278

Papers

10/10

Roadblocks Active

0

Connections

Researchers fine-tuned a Whisper speech model on ~34,000 subjects to extract depression and anxiety signals directly from 30-second audio clips, without using any spoken content — just acoustic patterns. The model reaches 71% sensitivity and specificity on a held-out set of ~5,000 people, and the weights are publicly released on HuggingFace. This matters because it offers a passive, scalable screening tool that could work over phone calls or apps without requiring patients to answer questionnaires.

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

ReadSaveConnections

This framework paper argues that current PTSD therapies like CBT and EMDR address how people respond to trauma triggers but do not dissolve the underlying over-reactive neural pathway itself — a distinction with real treatment implications. The authors propose using a lightweight on-device language model to deliver daily micro-exposures timed to intercept the moment between an unconscious stress signal and conscious elaboration, aiming to weaken the pathway through repeated non-reinforced activation. No clinical trial has been run yet, but the paper outlines an RCT design and is notable for grounding a conversational AI intervention in a specific neuroscientific mechanism rather than generic CBT prompts.

██████████ 0.9 neuroplasticity-interventions Preprint

ReadSaveConnections

TimeSRL converts raw smartphone sensor streams (movement, sleep, app use) into plain-language descriptions, then uses reinforcement learning to train a language model to predict anxiety scores from those descriptions alone — never from the raw numbers directly. Tested in a leave-one-study-out protocol across multiple passive-sensing datasets, it reduces prediction error by 3–10% over standard machine learning baselines and by up to 44% over other LLM approaches. The cross-dataset generalization result is the key contribution: most mental health sensing models fail on new populations, and this approach shows a meaningful step toward portability.

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

ReadSaveConnections

ADAPTS breaks clinical interview transcripts into symptom-by-symptom reasoning tasks using a network of LLM agents, then assembles a depression severity score — mimicking how a trained clinician would work through a structured interview. On high-disagreement cases where two human raters differed most, the automated system came closer to an expert benchmark (mean absolute error 22) than the original human raters did (error 26). This suggests automated tools may have real utility in clinical settings where inter-rater disagreement is the limiting factor, though the reliance on unspecified LLM APIs is a reproducibility concern.

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

ReadSaveConnections

EmoTrack uses LLM-extracted clinical cues combined with frozen semantic embeddings — deliberately avoiding full fine-tuning — to predict PHQ-8 depression scores from therapy session transcripts. It achieves a 13.5% reduction in prediction error over the best prior method on a standard single-session benchmark, and remains competitive on a new multi-session longitudinal dataset. The design choice to freeze embeddings rather than fine-tune is practically important: it prevents models from overfitting to specific therapy protocols and could allow deployment across diverse clinical settings.

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

ReadSaveConnections

PULSE uses an LLM agent that autonomously explores smartphone sensor data — choosing which signals to examine rather than following a fixed script — to predict when cancer survivors want emotional support or are available for a digital health intervention. The agentic approach reaches 74% balanced accuracy for emotion regulation need prediction and 71% for intervention availability, both from passive sensing without requiring the user to fill in mood diaries. The cancer survivorship context is underserved in digital mental health, and the passive-only prediction result is directly relevant to designing interventions that do not add burden to already fatigued patients.

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

ReadSaveConnections

This study tests three LLMs (Phi-4, Gemma-2-9B, Llama-3.1-8B) on estimating anxiety and depression scores from spontaneous speech transcripts, using 111 participants and four different speech-to-text transcription qualities to simulate real-world noise. Phi-4 and Gemma-2-9B maintained strong consistency (ICC > 0.89) even at 10% word error rates, while Llama-3.1-8B collapsed to ICC 0.36 under the same conditions. The practical takeaway is that model choice matters enormously for clinical reliability — a widely used open model may be unsuitable for any speech-based mental health screening pipeline without robustness testing.

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

ReadSaveConnections

This perspective paper identifies a persistent gap in computational neuroscience: biologically detailed brain simulations can reproduce brain structure but cannot perform cognitive tasks, while AI models that perform tasks have no meaningful biological grounding. The authors propose a framework called functional whole-brain models (fWBMs) to bridge the two traditions, with a roadmap for implementation. For psychiatry, this matters because models that are both biologically realistic and task-capable would enable more meaningful simulation of how disorders like depression alter information processing — rather than just altering connectivity statistics.

██████████ 0.8 computational-psychiatry Preprint

ReadSaveConnections

Federated learning — training models across devices without centralizing sensitive data — achieves depression detection F1 of 83.2 versus 85.6 for centralized training, a small and acceptable gap. However, adding differential privacy (a mathematical guarantee against data leakage) causes F1 to drop by up to 27 points even at the loose privacy budget of epsilon=50, and disproportionately destroys the sparse emotion and health-related word features that carry the most diagnostic signal. This is a concrete quantification of a trade-off that will affect every attempt to build privacy-compliant mental health AI at scale.

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

ReadSaveConnections

This study creates continuous depression, anxiety, and worry scores from text by projecting language onto axes defined by items from validated clinical scales — no labeled training data required. Tested on 247 observations from 145 participants, structured formats like selected words and short phrases correlate more strongly with PHQ-9 and GAD-7 scores than free-text entries. The unsupervised approach is notable because it could be deployed immediately in any text-collection context without a labeled clinical dataset, lowering the barrier for research in under-resourced settings.

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

ReadSaveConnections