine-tuned Whisper backbone on ~65,000 utterances — can detect depression and anxiety at 71% balanced sensitivity and specificity without analyzing the words spoken, only the acoustic signal. This matters because it points toward passive, low-friction screening that could work in telehealth or consumer apps without requiring users to answer questionnaires. The limitation is real: labels come from self-report scales, not clinician diagnoses, and the training data is proprietary, so the 71% figure cannot yet be independently verified.

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

ReadSaveConnections

Using spiking neural networks as a computational brain model, researchers found that removing Von Economo neurons (VENs) — a cell type lost early in frontotemporal dementia and implicated in autism and schizophrenia — caused learning to fail completely in 30% of simulations, compared to only 2% failure when VENs were intact (p=8.7e-5). The key insight is that VENs are not just speed regulators but reliability gatekeepers: networks without them either learned normally or not at all, with nothing in between. This gives computational psychiatry a specific, testable prediction about why VEN-depleting conditions produce such variable clinical presentations.

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

ReadSaveConnections

In a study of 50 cancer survivors — who face elevated rates of depression and anxiety — an LLM agent that autonomously explored smartphone sensing data (location, sleep, screen use) outperformed structured single-call approaches, reaching 74% balanced accuracy for predicting when a person wanted to regulate their emotions. The work directly addresses the 'diary paradox': people are least likely to log their mental state exactly when it's worst, so passive sensing must carry the load. An agent that can intelligently query its own data tools, rather than apply a fixed formula, proves significantly better at that job.

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

ReadSaveConnections

This systematic review highlights that bipolar disorder affects 2-3% of the global population but diagnosis is routinely delayed by years, with serious downstream consequences including suicide attempts and prolonged social dysfunction. The authors survey modern diagnostic tools and find that despite their existence, none have achieved broad clinical adoption — the gap is implementation, not invention. This is relevant context for any biomarker or digital tool research: the translation pipeline from 'technically works' to 'routinely used in clinics' is where most progress is currently blocked.

██████████ 0.8 depression-biomarkers Peer-reviewed

Read

This paper proposes a conversational AI system for PTSD that targets an earlier point in the stress response chain than existing therapies: instead of working through traumatic memories after the fear cascade fires (as in CBT or EMDR), MindGap aims to interrupt the moment between an incoming sensory signal and the automatic reactive elaboration. The mechanism draws on Hebbian plasticity — the same process that entrenches PTSD pathways can theoretically dissolve them with repeated interception. Critically, this is entirely theoretical with no clinical data; the value is the mechanistic target it identifies, not any demonstrated outcome.

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

ReadSaveConnections

Training depression detection models across distributed devices (federated learning) achieves F1 scores only 2.5 points below centralized training (83.2 vs 85.6), but adding differential privacy — the technique that prevents the model from leaking individual user data — causes accuracy to collapse by up to 27 F1 points even at relatively lenient privacy settings. The reason is specific: mental health language is sparse but highly distinctive, and privacy noise disproportionately destroys exactly those rare emotional and health-related word signals. This is a quantified demonstration that current privacy-utility trade-offs are not acceptable for clinical deployment.

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

ReadSaveConnections

ADAPTS uses a mixture of LLM agents to rate depression and anxiety severity from clinical interview transcripts by breaking the task into symptom-by-symptom sub-problems rather than asking a single model to rate everything at once. On interviews where human raters disagreed most, the automated system (absolute error 22) actually outperformed the original human ratings (absolute error 26), achieving an ICC of 0.877 when using clinical conventions. This matters because inter-rater disagreement is a structural problem in psychiatric assessment, and an AI that performs better precisely in the hard cases could improve trial quality and reduce measurement noise.

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

ReadSaveConnections

TimeSRL converts raw passive sensing streams (sleep, mobility, device use) into natural language descriptions, then predicts depression and anxiety scores from those descriptions alone using a reinforcement-learning-tuned LLM — no direct numerical feature engineering. Tested using a rigorous leave-one-study-out protocol across multiple datasets, it reduces prediction error 3-10% over standard machine learning baselines. The practical value is generalizability: by working in language space rather than feature space, the system adapts to different sensor configurations and study populations without retraining from scratch.

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

ReadSaveConnections

EmoTrack combines LLM-extracted clinical signals with a compact memory of the previous therapy session to track depression severity across counseling conversations, reducing PHQ-8 prediction error by 13.5% over the best existing DAIC-WOZ baseline in single-session settings. The memory mechanism is key: it lets the model ask 'has this patient changed since last time?' rather than treating each session in isolation, which mirrors how skilled clinicians actually track patients. A notable caveat is that multi-session evaluation used a synthetic dataset with LLM-generated transcripts, so real-world performance across sessions remains unverified.

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

ReadSaveConnections

This study shows that you can measure depression, anxiety, and worry from text without any labeled training data by projecting sentence embeddings onto axes defined by existing clinical scale items — essentially using validated questionnaires as a compass for meaning space. In 247 observations from 145 participants, structured word-based responses correlated substantially better with clinical scores than free text, suggesting that small amounts of scaffolding (asking people to list words rather than write freely) dramatically improve signal quality. This is an unusually clean, theory-grounded approach in a space crowded with black-box models.

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

ReadSaveConnections