273

Papers

9/9

Roadblocks Active

0

Connections

The study finds that vocal irregularities like pitch instability (jitter, shimmer) and patterns in word choice and sentence structure reliably track symptom severity for depression, anxiety, and ADHD across five separate datasets. Crucially, it uses interpretable models with SHAP explanations so clinicians can see which specific speech cues drove each prediction — not just a black-box score. This matters because clinician-interpretable speech biomarkers could support screening in settings where trained professionals are scarce, as long as the proprietary dataset limitation is addressed.

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

ReadSaveConnections

TimeSRL converts raw passive-sensing data (phone usage, accelerometer, etc.) into natural-language behavioral descriptions, then uses reinforcement learning to train a language model to predict anxiety and depression scores from those descriptions rather than raw numbers. Testing the model on datasets it was never trained on — a stringent real-world simulation — showed 3–44% error reduction over prior approaches. The semantic bottleneck is the key mechanism: by forcing the model to reason in human-readable behavioral terms, it learns representations that transfer across different populations and sensing setups.

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

ReadSaveConnections

This paper tackles a core practical barrier in EEG-based depression screening: small clinical samples and incompatible electrode configurations across hospitals. Instead of fabricating extra training data, it trains a generative model only on healthy brain signals, then measures how abnormal a patient's EEG looks relative to that healthy baseline — a 'pathological prior' score that supplements the classifier. A spatial adaptation module also handles the problem of different hospitals using different electrode placements, enabling multi-center data pooling that was previously infeasible.

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

ReadSaveConnections

Across all 12 tested combinations of EEG foundation models and datasets, frozen model representations encoded individual subject identity 13 to 89 times more strongly than actual clinical labels — and fine-tuning made this worse, not better. When a linear erasure step removed the identity axis before classification, label decoding improved by 6–27 percentage points. This is a structural validity problem: models may appear to detect depression or other conditions when they are actually recognizing individual people, which would make them useless for any new patient.

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

ReadSaveConnections

InfoShield strips demographic information (gender, age) from speech features used for depression screening, reducing gender inference from 92.6% to near-chance (55.5%) while maintaining depression classification performance. The key technical contribution is a new estimator called TimeAwareMINE that handles the mismatch between sequential speech signals and static demographic attributes — a problem that caused standard methods to fail. Privacy-preserving speech analysis is a prerequisite for deployment of speech-based mental health tools in regulated healthcare settings.

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

ReadSaveConnections

EmoTrack combines LLM-extracted clinical signals with sentence-level semantic embeddings to predict PHQ-8 depression scores from therapy transcripts, achieving 13.5% lower error than the best single-session baseline. An attention-based memory module lets the system also track how symptoms change across multiple therapy sessions — something most prior work ignored. The ability to monitor depression trajectories rather than just snapshots is directly relevant to measuring whether therapy is working.

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

ReadSaveConnections

Most algorithms for personalizing digital therapeutics treat patient engagement as a fixed context variable, but this paper argues adherence is endogenous — past behavior shapes future engagement capacity in a dynamic feedback loop. It models this with a linear dynamical system and proves that the UCB-BOLD algorithm achieves provably sublinear regret even under this more realistic setup. The practical implication is that treatment recommendation systems should account for engagement momentum rather than treating each session as independent, particularly for patients prone to dropout.

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

ReadSaveConnections

Testing five LLMs on 555 structured clinical interview transcripts, accuracy ranged from near-chance (0.49) to 0.86 depending on the disorder and model, with Matthews correlation coefficients — a more stringent metric — topping out at only 0.38. False negatives in anxiety and PTSD screening often occurred not because symptoms were absent but because LLMs weighted protective factors (coping ability, social support) too heavily, mirroring a known clinician bias. This tells us current LLMs cannot be used as standalone screeners but may have value as decision support if their evidence-weighting tendencies are corrected.

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

ReadSaveConnections

Dep-LLM improves LLM-based depression diagnosis without any model retraining by decomposing clinical interview text into five clinically-aligned themes and using token-level probability entropy to flag uncertain reasoning before aggregating evidence. Tested across 21 language models, it consistently outperforms zero-shot prompting on nine classification metrics. The training-free approach is significant for practical deployment: clinical teams could apply it to existing frozen models without the cost or safety concerns of fine-tuning on sensitive patient data.

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

ReadSaveConnections

Applying text analysis to a large Reddit dataset of self-reported mental health conditions, the study finds that cognitive distortions — systematic errors in thinking like catastrophizing or mind-reading — are more prevalent across all nine mental health groups compared to controls, but the patterns of distortions look surprisingly similar across different diagnoses. This challenges the assumption that distinct disorders have distinct cognitive profiles, at least as measured in online self-disclosure text. The finding is relevant for transdiagnostic therapy approaches that target distortions across multiple conditions simultaneously.

██████████ 0.7 computational-psychiatry Preprint

ReadSaveConnections