278

Papers

9/9

Roadblocks Active

0

Connections

Depression alters how different brain regions communicate, creating hierarchical network structures that standard graph neural networks — which assume flat, Euclidean geometry — systematically miss. This paper uses hyperbolic geometry, which naturally represents tree-like hierarchies, combined with a personalized graph construction that adapts to each individual's EEG data rather than applying a one-size-fits-all brain network template. The result is an EEG-based depression classifier that better captures the actual architecture of depressed brain connectivity, which matters because a more accurate biomarker could eventually reduce reliance on subjective self-report for diagnosis.

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

ReadSaveConnections

Rather than feeding dozens of separate behavioral measures (sleep timing, napping, physical activity, social engagement) into a depression classifier, this study compresses them all into a single Circadian Rhythm Score that preserves nearly all the predictive information. Tested on over 15,000 adults in a Chinese longitudinal survey, the resulting model achieves an AUC of 0.825 — meaning it correctly ranks a depressed person above a non-depressed person about 82% of the time. The approach also uses counterfactual modeling to estimate how much behavioral change would be needed per individual, moving toward actionable intervention recommendations rather than just screening labels.

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

ReadSaveConnections

By linking PHQ-8 depression scores from 766 participants to their actual donated ChatGPT conversation logs — nearly 190,000 conversations in total — this study finds that people with higher depression scores use the chatbot differently: more late-night sessions, more topics around loneliness and interpersonal problems, and language patterns featuring more first-person singular pronouns and absolutist words like 'always' and 'never'. This is significant because it shows that conversational AI is already functioning as an informal mental health resource for people who are struggling, and that passive behavioral signals in those interactions may carry diagnostic information — with major implications for both opportunity and privacy risk.

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

ReadSaveConnections

This paper systematically induces seven psychiatric disorder-like states — including anxiety, depression, OCD, addiction, and PTSD — in AI reinforcement learning agents by tuning a single internal 'appraisal' signal for each, producing graded dose-response behavioral shifts that mirror human clinical profiles. The fact that these disorders self-organize into a two-dimensional affective space (with mania as a mirror image of anxiety) aligns with longstanding dimensional theories of psychopathology. Beyond theoretical interest, this creates a controlled test environment where candidate therapeutic interventions can be trialed on a model system before moving to human studies, potentially accelerating the translation of computational psychiatry into clinical tools.

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

ReadSaveConnections

The study identifies specific internal units in vision-language AI models that behave like the brain's reward-anticipation circuitry (the nucleus accumbens), and shows that directly suppressing those units causes the model to avoid high-effort, high-reward choices — a behavioral signature of anhedonia, the loss of pleasure and motivation seen in depression. Critically, the perturbed models still perform normally on tasks without reward framing, confirming the deficit is specific to reward valuation rather than general capability loss. This mechanistic approach to AI anhedonia offers a novel proxy system for studying dopamine-related motivation deficits, and may help reverse-engineer which neural circuit properties are necessary and sufficient for reward-based decision-making.

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

ReadSaveConnections

This adversarial audit of eight major commercial AI chatbots finds that safety guardrails reliably block only suicide and self-harm content, while failing at rates up to 100% for eating disorders, substance use, and major depressive disorder when researchers use realistic attack strategies like fiction framing or journalism roleplay. The study tests across all 16 DSM-5 mental health condition categories, providing systematic coverage that prior work lacked. The implication is urgent: millions of people with eating disorders or depression are already using general-purpose chatbots for mental health support, and the safety infrastructure protecting them is condition-specific in ways that do not match actual clinical need.

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

ReadSaveConnections

People express depression online in very different ways — some disclose directly, others only hint at it through behavioral patterns like posting at unusual times or making vague negative comments — and a single classifier averaged across all users systematically misses the non-disclosing group. This paper addresses that problem by training a mixture of specialized sub-models (experts), with each user's data routed to the most appropriate expert based on three weak behavioral signals detectable at inference time. Tested on Chinese and English social media datasets totaling nearly a million posts, the approach outperforms strong baselines and provides interpretable routing decisions that explain why a particular user was flagged.

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

ReadSaveConnections

One of the core questions in precision psychiatry is whether tailoring a treatment to an individual actually produces better outcomes than simply giving everyone the treatment that works best on average — and until now there was no rigorous statistical test for this. This paper provides exactly that: a method that uses historical trial data to test whether a personalized policy significantly outperforms the single best universal intervention, with formal error control. Applied to a depression treatment dataset, among others, it provides a principled framework for deciding when personalization is worth the added complexity and cost — a question that has been largely argued on theoretical grounds rather than tested empirically.

██████████ 0.8 treatment-resistant-depression Preprint

ReadSaveConnections

This paper dissects how a widely used open-source language model (Llama-3.1-8B) processes suicidal language, finding that the model cannot reliably distinguish implicit suicidal intent from ordinary emotional distress — a critical failure for any clinical deployment. Using interpretability techniques, the authors identify a specific mid-network semantic feature that is causally responsible for binary suicide detection, and show that this feature is low-rank and generalizes across three different suicide datasets and three model families. The finding matters because it reveals a concrete, mechanistic vulnerability rather than just documenting that the model fails, opening a path toward targeted fixes rather than blanket retraining.

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

ReadSaveConnections

The error-related negativity (ERN) is a brain signal measurable via EEG that spikes when someone makes a mistake — and it is known to be elevated in youth at risk for anxiety and depression. This longitudinal study in female adolescents shows that high ERN at baseline predicts worsening internalizing symptoms over time, but the pathway differs by social domain: elevated ERN leads to more negative interpretations of family events (which then worsen symptoms), while for peer relationships the direction reverses and symptoms come first, then distorted perception. This domain-specific dissociation suggests that early interventions targeting how adolescents interpret family interactions might be more effective than generic cognitive retraining, particularly for girls with neurobiological markers of elevated error-sensitivity.

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

Read