266

Papers

10/10

Roadblocks Active

0

Connections

An AI system combining large language models with statistical analysis autonomously scanned wearable sensor data from over 9,000 participant observations and identified 41 candidate mental health biomarkers, with sleep timing variability emerging as a depression signal in two independent datasets. The replication of circadian-instability features across distinct cohorts (DWB and GLOBEM) gives these findings more credibility than single-dataset results typical in this space. This matters because it demonstrates that AI-assisted hypothesis generation can surface biologically plausible, reproducible signals from passively collected wearable data — a potential path toward scalable, low-cost depression monitoring.

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

ReadSaveConnections

This paper introduces a framework that jointly analyzes structural brain MRI and resting-state functional MRI using bidirectional cross-attention, allowing each modality to inform interpretation of the other rather than simply concatenating features. The best model achieved 84.71% accuracy and 86.42% sensitivity for MDD detection, outperforming conventional fusion approaches. The clinical implication is that combining brain shape and brain activity patterns captures complementary depression-related information that neither modality contains alone.

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

ReadSaveConnections

Using mice with surgically induced liver disease, this study found that the thalamus — a brain region central to arousal, mood, and cognition — shrinks in volume and shows gene expression patterns consistent with suppressed neural growth and increased cell death. Blocking the inflammatory protein TNF reversed many of these changes, directly linking peripheral liver-driven inflammation to thalamic damage via a targetable molecular pathway. For mental health, this provides a potential mechanism by which systemic inflammatory conditions (including those associated with depression) structurally alter brain regions involved in emotional regulation.

██████████ 0.9 neuroinflammation Peer-reviewed

Read

Across three established depression detection datasets, AI models trained exclusively on the interviewer's words — not the patient's — could still classify depression at above-chance accuracy, because interviewers use predictable prompts at predictable positions that correlate with study labels. This exposes a fundamental validity problem: many published depression detection models may be learning the structure of the interview rather than genuine linguistic markers of depression. The fix proposed — training only on participant speech and using attribution analysis to verify signal source — is simple but would invalidate many existing benchmark comparisons.

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

ReadSaveConnections

This framework uses a large language model to generate progressively detailed clinical summaries of patient interviews — covering emotional tone, specific PHQ-8 symptom dimensions, and inferred causes — and then uses those summaries to guide fusion of audio, video, and text signals for depression severity scoring. The approach improves both accuracy and interpretability compared to existing baselines on the E-DAIC benchmark. However, the reliance on GPT-o3, a proprietary non-deterministic model, means results cannot be independently reproduced, which significantly limits clinical translation confidence.

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

ReadSaveConnections

TAAC proposes splitting voice recordings into separate streams — one carrying depression-related acoustic features, one carrying speaker identity — so that a depression detection system can analyze the clinical signal without storing or transmitting identifying information. The adversarial disentanglement approach is tested on DAIC-WOZ, D-Vlog, and MODMA datasets. Privacy-preserving depression detection from voice is practically important for real-world deployment, though the lack of shared code and incomplete methodology reporting limit independent verification of the claimed performance.

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

ReadSaveConnections

This conceptual paper argues that current brain modeling efforts fail because they treat accuracy as the primary goal while ignoring 'execution semantics' — whether a model correctly handles time, events, and persistent state in the way a real brain does. It proposes a taxonomy ranging from static offline models to real-time co-simulations that run in parallel with biological tissue. For computational psychiatry, this reframes the problem: a model that predicts symptoms accurately from snapshots may still be useless for understanding or intervening in dynamic psychiatric conditions.

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

ReadSaveConnections

The DRIFT framework addresses a core challenge in precision psychiatry: how to estimate which treatment will work best for a given patient when outcomes are measured across many symptom domains simultaneously and the underlying constructs (like 'depression severity') are latent. The method uses adversarial optimization to find treatment recommendations that remain robust even under uncertainty about how symptoms cluster together. Applied to the EMBARC sertraline trial, it provides a principled statistical approach to personalizing antidepressant treatment decisions from high-dimensional symptom data.

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

ReadSaveConnections

Deep learning models trained to classify sleep stages from EEG recordings of healthy people perform poorly when applied to stroke patients with disrupted sleep architecture — and attention map analysis shows the models are literally looking at uninformative parts of the signal in patient data. This is not a minor technical detail: the psychiatric relevance of sleep staging depends on clinical populations, precisely where these models break down. The paper introduces a stroke patient dataset (iSLEEPS) to benchmark this gap, which is a practical contribution for future model development.

██████████ 0.8 sleep-circadian-psychiatry Preprint

ReadSaveConnections

In 222 university students, using AI tools for task-focused purposes (writing, coding, research) correlated with larger prefrontal and hippocampal brain volumes and better grades, while using AI for social or emotional support correlated with smaller amygdala and superior temporal volumes and worse mental health scores including depression and social anxiety. The same technology produces opposite brain and wellbeing associations depending on how it is used. This cross-sectional design cannot establish causation — heavy social AI use may reflect pre-existing distress rather than cause it — but the neuroanatomical specificity of the associations warrants longitudinal follow-up.

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

ReadSaveConnections