288

Papers

10/10

Roadblocks Active

0

Connections

A fine-tuned Whisper speech model trained on ~65,000 recordings from ~34,000 people can detect depression and anxiety from speech acoustics alone — without analyzing the words spoken — achieving 71% sensitivity and specificity across multiple severity thresholds. The key mechanism is that depression and anxiety alter vocal dynamics in ways captured in time- and frequency-domain representations of raw audio. A pre-trained model is publicly released on HuggingFace, enabling others to test it, though the proprietary training data means independent replication of the full pipeline is not possible.

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

ReadSaveConnections

This paper applies nonlinear dynamical systems analysis to speech, modeling how the vocal system revisits acoustic states over time (recurrence structure) rather than using standard acoustic averages. Depression-related recurrence biomarkers achieved a cross-validated AUC of 0.689 and outperformed several competing acoustic feature classes, suggesting depression disrupts the temporal organization of voice production in a measurable way. This independently corroborates the deep-learning voice biomarker approach found in another paper today, increasing confidence that vocal dynamics carry genuine diagnostic signal.

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

ReadSaveConnections

MindGap proposes a conversational AI grounded in Buddhist psychological theory (dependent origination) that aims to interrupt PTSD's over-reactive amygdala-HPA stress cascade before it fully activates, rather than managing it after the fact as current therapies do. The system is designed to run on-device using a lightweight LLM, making it potentially accessible without internet connectivity or a therapist. This is a theory-and-design paper only — no clinical data yet — and its proposed RCT has not been conducted, so efficacy claims are entirely speculative at this stage.

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

ReadSaveConnections

This perspective paper argues that computational psychiatry is stuck between two incomplete traditions: detailed brain simulations that cannot perform cognitive tasks, and AI models that perform tasks but ignore brain biology. It proposes 'functional whole-brain models' as a unifying framework requiring both structural realism (real connectomes) and functional task competence. If realized, such models could help identify which brain circuit properties cause specific psychiatric symptoms, guiding drug and neuromodulation targets — but the framework currently lacks operational specificity for independent implementation.

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

ReadSaveConnections

ADAPTS uses a multi-agent LLM system to automatically score depression severity from clinical interview transcripts by breaking the interview into symptom-by-symptom reasoning tasks, mimicking how a trained clinician approaches structured assessment. On difficult cases where human raters disagreed, the system outperformed the original human ratings (absolute error 22 vs. 26 compared to an expert benchmark), and achieved strong inter-rater reliability (ICC=0.877) with an extended scoring protocol. This could reduce the burden of manual symptom tracking in large clinical datasets, though reproducibility is limited by undisclosed models and datasets.

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

ReadSaveConnections

Analyzing how people structure their stories — the macro-level narrative organization — predicts mental health outcomes better than counting specific words (lexical features) or measuring sentence-level meaning (semantic embeddings), across 830 therapeutic writing samples covering depression, anxiety, and PTSD. The finding suggests that psychological distress shows up not just in what people say but in how they organize their accounts, which is a meaningful mechanistic insight for clinical NLP. Results come from Chinese-language data across six heterogeneous interventions, so generalizability to other languages and contexts is unconfirmed.

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

ReadSaveConnections

Cancer survivors frequently experience depression and anxiety but rarely log their distress at the moments it occurs — PULSE addresses this by using an LLM agent that autonomously queries wearable and smartphone sensor data to infer when a survivor might want support. The agentic approach, which can ask follow-up questions of the data much as a clinician might, achieved 74.3% balanced accuracy for predicting desire for emotion regulation support, outperforming simpler structured methods. The dataset involves 50 participants with protected health data, making independent replication unlikely in the near term.

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

ReadSaveConnections

When 15 major AI chatbots were tested on psychiatric triage vignettes, overall accuracy ranged from 42% to 72%, but critically, 5.6% of genuine psychiatric emergencies were classified as non-emergency — a pattern called under-triage that could result in delayed life-saving care. Performance was best for clear emergencies (94% accuracy) and worst for moderate-urgency cases like 'assessment needed within a week' (20% accuracy), revealing that current models struggle most with the nuanced middle ground that dominates real clinical settings. This is one of the first large-scale systematic evaluations of AI chatbot safety specifically for psychiatric triage.

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

ReadSaveConnections

This paper shows that depression, anxiety, and worry can be measured from text without any supervised machine learning training, by projecting text embeddings onto geometric axes derived from clinical scale items in a semantic vector space. The method works best when people write structured responses (word lists or short phrases) rather than open-ended paragraphs, though sentence-level aggregation of free text partially closes the gap. This unsupervised approach could enable low-cost screening using clinical frameworks that already exist, though the 247-observation dataset limits confidence in generalizability.

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

ReadSaveConnections

A theoretical paradox in consciousness research — that deep meditation states show high brain entropy despite minimal subjective experience, seemingly contradicting a leading theory of psychedelic-assisted therapy — is resolved here by distinguishing brain entropy from brain complexity. The paper argues that complexity (structured, differentiated neural activity) rather than raw entropy is what tracks the richness of experience, which has direct implications for understanding why psychedelics produce therapeutic effects and how to measure them. This matters for psychedelic-assisted therapy research because it suggests current neuroimaging metrics may be measuring the wrong quantity when evaluating treatment response.

██████████ 0.8 psychedelic-mechanisms Preprint

ReadSaveConnections