286

Papers

10/10

Roadblocks Active

0

Connections

This paper trains a generative model exclusively on healthy brain signals to learn what 'normal' EEG looks like, then uses the degree of deviation from that norm as an explicit input feature for a depression classifier — rather than trying to augment scarce patient data. The approach sidesteps the chronic shortage of labeled clinical EEG by repurposing healthy controls as the reference anchor. It also introduces a spatial adaptation module to reconcile differences in electrode placement across recording sites, which is a practical barrier to combining multi-center datasets.

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

ReadSaveConnections

In a small pilot RCT (n=10 randomized veterans), combining a smartwatch-based PTSD hyperarousal detector with personalized digital self-management tools produced more stable symptom trajectories in anxiety, depression, and PTSD severity than physical activity alone. The study is important because it attempts a real-world test of closed-loop sensing — where the device detects a physiological warning sign and immediately prompts a coping intervention — rather than passive monitoring. The sample is too small to draw firm conclusions, but the registered trial design and use of validated measures (GAD-7, PHQ-8, PCL-5) set a reasonable template for scaling.

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

ReadSaveConnections

Most digital therapy optimization algorithms assume patient adherence is a fixed backdrop; this paper argues it is itself shaped by prior treatment decisions and models it as a linear dynamical system. The UCB-BOLD algorithm derived from this model is shown in synthetic experiments to outperform standard reinforcement learning approaches at selecting which intervention to recommend next. The implication is that ignoring the feedback loop between recommendation and engagement systematically underestimates how much personalization matters for long-term digital therapy outcomes.

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

ReadSaveConnections

Analyzing voice recordings for subtle acoustic properties — pitch irregularity, pause patterns, emotional tone — and linguistic patterns such as vocabulary diversity and sentence complexity yields consistent associations with clinically measured depression, anxiety, and ADHD severity across five separate datasets. The study uses explainable ML (SHAP) to identify which specific voice features drive predictions, which matters for clinical credibility. Confidence is limited because sample sizes and train/test split details are not reported, but the multi-dataset consistency is a stronger signal than single-cohort studies.

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

ReadSaveConnections

Using a single mathematical property of brain signals — how self-similar the EEG waveform is across different time scales, quantified via the Hurst exponent — a simple probabilistic classifier can identify deep sleep (N3) with 87% balanced accuracy in a large public dataset of older women. The finding that linear classifiers fail badly (57% and 51%) while probabilistic ones succeed reveals the underlying feature space is nonlinearly structured, which has design implications for sleep-monitoring wearables. A reliable passive deep-sleep detector could enable closed-loop neurofeedback without requiring users to actively engage with the device.

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

ReadSaveConnections

Brain signals during non-REM sleep carry detectable signatures of future cognitive decline: in a five-year longitudinal study of 290 older women, those who later developed dementia had measurably less 'critical' EEG dynamics at baseline compared to those who remained cognitively healthy. The method — multifractal analysis of 30-second sleep EEG epochs — uses a publicly accessible dataset and is non-invasive, making it a plausible screening tool. The limitation is that no classifier accuracy metrics are reported, so the magnitude of predictive value remains unclear from this paper alone.

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

ReadSaveConnections

When five different 'explanation' methods are applied to the same deep learning EEG depression classifier, they converge on highlighting frontal and right-hemisphere brain regions — consistent with existing neuroscience theory about emotional regulation. This convergence across methods (gradient-based, perturbation-based, and Shapley-value approaches) strengthens confidence that these regions are genuinely informative rather than artifacts of any single technique. For the field, this provides a practical argument for using explainability as a model-validation tool rather than just a post-hoc communication aid.

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

ReadSaveConnections

Voice-based depression screening inadvertently leaks demographic information: a model trained to detect depression from speech can also infer a speaker's gender with 93% accuracy as a side effect. InfoShield uses an information-theoretic training objective to mathematically suppress this leakage, reducing gender inference to near chance (56%) while preserving depression detection performance. This is a direct response to regulatory and ethical pressure on mental health AI — showing that privacy protection and clinical utility are not necessarily in conflict, at least on a single benchmark dataset.

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

ReadSaveConnections

Rather than fine-tuning a language model on clinical data — which requires expensive labeled datasets — Dep-LLM prompts frozen general-purpose LLMs to decompose clinical interview transcripts into five structured symptom themes, then weights each theme's contribution by how confident the model's token-level outputs appear. Across 21 different open-source LLMs and two established benchmark datasets, this zero-training approach outperforms supervised depression-specific models. The key caveat is that no code or prompt templates are released, making independent verification currently impossible.

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

ReadSaveConnections

Mental-R1 trains a language model to reason about mental health assessments using reinforcement learning with a reward structure that mimics how clinicians progressively narrow diagnostic uncertainty — starting broad and becoming more confident as evidence accumulates. The resulting model improves weighted F1-score by 10.4 percentage points over the best existing RL baseline across eight mental health classification datasets. The method is notable for explicitly modeling the uncertainty structure of clinical reasoning rather than treating diagnosis as a simple pattern-matching task.

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

ReadSaveConnections