174

Papers

10/10

Roadblocks Active

3

Connections

A multi-agent AI system autonomously analyzed wearable sensor data from over 9,000 participant-observations and identified 41 candidate digital biomarkers for mental health, including sleep duration variability and sleep onset variability as cross-validated markers of depression severity in two independent cohorts. The system combines hypothesis generation, statistical testing, adversarial validation, and literature-grounded reasoning, reducing researcher time while maintaining scientific rigor. This matters because it demonstrates that passive, continuous signals from everyday devices can reliably flag depression risk — a critical step toward scalable, low-burden mental health screening.

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

ReadSaveConnections

A case report of deep TMS targeting the left prefrontal cortex in an elderly patient with treatment-resistant depression and dissociative identity disorder found that depressive symptoms improved substantially (severity score dropped from 24 to 11-12) while PTSD symptoms only modestly improved and dissociative symptoms actually worsened. This dissociation of symptom responses is clinically important: it shows that TMS can successfully target depression circuitry even in complex comorbid presentations, but also signals that dissociation may be a distinct and fragile dimension requiring separate management. The case highlights the risk of under-assessing dissociative disorders before initiating brain stimulation protocols.

██████████ 0.9 treatment-resistant-depression Peer-reviewed

Read

This narrative review synthesizes longitudinal and experimental evidence showing that adolescent depression, anxiety, self-harm, and suicidality have risen sharply across high-income countries since the early 2010s, tracking closely with mass smartphone and social media adoption. It identifies specific mechanisms — social comparison, cyberbullying, sleep disruption, and algorithmic amplification of harmful content — that link platform use to deteriorating mental health. While the review is brief and lacks systematic search methodology, it consolidates the current public health consensus and provides a framework for identifying intervention targets.

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

Read

This paper proposes a framework that simultaneously analyzes structural MRI (brain anatomy) and resting-state fMRI (brain connectivity patterns) using bidirectional attention mechanisms, achieving 84.7% accuracy in classifying patients with major depressive disorder from healthy controls in a large multi-site dataset. The key insight is that letting each imaging modality inform the other's feature extraction — rather than simply combining them after the fact — captures complementary information that improves detection. Higher diagnostic accuracy from non-invasive neuroimaging reduces reliance on subjective clinical interviews, which is particularly valuable for treatment planning and monitoring.

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

ReadSaveConnections

This paper uses large language models to generate progressively detailed clinical summaries from a patient's audio, video, and text signals, then feeds those summaries back into a detection pipeline that classifies depression severity across three levels — from binary screening to fine-grained continuous scoring. The system outperforms current state-of-the-art baselines and, crucially, produces human-readable rationales for each prediction, addressing the 'black box' problem that limits clinician trust in AI tools. Interpretability is not a luxury here: regulators increasingly require explainability for AI used in clinical decision support.

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

ReadSaveConnections

In a cross-sectional MRI study of 222 university students, higher functional AI use (e.g., for academic tasks) correlated with larger gray matter volume in prefrontal and visual regions and better academic performance, while the relationship with mental health outcomes differed by use type. The neuroanatomical differences suggest that how young people use AI — not merely how much — shapes brain structure and psychological outcomes, a distinction that gets lost in blanket debates about screen time. This is an early and correlational finding, but it raises an important question for digital therapeutics: could AI-delivered mental health tools designed around functional rather than socio-emotional interaction patterns produce better outcomes?

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

ReadSaveConnections

Using four years of follow-up data from nearly 12,000 children in the ABCD Study, this paper shows that impulsivity, sleep disturbance, and reduced parental monitoring dynamically predict when a child first tries substances, with genetic risk for nicotine dependence also accelerating initiation. The use of time-varying statistical models means the analysis captures risk as it changes over childhood — not just as a static snapshot — which is closer to how real prevention windows work. Identifying these modifiable environmental levers (sleep, parental monitoring) alongside genetic risk creates a roadmap for early, targeted prevention in digital and school-based programs.

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

ReadSaveConnections

This conceptual survey argues that current brain modeling efforts — from neural simulations to neuromorphic chips — are fragmented because they lack a shared definition of what it means for a model to actually run in time alongside a real brain. The author introduces a taxonomy of 'execution regimes' ranging from offline static models to systems that co-simulate with live biological data, providing a vocabulary for comparing otherwise incommensurable approaches. For computational psychiatry, this framework matters because it clarifies which kinds of brain models are close enough to real-time dynamics to be useful for intervention targeting versus which are purely descriptive.

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

ReadSaveConnections

This paper introduces a continual learning system for diagnosing brain disorders from fMRI that can incorporate data from new clinical sites without forgetting what it learned from previous sites — a problem called 'catastrophic forgetting' that makes most AI models impractical to deploy across hospitals. The system uses a generative model to synthesize realistic brain connectivity patterns from past sites and replays them during new training, effectively giving the model a synthetic memory. For depression diagnosis specifically, multi-site deployment is critical because brain scanning parameters vary significantly across institutions, and a model that degrades when moved between sites cannot be used in real clinical workflows.

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

ReadSaveConnections

This PhD thesis-derived work finds that machine learning classifiers trained on EEG signals recorded during an active cognitive task (stop-signal) outperform resting-state EEG in classifying adults with ADHD, with gamma-band activity over frontal and parietal regions being most diagnostic. It also links white matter tract integrity to computationally modeled differences in how individuals weigh effort against reward — a core dimension of depression and motivational disorders. The implication is that task-evoked brain signals, not just passive resting-state recordings, carry more clinically actionable information for conditions defined by impaired motivation and control.

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

ReadSaveConnections