281

Papers

9/9

Roadblocks Active

0

Connections

A deep learning model fine-tuned on raw 30-second speech clips — without reading the words themselves — achieves 71% sensitivity and specificity for detecting depression and anxiety across roughly 5,000 people, outperforming hand-crafted acoustic features. The model adapts a pre-trained speech transcription backbone (Whisper) using a parameter-efficient technique (LoRA) and is released publicly on HuggingFace, making it one of the few voice-biomarker tools others can actually test. This matters because it demonstrates that content-agnostic vocal patterns carry real clinical signal at a scale and accessibility that could enable passive screening.

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

ReadSaveConnections

Rather than averaging acoustic features across a recording, this study tracks how unpredictable (high-entropy) a person's vocal patterns are moment-to-moment during conversation, finding that this temporal disorder measure outperforms static pooled features for depression detection (AUC 0.646 vs 0.593) with statistical significance confirmed by permutation testing. The finding implies that depression leaves its mark not in how loud or flat someone sounds on average, but in the irregular rhythm of how their voice changes over time. Reproducibility is limited by use of a restricted dataset and absent code sharing, so independent confirmation is needed before clinical interpretation.

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

ReadSaveConnections

This paper applies recurrence quantification analysis — a technique from nonlinear dynamics that measures how often a system revisits the same states — to voice recordings and achieves a cross-validated AUC of 0.689 for depression detection, beating entropy-based, fractal, and simple acoustic baselines on the same DAIC-WOZ dataset. The intuition is that a depressed vocal system gets 'stuck' in repetitive acoustic patterns in ways that classical averages miss. Together with the entropy paper above, this suggests that time-structure features of voice, not just average properties, are where the depression signal lives.

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

ReadSaveConnections

Using clinically validated personas representing high-risk users (e.g., users with self-harm ideation or disordered eating), this study finds that Replika — one of the most widely used AI companion apps — frequently mirrors or normalizes dangerous content rather than deflecting it, and shows a narrow emotional palette dominated by curiosity and care that fails to adapt to crisis signals. This is important because millions of vulnerable people use AI companions as a first point of contact for distress, and the findings expose a concrete safety gap that is not yet addressed by existing guardrails. The methodology of persona-based adversarial evaluation provides a replicable template for auditing other companion systems.

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

ReadSaveConnections

This study creates continuous depression, anxiety, and worry scores from written language by projecting sentence embeddings onto semantic axes built from clinical scale items — no labelled training data required. The resulting scores correlate strongly with validated clinical measures (PHQ-9, GAD-7, CES-D) and work best when participants write structured responses (selected words, short phrases) rather than open free text, though sentence-level aggregation rescues much of the signal in free text. The unsupervised design is significant because it means the approach can in principle be applied to any language sample without requiring expensive annotated datasets.

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

ReadSaveConnections

Two leading vision-language models (Phi-3.5-Vision and Qwen2-VL) are evaluated for binary depression detection using audio, video, and text from clinical interviews, revealing dramatic performance swings — 80% accuracy on one dataset, 34% on another — and systematic bias where models over-predict depression and show measurable gender and racial disparities. The study then tests fairness-aware prompting (chain-of-thought fairness rules and counterfactual loss) as a mitigation strategy and finds partial but incomplete improvement. This matters because AI depression screeners that perform unevenly across demographic groups could cause worse harm than no screening at all.

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

ReadSaveConnections

ADAPTS uses a multi-agent large language model architecture to automatically rate depression and anxiety severity from clinical interview transcripts by breaking the task into symptom-specific sub-tasks, each handled by a dedicated reasoning agent. On interviews where human raters disagreed, the automated system's ratings were actually closer to a gold-standard expert re-rating (absolute error 22) than the original human ratings (error 26), with an inter-rater reliability (ICC) of 0.877. If this holds up under external validation, it could meaningfully reduce the cost and inconsistency of clinical rating in large-scale research or screening pipelines.

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

ReadSaveConnections

Patient simulators — AI systems that role-play as depressed patients for clinician training or therapy testing — currently produce responses that are too long, too emotionally uniform, and resolve distress far too quickly compared to real patient conversations, according to a benchmark that tests seven large language models against an actual patient conversation dataset. PSI-Bench introduces clinically grounded metrics at the turn, dialogue, and population level, validated by 20 mental health experts. This matters because unrealistic simulators could train therapists or tune AI therapists on false behavioural norms, propagating systematic errors into deployed tools.

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

ReadSaveConnections

This paper argues that existing PTSD treatments (prolonged exposure, EMDR, CBT) address distress after the stress cascade has already fired, and proposes a conversational AI framework aimed at the earlier moment — between an initial threat signal and the elaborative reactive thought — where neuroplastic rewiring could theoretically prevent the cascade from engaging. The mechanism draws on Hebbian learning and dependent origination (a Buddhist cognitive model) to identify a tractable intervention window. While the theoretical framing is substantive, the paper offers no clinical trial data, making this a hypothesis-generation contribution that requires prospective testing before its claims can be evaluated.

██████████ 0.8 neuroplasticity-interventions Peer-reviewed

ReadSaveConnections

Applying K-Means clustering to survey data from 551 participants, this study identifies six user segments that differ in their combination of social media usage intensity and psychological well-being, with a modest correlation (r=0.28) between hours of use and anxiety symptoms. The cluster separation quality (Silhouette Score 0.32) is only moderate, and the cross-sectional design cannot establish whether heavy use causes anxiety or vice versa. The study's main value is descriptive — it suggests heterogeneity in how social media relates to mental health rather than a single universal effect — but should be treated as hypothesis-generating given the weak methodology.

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

ReadSaveConnections