249

Papers

10/10

Roadblocks Active

0

Connections

This paper combines structural brain scans (sMRI) and functional connectivity data (rs-fMRI) using a bidirectional attention mechanism that lets each data type inform the interpretation of the other, rather than simply concatenating features. Tested on the large public REST-meta-MDD dataset with 10-fold cross-validation, it outperforms standard fusion methods. The importance is that MDD has no reliable single biomarker, and approaches that jointly model brain structure and function across many patients move us closer to an objective diagnostic signal.

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

ReadSaveConnections

Rather than treating depression detection as a black-box classification problem, this framework uses a large language model to generate step-by-step clinical summaries — first screening for binary depression presence, then estimating severity on a five-class scale, then producing continuous PHQ-8 scores — with each stage's narrative guiding the next. On the E-DAIC benchmark the approach improves over prior state-of-the-art while producing human-readable reports. Making AI reasoning legible to clinicians is a prerequisite for real-world deployment of any automated screening tool.

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

ReadSaveConnections

EviDep adds a principled uncertainty layer — using a Normal-Inverse-Gamma distribution — to multimodal depression severity estimation, so the model can flag when it is unsure rather than producing overconfident predictions. A wavelet-based module separates slow and fast behavioral signals before fusion, reducing redundancy across audio-visual inputs. Uncertainty quantification is clinically essential: a system that knows when to defer to a human clinician is far safer than one that does not.

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

ReadSaveConnections

DRIFT addresses a persistent problem in psychiatric treatment research: clinical trials measure many overlapping symptoms, but it is unclear which composite outcome to optimize when personalizing treatment. The method uses adversarial learning to find individualized treatment recommendations that perform well across all plausible ways of weighting symptom domains, validated on the EMBARC sertraline-vs-placebo RCT for major depressive disorder. This matters because treatment heterogeneity in depression is large, and robust personalization could rescue subgroups that average-effect analyses miss.

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

ReadSaveConnections

This paper generates 28,800 synthetic psychiatric profiles using four leading LLMs and compares their statistical distributions to real epidemiological databases (NHANES, NESARC-III). It finds that LLMs produce individually plausible cases but compress variance by 14–62%, erasing the distributional tails where clinical risk concentrates — and 37% of simulated cases flip diagnostic threshold between two identical prompts. If LLMs are used to train or test mental health AI tools, this systematic distortion could embed invisible biases into those systems.

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

ReadSaveConnections

Using task-based EEG during a stop-signal paradigm (rather than resting-state EEG), this work achieves better ADHD classification and identifies gamma-band activity over frontal-parietal regions as the key predictive signal. A separate analysis links white matter tract integrity — particularly tracts connected to the supplementary motor area — to individual differences in effort and reward sensitivity derived from computational behavioral models. This is relevant because it shows that biologically grounded computational models of motivation can connect brain structure to clinically meaningful variation in psychiatric populations.

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

ReadSaveConnections

Tracking nearly 12,000 children over four years in the ABCD Study, this paper uses survival analysis to identify when and why youth first try alcohol, nicotine, or cannabis. Impulsivity, low parental monitoring, and polygenic risk for nicotine use disorder are the strongest predictors of earlier initiation, with genetic effects strongest for nicotine. Because early substance initiation is one of the most robust predictors of later mental health disorders, identifying modifiable risk factors at scale offers actionable prevention targets.

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

ReadSaveConnections

This conceptual paper argues that the field of brain digital twins is fragmented because researchers use incompatible definitions of what it means to 'run' a brain model — differences in how time, causality, and updating are handled prevent meaningful comparison across systems. It proposes a taxonomy based on execution semantics — how a model steps forward in time and responds to inputs — as a unifying lens. While purely theoretical, this kind of framework work is necessary groundwork before computational psychiatry can build reusable, comparable brain simulations.

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

ReadSaveConnections

FORGE solves a practical deployment problem: when a brain disorder classifier is trained at one hospital site and then updated with data from a new site, it tends to forget what it learned before (catastrophic forgetting). The paper uses a generative model to synthesize realistic functional connectivity matrices from prior sites, replaying them during training on new data without requiring access to original patient records — a privacy-preserving solution. This matters for scaling psychiatric neuroimaging AI across clinical networks where data sharing is legally restricted.

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

ReadSaveConnections

In 222 university students, those who use AI tools for functional tasks (writing, coding, analysis) showed larger prefrontal and hippocampal gray matter volume and more efficient hippocampal network organization, along with better grades; those using AI for social-emotional support showed the opposite pattern. This cross-sectional neuroimaging study cannot establish causation, but it raises a concrete and testable concern: heavy reliance on AI for emotional support may correlate with, or contribute to, adverse brain and mental health outcomes in young people. The findings give neurobiological grounding to a policy-relevant question about how youth AI use should be designed.

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

ReadSaveConnections