273

Papers

9/9

Roadblocks Active

0

Connections

This paper tackles a fundamental problem in digital therapeutics: patients who disengage from an app affect which treatments get recommended to them, creating a feedback loop that naive algorithms ignore. The authors model this using linear dynamical systems with a logit adherence link and propose an online learning algorithm (UCB-BOLD) that provably identifies the best treatment policy in finite time even under this feedback. This matters because most digital health trials assume adherence is independent of treatment history — an assumption this work formally shows is wrong and provides tools to fix.

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

ReadSaveConnections

This is one of the few registered randomized trials (NCT06993012) testing a wearable-plus-app digital intervention for PTSD symptoms in veterans. Veterans using the smartwatch-based hyperarousal detection system maintained symptom improvements after an endurance cycling event, while the cycling-only group showed late-study escalation. The trial is very small (n=13 randomized) so results are hypothesis-generating, but the continuous physiological sensing and nonlinear trajectory modeling via GAMMs represent a methodological step up from typical self-report-only pilots.

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

ReadSaveConnections

The paper tests whether vocal properties — roughness (shimmer, jitter), pitch patterns, and lexical complexity — reliably track depression, anxiety, ADHD, and stress severity across five separate datasets including a real-world clinical sample. Using XGBoost with SHAP explainability, the authors find stable associations between vocal irregularities and symptom scores that hold across different recording contexts. The cross-dataset consistency is what elevates this above typical single-cohort speech studies, though the proprietary clinical dataset cannot be shared, limiting full replication.

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

ReadSaveConnections

EEG-based depression classifiers are increasingly accurate but act as black boxes, which is a serious barrier to clinical adoption. This paper applies five different explanation methods to the same InceptionTime classifier and asks whether they agree on which brain regions drive the predictions. Encouragingly, gradient-based and perturbation-based methods converged on frontal and right-hemisphere temporal regions — consistent with known depression neuroscience — which strengthens confidence that the model is capturing real signal rather than noise artifacts.

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

ReadSaveConnections

Dep-LLM structures clinical interview transcripts into five themes aligned with standard diagnostic criteria, then uses token-level uncertainty (entropy) from a frozen language model to weight how much to trust each piece of reasoning before making a final depression call — all without any additional training. Tested across 21 open-source language models on the DAIC-WOZ and E-DAIC datasets, it consistently outperforms zero-shot baselines. The practical value is that any clinic could apply this to existing interview recordings using publicly available models, though the CoT prompts and entropy thresholds are not yet fully disclosed.

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

ReadSaveConnections

This pilot study simultaneously records brainwave (EEG) and blood-oxygenation (fNIRS) signals from participants doing a memory task with emotional content, then applies an end-to-end deep learning classifier to detect depressive state. Using both signal types together is important because EEG captures electrical timing and fNIRS captures slower metabolic changes — they carry complementary information. The study is very small (11 healthy students) and lacks a clinical depression group, so the results are proof-of-concept, but the combined modality approach is worth watching as hardware costs fall.

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

ReadSaveConnections

Most EEG depression classifiers rely on data augmentation to compensate for small clinical datasets, which risks amplifying noise. This paper instead trains an unsupervised model exclusively on healthy controls to learn what normal brain activity looks like, then flags depression by measuring how far a new recording deviates from that normal pattern. A secondary module also handles the practical problem of different hospitals using different numbers of EEG electrodes. Tested on two public datasets, the approach avoids the synthetic-data pitfalls of augmentation, though sample sizes remain modest.

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

ReadSaveConnections

Rather than collecting a large PTSD dataset from scratch, this paper borrows a model trained on fear responses to spiders and transfers it to estimate PTSD severity from heart rate and skin conductance in military veterans during a simulation. The logic is that both spider phobia and PTSD involve similar physiological fear responses. Achieving 86% classification accuracy and a mean error of 5.6 points on the clinical severity scale from just two cheap sensors is noteworthy, though the 21-person military sample and unavailability of the dataset limit how far these numbers can be trusted.

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

ReadSaveConnections

Mental-R1 trains a language model to reason about mental health the way a clinician does — moving from uncertain initial impressions toward a confident diagnosis — by using a reinforcement learning reward that penalizes overconfidence early in reasoning and rewards it later. Across eight open-source mental health classification datasets, this staged uncertainty approach improved weighted F1 by 10.4 percentage points over standard reinforcement learning baselines. The key innovation is that the reasoning process itself is shaped to match human cognitive patterns in clinical assessment, not just the final answer.

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

ReadSaveConnections

The brain at rest operates near a critical point between order and chaos, and this paper tests whether people who later develop dementia show detectable deviations from that critical state in their sleep EEG years earlier. Using multifractal analysis of slow-wave sleep signals, those who remained cognitively healthy showed Hurst exponent distributions closer to the theoretical critical point than those who progressed to dementia (p ≤ 0.001). Sleep EEG is already recorded in clinical settings, so if validated in larger cohorts, this could become a low-cost early warning signal for cognitive decline without requiring expensive imaging.

██████████ 0.7 sleep-circadian-psychiatry Preprint

ReadSaveConnections