282

Papers

10/10

Roadblocks Active

0

Connections

A multi-agent AI system analyzed wearable sensor data from over 9,000 participant-observations and identified 41 candidate digital biomarkers for mental health, with sleep duration variability and sleep onset variability surfacing consistently across two independent depression cohorts. The finding matters because it suggests the circadian system—not just average sleep amount—carries depression-relevant information that passively collected wrist data can detect without clinical visits. Caveat: the primary dataset is proprietary and the LLM-based pipeline is non-deterministic, so the specific biomarker list should be treated as hypothesis-generating rather than validated.

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

ReadSaveConnections

Using 142 labeled clinical interview recordings, this study shows that Shannon entropy of how a speaker's voice changes over conversation time outperforms averaging acoustic features for detecting depression (AUC 0.646 vs 0.593), with statistical validation via permutation testing. The key insight is that depression lives in the dynamics—the irregularity of vocal patterns—rather than in whether someone's voice is on average higher or lower pitched. The dataset is small and the paper is a solo preprint with a mid-study data corruption event, so effect sizes should be interpreted cautiously.

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

ReadSaveConnections

This system uses multiple LLM agents, each responsible for reasoning about a single depression or anxiety symptom from a clinical interview transcript, and achieves an intraclass correlation of 0.877 for severity rating—approaching expert-level agreement even when the original human raters disagreed. The significance is that it works across different structured interview formats without retraining, which is a practical barrier for deploying automated rating in diverse clinic settings. Reproducibility is limited because the specific LLM models and full prompts are undisclosed.

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

ReadSaveConnections

Evaluating two vision-language models on depression classification using combined audio, facial movement, and text features, this study finds performance varies wildly (33.9% to 80.4% accuracy) depending on the dataset, and that both models show systematic demographic biases—one skewing by gender, one by race. This matters because AI depression screeners that look unbiased on aggregate benchmarks may still fail specific groups in ways that would cause clinical harm. Fairness-aware prompting and counterfactual loss partially but incompletely corrected these disparities.

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

ReadSaveConnections

Generating 28,800 synthetic patient profiles from four major LLMs and comparing them to US epidemiological baselines, this study finds that LLMs produce superficially plausible individuals but systematically eliminate the extremes of clinical reality—symptom variance is compressed by 14–62% and 37% of cases change their diagnostic status between runs. For researchers using LLMs to simulate patients for training data or study design, this is a significant validity problem: the models reflect the statistical middle of the training distribution, not the full range of human suffering that clinicians actually treat.

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

ReadSaveConnections

PsyGAT models clinical interview conversations as temporal graphs where each utterance is a node encoding clinical evidence, and a separate module traces which conversational moments triggered which depression symptoms—achieving 89.99 Macro F1 on the DAIC-WoZ benchmark while outperforming GPT-class models. The clinical value is the interpretability layer: unlike black-box classifiers, this system can point to specific exchanges that drove a severity assessment, which matters for clinician trust and audit. The benchmark comparison to 'GPT-5' cannot currently be verified, and the training data relies on LLM-generated synthetic augmentation.

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

ReadSaveConnections

A structured pipeline of three specialized LLM agents (worker, risk, legal) with an adaptive sampling strategy reduced false positives for self-harm risk screening by 40% compared to a single-agent model, across two public behavioral health datasets totaling 411 samples. False positives in self-harm screening have real costs—unnecessary interventions, alert fatigue, erosion of trust—so precision improvements matter clinically even if sensitivity holds steady. The theoretical regret bounds add credibility, but the full prompting and model specifications are not disclosed, limiting external validation.

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

ReadSaveConnections

Using psychometrically validated synthetic personas representing vulnerable users (depression, PTSD, eating disorders, PTSD), this study found that Replika—a widely deployed AI companion app—frequently mirrors or normalizes self-harm narratives and validates disordered eating behaviors across multi-turn conversations. This documents a concrete harm pathway: emotionally vulnerable users seeking connection receive responses that may reinforce dangerous cognitions rather than redirect them. Because Replika's backend is proprietary and frequently updated, findings reflect a snapshot and may not persist, but the evaluation methodology itself is reusable.

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

ReadSaveConnections

This study demonstrates that heart rate variability metrics derived from a wearable ECG device can classify psychiatric disorders using machine learning, adding to a growing body of evidence that the autonomic nervous system carries diagnostic signal for conditions like depression and anxiety. The significance is that wearable ECG is already consumer-accessible (smartwatches), so a validated HRV-based classifier could enable passive, continuous psychiatric monitoring without clinical visits. Full methodological details are not available from the abstract alone, limiting confidence in effect sizes.

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

Read

Applying K-Means clustering to survey data from 551 social media users, this study identified six distinct behavioral-psychological profiles and found a modest correlation (r=0.28) between daily social media hours and anxiety symptoms. The value is descriptive rather than causal: it suggests that the relationship between social media and mental health is not uniform across users, and that subgroup segmentation may be more informative than population-average effects for intervention design. Reproducibility is very low—no code, data, or instrument is shared—and the cluster solution (k=6, Silhouette=0.32) is relatively weak.

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

ReadSaveConnections