280

Papers

10/10

Roadblocks Active

0

Connections

This paper releases a carefully stratified dataset of 207 adolescents aged 12–17 divided into three groups: those in acute suicidal crisis, those receiving non-suicidal psychiatric care, and healthy controls, all assessed with seven validated instruments covering suicidal ideation, depression, bullying, and problematic internet use. Having a public, labelled dataset that includes a genuine acute-crisis group is rare — most research either lacks clinical severity stratification or cannot be shared due to privacy constraints. This resource directly enables the development and benchmarking of predictive models for adolescent suicide risk, a problem where training data scarcity has long been the primary bottleneck.

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

Read

A companion Zenodo release of the same VisIA-Q dataset, providing item-level responses alongside aggregated scores from all seven instruments plus sociodemographic variables for the same 207-adolescent cohort. Item-level access matters because it allows researchers to examine which specific questions or response patterns best discriminate between risk groups, enabling more granular feature engineering for clinical decision tools. The dual-version release suggests the authors are actively maintaining and versioning the resource for community use.

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

Read

This paper formalizes a problem that undermines most digital therapeutic trials: patient adherence is not static but changes over time depending on what interventions were previously recommended and whether the patient followed them. The authors model this as a dynamical system and derive an algorithm (UCB-BOLD) that learns optimal treatment sequences while accounting for this feedback loop, achieving provably sublinear regret. The practical implication is that recommendation systems for apps like mood trackers or CBT tools should factor in a user's engagement history, not just their current symptoms, to avoid burning out compliance.

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

ReadSaveConnections

TimeSRL converts raw smartphone sensor streams (movement, phone use patterns, etc.) into natural language descriptions first, then uses a large language model trained with reinforcement learning to predict anxiety and depression scores — and tests this across multiple independent study cohorts it was never trained on. The key finding is that the semantic translation step dramatically improves how well the model generalises to new populations, reducing anxiety prediction error by up to 44% over other LLM approaches. This matters because most mental health AI tools fail in deployment when the user population differs from the training sample; this architecture is specifically designed to close that gap.

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

ReadSaveConnections

EmoTrack extracts structured clinical cues from therapy session transcripts using a large language model, combines them with sentence-level meaning representations, and predicts standardized depression severity scores (PHQ-8), achieving a 13.5% error reduction over the best existing single-session baseline on the widely-used DAIC-WOZ dataset. The model also handles multi-session tracking by compressing memory of past sessions through an attention mechanism — addressing a real gap since most NLP depression tools treat each session in isolation. The reliance on a synthetic multi-session evaluation dataset is a limitation that will need independent validation.

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

ReadSaveConnections

Among cancer survivors — who experience elevated depression and anxiety — people rarely fill in mental health diaries precisely when they are struggling most (the 'diary paradox'). PULSE uses a smartphone-sensing system where an AI agent actively queries different data streams (location, activity, phone usage) rather than following a fixed analysis script, achieving 74% balanced accuracy in detecting when a survivor wants emotional support. The agentic approach consistently outperformed structured pipelines, suggesting that flexible, hypothesis-driven data interrogation is better suited to the irregular and context-dependent nature of mental health signals than predefined feature extraction.

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

ReadSaveConnections

This study tests whether voice characteristics — small pitch irregularities, speaking rhythm, word choice complexity — can reliably signal depression, anxiety, and ADHD severity across five different datasets using open-source tools. It finds stable relationships between vocal irregularities (shimmer, jitter) and symptom severity that replicate across datasets, and uses explainability methods (SHAP, LIME) to show which features drive predictions. The multi-dataset approach is a genuine strength because single-dataset voice biomarker studies rarely generalise; the main limitation is that one dataset is proprietary clinical data that cannot be independently accessed.

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

ReadSaveConnections

Deep learning models can classify depression from brainwave (EEG) recordings with reasonable accuracy, but clinicians cannot use them if they don't know which brain signals drove the decision. This paper applies four different explanation methods to the same EEG classifier and asks whether they agree — finding that gradient-based and perturbation-based methods converge on similar brain regions (frontal, temporal, right hemisphere), which provides tentative validation that the model is picking up on neurobiologically plausible patterns rather than artifacts. Convergence across explanation methods is important because any single method can be misleading; agreement between them increases trust in the identified features.

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

ReadSaveConnections

This paper describes a fully open-source wearable headband for home sleep monitoring built from off-the-shelf components for under €40, achieving four-stage sleep classification (wake, light, deep, REM) at 77.6% accuracy — comparable to consumer devices costing hundreds of euros. Sleep disruption is both a symptom and a cause of depression, anxiety, and PTSD, and most objective sleep monitoring requires either expensive clinical equipment or costly proprietary devices that create access barriers for research. The hardware designs, firmware, and software are all openly published, meaning any researcher can build and modify the device; the key current limitation is that it has only been validated in one person over 15 nights.

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

ReadSaveConnections

A dominant theory of psychedelic-assisted therapy holds that these drugs work by increasing brain entropy (randomness of neural activity), and that more entropy equals richer experience. This paper challenges that by showing that deep meditation produces similarly elevated brain entropy to psychedelics like 5-MeO-DMT, yet meditation produces minimal, content-free awareness rather than vivid hallucinations — meaning entropy alone cannot explain what psychedelics do to the mind. The authors propose that brain complexity (structured variation across regions) better explains the richness of psychedelic experience, which has implications for understanding how and why psychedelics might treat depression and PTSD rather than just that they do.

██████████ 0.7 psychedelic-mechanisms Preprint

ReadSaveConnections