288

Papers

10/10

Roadblocks Active

2

Connections

In 222 university students, using AI tools for task-oriented purposes (writing, research) was linked to better grades and larger prefrontal and hippocampal brain regions, while using AI for emotional companionship was associated with depression, social anxiety, and reduced volume in brain areas that process social cues and emotion. The same apps produce opposite neural and psychological signatures depending on why you use them. This is practically important because it suggests mental health risk from AI use is not about screen time but about whether AI is substituting for human social connection.

██████████ 0.9 youth-mental-health-crisis Preprint

ReadSaveConnections

A three-stage AI pipeline uses a large language model to generate plain-language clinical summaries at each step — screening, severity rating, and continuous scoring — which then guide the fusion of audio, video, and text signals from patient interviews to detect depression. The system outperforms prior benchmarks on two datasets and, crucially, produces human-readable justifications for each prediction rather than a black-box score. Interpretability is the key advance here: clinicians can inspect why the system flagged a patient, which is a prerequisite for real-world adoption in any clinical workflow.

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

ReadSaveConnections

By explicitly modeling how structural brain anatomy and functional brain activity inform each other — rather than simply concatenating them — this framework detects major depressive disorder from MRI scans with 84.7% accuracy and 86.4% sensitivity on a large public dataset. The bidirectional cross-attention mechanism lets each imaging modality correct and sharpen the signal from the other, which mirrors how clinicians integrate different types of scan information. Higher sensitivity (86%) than specificity (83%) is clinically meaningful: the system misses fewer true cases, which matters most in screening contexts.

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

ReadSaveConnections

A sparse Bayesian model that combines EEG, eye-tracking, and facial muscle data during implicit association tests achieved AUCs of 0.73–0.79 for predicting mental health symptoms, vastly outperforming the standard D-score method (AUC 0.50–0.53) that clinicians currently use. The key insight is that the traditional single-number D-score throws away most of the multimodal signal; the Bayesian approach integrates it probabilistically. Sample sizes are small (n=34–39), so this is proof-of-concept rather than clinical-ready, but the performance gap over existing tools is large enough to warrant follow-up.

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

ReadSaveConnections

AI models trained on three widely used depression interview datasets (ANDROIDS, DAIC-WOZ, E-DAIC) can classify depressed versus control participants at high accuracy using only the interviewer's scripted questions — not what patients say at all. This means that fixed interviewer prompts carry enough structural information to distinguish patient groups, and any model benchmarked on these datasets may be measuring interview design artifacts rather than genuine depression signals. The finding is a red flag for the entire subfield of automated depression detection from clinical interviews: previously reported accuracy numbers may be substantially inflated.

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

ReadSaveConnections

Using EEG recorded while people discriminated happy from sad faces, sad emotional processing produced significantly more stable (lower energy) brain network states in delta, theta, and alpha frequency bands, with the strongest effect in alpha (Cohen's d = 0.83, a large effect size). Networks with higher global efficiency — meaning information travels more efficiently across the brain — corresponded to these more stable sad-processing states. This provides an objective, session-level neural measure of emotional processing that could eventually serve as a biomarker for disorders characterized by negative affect bias, such as depression.

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

ReadSaveConnections

This framework separates the voice features that signal depression from those that reveal personal identity, then encrypts the identity component before any analysis occurs — enabling privacy-preserving automated depression screening from audio. The adversarial training approach forces the depression detector to learn only clinically relevant acoustic patterns, tested across three benchmark datasets (DAIC-WOZ, D-Vlog, MODMA). Stripping identity signals from the feature space also has a secondary benefit: it could reduce demographic bias in depression detection models that currently perform unevenly across age, sex, and accent groups.

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

ReadSaveConnections

SynSym uses large language models to generate realistic patient expressions of psychiatric symptoms — in both clinical and everyday language — by first expanding each symptom into clinically grounded sub-concepts, then composing realistic multi-symptom descriptions. Models trained only on this synthetic data matched the performance of models trained on real patient data across three depression symptom benchmarks, and improved further when fine-tuned on real data afterward. The practical payoff is that researchers can now train and test depression-related NLP systems without needing access to sensitive patient records, which is a persistent bottleneck for digital therapeutic development.

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

ReadSaveConnections

DRIFT is a statistical method that estimates which treatment will work best for a specific patient with depression by first learning latent symptom dimensions from questionnaire items, then finding the treatment assignment that performs best even in the worst-case symptom domain — guarding against the risk of optimizing for measured symptoms while ignoring unmeasured ones. Validated on the EMBARC randomized trial of sertraline versus placebo for major depressive disorder, DRIFT offers a principled way to handle the heterogeneity that makes one-size-fits-all treatment guidelines a poor fit for many patients. The closed-form solution makes it computationally accessible without requiring specialized optimization software.

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

ReadSaveConnections

Deep learning sleep-staging models trained on healthy people generalize poorly to patients with ischemic stroke — and visualization tools reveal the models are paying attention to physiologically uninformative EEG regions in patient data, meaning they are pattern-matching noise rather than real sleep signals. Sleep disruption is both a core symptom and a causal contributor to depression and anxiety, so this generalization failure matters clinically: AI tools marketed for mental health monitoring cannot be assumed to work in people with neurological or psychiatric comorbidities. The paper highlights that benchmark accuracy in healthy samples is a misleading proxy for real-world clinical performance.

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

ReadSaveConnections