289

Papers

9/9

Roadblocks Active

0

Connections

This paper trains two generative models exclusively on healthy EEG data to learn what normal brain activity looks like, then uses how far a patient's EEG deviates from that norm as an input signal for depression classification — without synthesizing any fake patient data. The approach sidesteps the chronic data scarcity problem in clinical EEG research and introduces a hardware-alignment module to handle the messy reality of multi-site datasets recorded with different electrode setups. It matters because it reframes the diagnostic question: instead of asking 'does this look depressed?' it asks 'how far does this deviate from healthy?' which may generalize better across clinics.

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

ReadSaveConnections

Most adaptive digital health algorithms treat patient engagement as a background factor unrelated to what the app recommends, but this paper proves mathematically that engagement and recommendations are causally entangled — each affects the other over time. The authors build a model that captures this feedback loop and design a treatment-selection algorithm (UCB-BOLD) with formal performance guarantees, tested against a real workplace behavior trial. This matters because digital mental health apps that ignore adherence endogeneity will optimize toward recommendations that patients happen to follow rather than recommendations that actually help them.

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

ReadSaveConnections

The study extracts acoustic voice features (pitch stability, jitter, shimmer) and linguistic patterns from speech and finds stable associations with symptom severity across depression, anxiety, and ADHD, validated across five separate datasets including one real-world clinical dataset. Using XGBoost with SHAP explanations, the work identifies which specific features drive predictions — making it auditable rather than a black box. Cross-disorder replication across five datasets is rare in this literature and meaningfully raises the bar for what a speech biomarker claim needs to demonstrate.

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

ReadSaveConnections

TimeSRL converts raw smartphone sensor streams (step counts, screen time, etc.) into plain-language behavioral summaries before predicting anxiety and depression scores — a detour through language that turns out to generalize better across different study populations than feeding numbers directly into models. The key methodological advance is using reinforcement learning to train the natural-language translation step without requiring human annotation of intermediate descriptions. The leave-one-dataset-out evaluation protocol is notably rigorous and addresses a core weakness in passive sensing research: models that work on one cohort often fail completely on another.

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

ReadSaveConnections

Tested against 555 real diagnostic interviews with gold-standard SCID labels, five leading LLMs showed wide accuracy variation (49–86%) and frequently missed anxiety and PTSD diagnoses not because they failed to detect symptoms, but because they over-weighted mentions of coping ability and social support as evidence of health. This is clinically important: the models appear to be making the same reasoning error that clinicians are trained to avoid — confusing functional preservation with absence of disorder. The study gives actionable guidance on where LLM psychiatric tools currently fail and why.

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

ReadSaveConnections

EmoTrack reduces prediction error on PHQ-8 depression scores from therapy transcripts by 13.5% relative to the best prior approach, combining LLM-extracted clinical signals with semantic embeddings and a memory mechanism that carries forward information across multiple sessions. Tracking symptom change over the course of treatment — rather than single-session snapshots — is what actually matters clinically, and this is one of the first systems to address it directly. The reliance on a self-constructed synthetic multi-session dataset for the longitudinal evaluation is a limitation that future work will need to address with real clinical data.

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

ReadSaveConnections

Von Economo neurons (VENs) are large, spindle-shaped brain cells found in regions associated with social cognition and known to be reduced in conditions like autism and frontotemporal dementia. Simulating their removal in a spiking neural network reduced reliable learning convergence from 98% to 70% of runs (Fisher's exact p=8.7e-5), with the disruption concentrated in mid-training when circuits are forming interdependencies. The computational specificity here — identifying a narrow developmental window where VEN loss is most disruptive — generates testable predictions for when social skill deficits might emerge or be most amenable to intervention.

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

ReadSaveConnections

OSSMM is a wearable headband that classifies four sleep stages (Wake, Light, Deep, REM) with 77.6% accuracy and moderate agreement with gold-standard polysomnography (κ=0.63), using conductive electrodes from off-the-shelf fitness chest straps rather than clinical-grade hardware. The open-source design is significant because the cost and access barriers of standard sleep monitoring are a major bottleneck in sleep-psychiatry research, particularly for longitudinal studies. A simplified two-frontal-electrode configuration that still captures sleep spindle signatures opens the door to practical at-home sleep research at scale.

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

ReadSaveConnections

In 50 cancer survivors — a group with high depression and anxiety burden who often go undetected — an LLM agent equipped with smartphone sensing tools predicted emotion regulation needs with 74.3% balanced accuracy, compared to 52–60% for standard machine learning pipelines. The key finding is that giving an AI agent the ability to dynamically query different sensing streams (rather than using a fixed feature set) produces meaningfully better predictions. This matters because it demonstrates a path toward proactive mental health monitoring that does not require patients to actively self-report when they are least likely to do so.

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

ReadSaveConnections

InfoShield reduces the ability to infer a speaker's gender from a depression-screening model's internal representations from 92.6% down to 55.5% (near chance), while keeping depression classification performance largely intact — by mathematically removing demographic information from the representation rather than adding privacy noise. The paper also diagnoses why standard privacy estimation methods fail on speech: temporal audio signals and static demographic labels don't align, requiring a new cross-modal attention estimator. This addresses a concrete deployment barrier for speech-based mental health tools, where demographic leakage creates both ethical and regulatory risks.

██████████ 0.7 depression-biomarkers Preprint

ReadSaveConnections