274

Papers

9/9

Roadblocks Active

0

Connections

This framework trains exclusively on healthy EEG data to build a model of normal brain activity, then scores patients by how far their signals deviate from that norm — flagging depression without needing large labeled patient datasets. A separate module handles the common real-world problem of EEG devices with different electrode counts, making it more deployable across clinical sites. If it holds up beyond the two public datasets tested, it addresses a persistent bottleneck in EEG-based depression screening: the scarcity of labeled patient recordings.

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

ReadSaveConnections

Adolescents carrying the A2 allele of the DRD2 dopamine receptor gene show greater susceptibility to suicidal behavior when exposed to family trauma and school bullying, with depression acting as a bridge between these stressors and suicidal outcomes. This gene-environment interaction means at-risk youth are not a uniform group — the same adverse experiences carry different biological risk depending on dopamine signaling genetics. The finding adds mechanistic specificity to broad adverse-childhood-experiences frameworks and points toward genetically-informed triage in youth mental health settings.

██████████ 0.9 youth-mental-health-crisis Peer-reviewed

Read

Researchers combined brainwave (EEG) and frontal blood oxygenation (fNIRS) signals in a joint machine learning pipeline to detect depressive states without relying on self-reports or clinical interviews, with the stated ambition of catching depression before individuals recognize it themselves. The study also proposes this approach for early differentiation between depression and dementia in aging populations, where clinical overlap creates diagnostic delays. Critically, the sample is only 11 healthy university students — this is a proof-of-concept for the measurement paradigm, not a clinical validation, and results should be interpreted accordingly.

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

ReadSaveConnections

A fine-tuned large language model predicts PHQ-9 depression scores from AI therapy conversation transcripts with a correlation of 0.80 and a clinical-threshold AUC of 0.91 — approximating what a clinician would score without asking directly. The training set was expanded from 3,111 to 6,283 examples by using a second AI model to generate pseudo-labels, a practical technique for working with scarce labeled clinical data. The key implication is passive symptom monitoring: patients who are worsening between sessions could be flagged automatically without waiting for a scheduled assessment.

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

ReadSaveConnections

Open-weights language models can estimate depression severity from clinical interview transcripts in zero-shot settings — with no task-specific training — achieving a mean absolute error of 0.60, and a structured feature-extraction step reduces dementia prediction error by up to 35% over zero-shot baselines. The study used 154 German-speaking older adults and tested robustness against automatically-generated transcripts, showing the approach survives transcription noise. This matters because it suggests LLMs could assist structured clinical documentation and severity staging in settings where labeled training data will never exist.

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

ReadSaveConnections

A small pilot RCT tested whether adding a smartwatch-based real-time hyperarousal detection system to a veteran cycling program improved PTSD outcomes beyond exercise alone; the wearable group showed more stable symptom trajectories while the cycling-only group saw hyperarousal escalate in later weeks. The smartwatch combined heart rate and movement data with participant confirmation to detect PTSD-related physiological events in daily life, not just in a clinic. With only 7 veterans in the digital arm the result is hypothesis-generating, but the trial establishes that continuous physiological monitoring during a community wellness program is operationally feasible.

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

ReadSaveConnections

A mixed-methods study combining focus groups and a cross-national survey of 435 mental health service users found that digital self-monitoring is most valued for relapse prevention and building personal autonomy, but four distinct user profiles — varying by tracking experience and technology comfort — mean a single app design will alienate large subgroups. High assessment frequency and limited clinician feedback were identified as the top barriers to sustained engagement. This is actionable for digital therapeutic developers: the evidence suggests the engagement bottleneck is interaction design and clinical integration, not the underlying technology.

██████████ 0.8 digital-therapeutics Peer-reviewed

Read

A reinforcement learning framework called CRPO trains language models to work through mental health assessments in a structured four-stage sequence — mirroring how a clinician moves from observing symptoms to reaching a diagnosis — rather than jumping directly to a label. Tested across 8 datasets covering depression, anxiety, stress, and suicide risk, it outperformed standard RL baselines by 10.4 percentage points in weighted F1. Explicit reasoning structure may improve generalization beyond surface text patterns, which is a persistent failure mode when fine-tuning LLMs on clinical data.

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

ReadSaveConnections

Dep-LLM diagnoses depression from clinical interview transcripts using chain-of-thought prompting that decomposes the interview into five clinically-aligned themes, then uses token-level entropy to weight how confident the model is in each piece of evidence — all without any fine-tuning on labeled patient data. Evaluated across 21 open-source language models on two public datasets (DAIC-WOZ and E-DAIC), the approach removes the labeled-data requirement that blocks clinical deployment in most real-world settings. The practical value is largest in low-resource contexts, though a reproducibility issue with comparisons against unverifiable commercial model versions weakens the benchmarking claims.

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

ReadSaveConnections

A formal mathematical analysis — verified by computer proof — shows that measurement noise attenuates nonlinear interaction signals exponentially by interaction order, while linear effects are only attenuated once; this explains why deep networks and gradient-boosted trees repeatedly fail to outperform logistic regression across 140 UK Biobank prediction tasks. For mental health biomarker research, the implication is direct: collecting more reliable measurements (better sensors, larger samples, standardized protocols) will likely yield more predictive improvement than switching to more complex models. This is an important corrective to the assumption that prediction gaps in psychiatric data reflect insufficient model expressiveness.

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

ReadSaveConnections