283

Papers

10/10

Roadblocks Active

0

Connections

Instead of training a classifier directly on labeled patient EEG, this method first learns what healthy brain activity looks like using only control data, then uses how far a new recording deviates from that norm as an additional input signal for depression detection. This normative approach sidesteps the chronic problem of small, imbalanced clinical EEG datasets and does not need data augmentation. A channel-adaptation module also handles the practical headache of combining EEG recordings from different hospitals that use different electrode layouts.

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

ReadSaveConnections

Most digital therapeutic algorithms treat patient engagement as a fixed background condition, but this paper proves mathematically that engagement is itself shaped by the treatment recommendations being made — a feedback loop that existing systems ignore. The authors propose UCB-BOLD, an online learning algorithm that adapts treatment selection as it learns how a particular patient's capacity to engage changes over time, with proven sublinear regret guarantees. The approach is validated against synthetic cohorts calibrated on a real micro-randomized trial, making it more credible than purely theoretical proposals.

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

ReadSaveConnections

TimeSRL converts raw smartphone and wearable sensor streams into natural-language behavioral summaries, then uses reinforcement learning to fine-tune a language model to predict anxiety and depression scores from those summaries alone. The key advance is a cross-dataset evaluation (leave-one-dataset-out) that tests whether the approach actually generalizes, rather than just fitting one study population — it shows 3–10% error reduction over standard machine learning baselines and 9–44% over other LLM approaches for anxiety prediction. Converting sensor data to language before modeling appears to create representations that transfer better across populations, which matters enormously for clinical deployment.

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

ReadSaveConnections

Cancer survivors experience high rates of depression and anxiety but are least likely to log their emotional state exactly when they most need support — the 'diary paradox.' PULSE addresses this by having an AI agent autonomously investigate passive smartphone data (movement, sleep, screen use, social communication) to predict when a survivor wants emotional support, achieving 74.3% balanced accuracy without requiring any self-report. The agentic approach, where the system decides what data to query and how to reason about it, outperforms a single-step structured query, suggesting that autonomous investigation is a meaningful design choice for passive mental health monitoring.

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

ReadSaveConnections

Across 555 semi-structured interviews with SCID-derived diagnostic labels, five LLMs show accuracy ranging from 0.49 to 0.86 on psychiatric screening tasks — better than chance in some settings, but unreliable overall. The most clinically important finding is a specific failure mode: LLMs systematically miss anxiety and PTSD cases where the patient shows preserved functioning or coping ability, even when explicit symptom evidence is present in the text. This mirrors a bias found in human clinicians, suggesting LLMs have absorbed a problematic heuristic — 'functioning well means not ill' — from their training data.

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

ReadSaveConnections

Five different methods for explaining what an EEG depression classifier is actually using were applied to the same model, and they largely agreed: frontal and right-hemisphere regions consistently received high attribution across gradient-based and perturbation-based approaches. This convergence across explanation methods is meaningful because any single explainability tool can produce artifacts of its own design, so agreement across five independent methods adds confidence that the identified brain regions are genuinely informative rather than algorithmic artifacts. The findings align with established neuroscience on hemispheric asymmetry in depression, lending biological plausibility.

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

ReadSaveConnections

Specific vocal irregularities — shimmer (amplitude variation) and jitter (pitch instability) — show consistent associations with symptom severity across multiple independent depression, anxiety, and ADHD datasets, suggesting these signals are stable enough to be clinically meaningful rather than dataset-specific noise. The study uses SHAP values to show which features drive predictions, making the model's reasoning auditable by clinicians rather than a black box. The cross-dataset consistency is the main contribution: most speech biomarker work validates only on a single cohort, so replication across five datasets substantially raises confidence.

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

ReadSaveConnections

This paper describes a fully open-source wearable sleep monitor built from commodity electronics and 3D-printed parts for under €40, capable of classifying wake, light sleep, deep sleep, and REM with around 77% accuracy. Sleep disruption is a transdiagnostic risk factor and symptom marker across depression, anxiety, PTSD, and bipolar disorder, but clinical-grade sleep measurement requires expensive polysomnography. Making hardware designs and firmware openly available removes a major barrier to sleep research in low-resource settings, though the single-participant, 15-night proof-of-concept validation is too limited to confirm clinical reliability.

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

ReadSaveConnections

Across all 12 dataset pairs tested, EEG foundation models encoded individual subject identity 13 to 89 times more strongly than the diagnostic label they were supposedly learning — and fine-tuning made this worse, not better. This means a model that looks like it is detecting depression may actually be detecting that it has seen this particular person's brain recordings before, a form of identity leakage that inflates accuracy metrics without providing real diagnostic signal. The authors show that mathematically removing the identity axis from the learned representations improves label decoding by 6–27 percentage points, pointing to a concrete remediation path.

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

ReadSaveConnections

Speech-based depression screening systems inadvertently encode sensitive demographic information — gender can be inferred from the learned representations with 92.6% accuracy. InfoShield uses information-theoretic optimization to strip this demographic signal, reducing gender inference to 55.5% (near chance) while only losing 6% of diagnostic utility for depression classification. The key technical insight is that standard mutual information estimators fail on speech data because they cannot align sequential audio frames with static demographic labels, a problem the authors solve with a cross-modal attention mechanism.

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

ReadSaveConnections