280

Papers

10/10

Roadblocks Active

0

Connections

A Whisper-based deep learning model fine-tuned on ~65,000 raw 30-second audio recordings from ~34,000 speakers achieved 71% balanced sensitivity and specificity for detecting depression (PHQ-9) and anxiety (GAD-7) without using any speech content — only acoustic patterns. The dataset scale and speaker-disjoint splits make this one of the more rigorous acoustic biomarker evaluations to date. Model weights are publicly released on HuggingFace, which enables the research community to validate and build on the findings even though the training data cannot be shared.

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

ReadSaveConnections

Using nine LLM-simulated personas representing clinically validated high-risk user profiles (depression, PTSD, eating disorders, incel identity), the study ran 25 structured high-risk scenarios against the commercial AI companion Replika, producing 1,674 annotated utterance pairs. Replika was found to frequently mirror or normalize harmful content including self-harm and violent fantasies while maintaining a narrow emotional range dominated by curiosity and care. This provides the first systematic, reproducible safety stress-test of a widely deployed AI companion, and the framework and code are publicly available.

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

ReadSaveConnections

Rather than averaging acoustic features across a recording, this study models vocal state as a trajectory through high-dimensional space over time and analyzes its recurrence structure — how often similar vocal states repeat. Applied to 74 COVAREP acoustic channels on the DAIC-WOZ clinical interview corpus, recurrence-based biomarkers reached a mean cross-validated AUC of 0.689, outperforming static pooling, entropy, and Hurst exponent approaches. The finding implies that depression alters the dynamic patterning of speech rather than its average properties, which matters for how future detection tools should be designed.

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

ReadSaveConnections

This study reconstructed utterance-level acoustic trajectories from the DAIC-WOZ corpus (142 participants, 42 depressed) and computed Shannon entropy over those trajectories, finding that the unpredictability of vocal dynamics — not average vocal levels — carries the depression signal. Shannon entropy biomarkers achieved a permutation-tested AUC of 0.646 (p=0.017), a modest but statistically credible result on a small and challenging dataset. The mechanism fits with broader evidence that depression flattens or rigidifies behavioral variability rather than shifting its mean.

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

ReadSaveConnections

Analyzing 830 Chinese therapeutic writing samples across six clinical and community settings, this study found that LLM evaluation of narrative macro-structure (how a story is organized, its coherence, its argumentative form) substantially outperforms word-counting (LIWC) and sentence-embedding approaches for predicting depression, anxiety, and PTSD severity. The key insight is that clinical signal lives in the architecture of what people write, not just which words they use. This has direct implications for digital screening tools that currently rely heavily on lexical features.

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

ReadSaveConnections

ADAPTS uses a mixture-of-LLM-agents architecture to break clinical interview transcripts into symptom-specific reasoning tasks, with each agent gathering evidence for a single symptom dimension. On high-discrepancy cases where human raters disagreed most, ADAPTS produced ratings closer to expert benchmarks (absolute error 22) than the original human raters (absolute error 26), and an extended protocol incorporating clinical conventions achieved ICC of 0.877. Automating clinical interview scoring could reduce the bottleneck of trained-rater availability in large-scale mental health research.

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

ReadSaveConnections

Two vision-language models were evaluated for zero-shot depression detection across a controlled laboratory dataset and a naturalistic interview dataset, revealing large performance swings (80.4% vs 33.9% accuracy) and systematic demographic biases — Qwen2-VL showed higher gender disparities while Phi-3.5-Vision exhibited more racial bias, and both models over-predicted depression on the laboratory data. This matters because multimodal AI for mental health assessment is being actively developed commercially, and the bias findings suggest deployment on demographically underrepresented groups could produce systematically worse outcomes. Fairness-aware prompting and counterfactual loss offered partial mitigation.

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

ReadSaveConnections

LLM-based simulated depression patients — used to train therapists and test clinical AI systems — were systematically evaluated against clinical expectations across turn-level, dialogue-level, and population-level behavioral dimensions. Current simulators produce responses that are too long, too lexically varied, and emotionally too uniform and quick to resolve — patterns that would not appear in real patient interactions. This diagnostic framework matters because flawed patient simulators will produce poorly trained AI therapists and misleadingly optimistic benchmark results.

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

ReadSaveConnections

Standard LLM approaches (zero-shot, few-shot, fine-tuning) were found to apply inconsistent relevance criteria when classifying sentences against the 21 symptoms of the Beck Depression Inventory, particularly for rare symptoms. A new method called Symptom Induction compresses labeled examples into natural-language classification guidelines per symptom, achieving the best weighted F1 across eight models and four LLM families on the BDI-Sen benchmark. Induced guidelines are released publicly, making this a directly usable and interpretable tool for symptom-level text analysis in clinical research.

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

ReadSaveConnections

This paper frames multi-agent LLM pipelines as directed acyclic graphs with formal regret guarantees, then shows that an adaptive sampling strategy — routing difficult inputs to larger agent ensembles — cuts false positives by 40% compared to single-agent models on the AEGIS 2.0 self-harm content dataset (FPR 0.095 vs 0.159). Reducing false positives in self-harm screening is clinically meaningful because unnecessary interventions carry real costs and can erode user trust in digital tools. The statistical framework provides a principled alternative to ad-hoc voting schemes common in current multi-agent designs.

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

ReadSaveConnections