214

Papers

10/10

Roadblocks Active

1

Connections

An AI pipeline combining multiple LLM agents automatically screened wearable data from over 9,000 participants and surfaced 41 candidate digital biomarkers for mental health, with sleep-timing variability emerging as a consistent depression signal across two independent cohorts. The key insight is that circadian disruption — not just sleep duration — tracks with depression severity, and this held up to adversarial validation checks built into the pipeline. If the biomarkers replicate, this approach could shift depression monitoring from clinic visits to continuous, passive wearable data.

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

ReadSaveConnections

The DRIFT framework addresses a core problem in precision psychiatry: predicting which patient will benefit from a specific antidepressant when trials only measure a subset of outcomes. By training on latent symptom constructs rather than individual questionnaire items, the method produces treatment-effect estimates that generalize to side effects and patient-reported symptoms never seen during model training. Validated on the EMBARC sertraline vs. placebo RCT, this is one of the more rigorous individualized treatment-effect methods applied to depression to date.

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

ReadSaveConnections

This system uses a large language model to generate progressively detailed clinical summaries at each stage of a three-step depression assessment pipeline — screening, severity classification, and continuous scoring — then fuses those summaries with audio and video signals. The approach outperforms prior state-of-the-art on the E-DAIC benchmark and, crucially, produces human-readable rationales for its predictions, which matters for clinical trust. The main caveat is dependence on a proprietary GPT API, making exact replication difficult.

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

ReadSaveConnections

Brain imaging diagnostic models trained at one hospital typically fail when deployed at another because scanner settings and patient populations differ. This paper proposes a continual learning system that synthesizes realistic fake fMRI connectivity data from previous sites and replays it during training on new sites, preventing the model from forgetting what it learned before. The method is the first specifically designed for this multi-site fMRI problem and could help psychiatric AI tools remain accurate as they are rolled out across health systems.

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

ReadSaveConnections

In 222 university students who underwent brain MRI, using AI chatbots for task-focused purposes was associated with larger prefrontal and memory-related brain regions and better academic performance, while using them for emotional support and social substitution was linked to higher depression and social anxiety scores. This is among the first studies to connect AI usage patterns to both brain structure and mental health outcomes in young people. The cross-sectional design means causality cannot be established, but the dissociation between functional and socio-emotional use is a clear signal worth tracking longitudinally.

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

ReadSaveConnections

Using EEG recorded while healthy adults categorized emotional faces, the study quantifies how stable the brain's network state is during sad versus happy processing by treating brain connectivity as an energy landscape — deeper energy wells mean more stable, harder-to-exit states. Sad processing drove the brain into significantly more stable states in the alpha band (effect size Cohen's d = 0.83), and deeper stability was associated with slower responses, suggesting a mechanism for why depressive cognition can feel 'stuck'. The sample is small (N=20) and healthy, but the energy-landscape approach offers a quantitative bridge between EEG biomarkers and computational models of rumination.

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

ReadSaveConnections

Models trained only on the interviewer's side of three major depression interview datasets — never seeing patient speech — still classified depression versus control at above-chance rates, revealing that fixed script structure and prompt positioning are leaking label information into these benchmarks. This means many published depression detection results may be inflated by artifacts in the interview design rather than genuine detection of patient language patterns. Any research team using DAIC-WOZ, E-DAIC, or ANDROIDS should treat prior accuracy numbers with caution until participant-only models are re-evaluated.

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

ReadSaveConnections

Wearable sensors alone (heart rate variability, electrodermal activity) classified physical stress well but struggled with psychological stress, achieving only 50% recall — a failure mode directly relevant to depression monitoring where psychological states are the target. Adding a single cortisol sample raised psychological stress recall to 83%, demonstrating that HPA-axis endocrine markers provide information that autonomic signals cannot. This multimodal fusion result directly informs depression biomarker design: purely digital or purely physiological approaches may need a cortisol anchor to reach clinical-grade specificity.

██████████ 0.7 depression-biomarkers Preprint

ReadSaveConnections

Deep learning sleep-staging models trained on healthy individuals performed poorly on patients with ischemic stroke, and attention maps showed the models were focusing on EEG patterns that carry no physiological meaning in the patient group. This generalisation failure matters for psychiatric care because most validated sleep AI tools were trained on healthy or simple clinical populations, yet the patients who most need sleep monitoring — those with depression, stroke, or other comorbidities — have the most atypical EEG. The finding calls for disorder-specific training data and domain-adaptation strategies before clinical deployment.

██████████ 0.7 sleep-circadian-psychiatry Preprint

ReadSaveConnections

This theoretical paper proposes measuring brain plasticity as a single ratio — the number of network nodes divided by total connection strength — and argues that an optimal intermediate value of this ratio predicts the sweet spot between being too rigid (unable to change) and too unstable (unable to maintain any state). While purely mathematical with no new data, the framework provides a concrete, measurable quantity that could be applied to existing connectome datasets to test whether antidepressants or psychedelics shift networks toward higher plasticity. The generality across neuroscience, ecology, and economics is interesting but also means psychiatric validation is still needed.

██████████ 0.7 neuroplasticity-interventions Preprint

ReadSaveConnections