282

Papers

9/9

Roadblocks Active

0

Connections

Researchers systematically attacked eight major commercial LLMs (like those powering popular AI assistants) using four adversarial techniques across all 16 DSM-5 clinical categories, finding that safety protections hold for suicide and self-harm but fail catastrophically — up to 100% of the time — for eating disorders, substance use disorder, and major depression. The attack methods were simple: framing requests as journalism or embedding harmful scenarios in fiction. This matters because millions of people with these disorders are already using general-purpose AI tools, and the uneven protection means the highest-volume conditions are the least defended.

██████████ 0.9 digital-therapeutics Preprint

ReadSaveConnections

This sham-controlled trial tested whether directing a near-infrared laser (980 nm) at the right prefrontal cortex for 10 minutes daily over one week could reduce insomnia in 37 college students, targeting the brain region known to show reduced activity in people with sleep disorders. Outcomes measured across four time points included sleep quality scales, EEG resting-state activity, and a cognitive task analyzed with drift diffusion modeling — an unusually rigorous multimodal battery for a pilot. Insomnia is both a major psychiatric risk factor and highly prevalent in young people, so a safe, non-pharmacological brain-stimulation approach addressing its neural substrate would be clinically meaningful if larger trials confirm these early signals.

██████████ 0.9 sleep-circadian-psychiatry Preprint

ReadSaveConnections

This pilot study had 11 healthy university students wear simultaneous EEG and near-infrared brain-imaging (fNIRS) headsets during a memory task, then used an end-to-end neural network to classify depressive states — bypassing the subjective self-report questionnaires that currently dominate depression diagnosis. The goal is a future where a brief cognitive task with wearable sensors provides an objective biological fingerprint of depression, which would be transformative for early detection and for distinguishing depression from conditions like early dementia that share symptom profiles. The sample is too small to draw conclusions yet, but the pipeline and methodology are clearly described for future scaling.

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

ReadSaveConnections

The study applied multifractal analysis to sleep EEG recordings and found that people who later develop dementia show measurably different brain signal dynamics during deep sleep (N2 and N3 stages) compared to those who remain cognitively healthy — specifically, their brain activity is farther from the optimal 'critical' state that maximizes information processing efficiency. Sleep EEG is already routinely collected in sleep labs and is non-invasive, meaning this approach could become a practical early-warning biomarker years before cognitive symptoms appear. The statistical separation between groups was strong (p ≤ 0.001), though independent replication is needed.

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

ReadSaveConnections

Veterans with PTSD, depression, and anxiety were randomized to either a multi-week endurance cycling program alone or cycling plus a smartwatch-based intervention that detected physiological hyperarousal in real time and prompted self-management responses. The digital intervention group showed more stable symptom trajectories — particularly for the hyperarousal characteristic of PTSD — while the cycling-only group showed symptom escalation late in the study. The trial is very small (n=13 randomized), so results are preliminary, but the real-time biosensor-to-behavior feedback loop is a promising design for scalable veteran mental health support.

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

ReadSaveConnections

This computational study trained neural network models of working memory and then systematically tested eight different ways of introducing 'stress' as a change in how brain cells communicate — finding that strengthening inhibitory-to-excitatory connections alone reproduced all three well-documented signatures of stress-impaired prefrontal function: inhibitory dominance, excitatory neuron underactivity, and degraded task performance. Crucially, networks trained while simultaneously exposed to the stress operator ('resilience training,' analogous to stress inoculation therapy) maintained performance under stress by staying in the same efficient computational regime. This gives a mechanistic neural-circuit explanation for why stress inoculation works, with implications for both understanding treatment-resistant depression and designing interventions.

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

ReadSaveConnections

A transformer-based foundation model pre-trained on ~3,000 subjects' EEG data was fine-tuned using a progressive layer-unfreezing strategy and then tested on entirely unseen individuals — achieving better cross-subject generalization in predicting cognitive performance (reaction time) than standard convolutional neural network approaches. The key advance is demonstrating that large-scale EEG models can generalize without requiring individual calibration sessions, which has been a major practical barrier to using brain signals as mental health biomarkers at scale. The work is methodologically significant infrastructure for future psychiatric biomarker pipelines.

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

ReadSaveConnections

This study tried to generate synthetic clinical mental health dialogues in Mandarin, Bengali, and Hindi by simply telling an LLM to use different nationalities and languages within a clinically validated English persona framework — and found this approach produces inconsistent results that LLM evaluators themselves struggle to score accurately in non-English text. The finding is practically important because most AI mental health tools are developed in English and then assumed to transfer to other languages, but this work shows the clinical validity of depression severity signals degrades significantly in that transfer. Equitable global mental health AI requires dedicated non-English training pipelines, not language-parameter hacks.

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

ReadSaveConnections

Researchers fine-tuned deep convolutional neural networks originally trained for action recognition — detecting whole-body movement patterns in video — to instead score depression severity from facial videos, and used saliency maps to identify which temporal facial expression patterns the model was using. This is notable because action recognition networks capture the dynamics of movement over time rather than static facial snapshots, which better reflects how depression actually manifests in slowed, reduced expressiveness. The approach moves depression video analysis toward a more interpretable and clinically grounded methodology.

██████████ 0.8 depression-biomarkers 🔗 2 cited Preprint

ReadSaveConnections

Using mechanistic interpretability techniques, this study looked inside Llama-3.1-8B to find which internal neural features actually drive its suicidality detection — and found a low-dimensional semantic feature causally responsible for binary suicide detection that recurs across three model families and three separate suicide datasets. Importantly, the study also found that the model cannot reliably distinguish implicit suicidal intent from general distress, revealing a gap between surface-level performance and genuine clinical understanding. Understanding what LLMs are actually 'detecting' when they flag suicidality is critical before these systems are deployed in crisis contexts.

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

ReadSaveConnections