281

Papers

10/10

Roadblocks Active

0

Connections

Five leading LLMs were tested on 555 real clinical interviews with verified psychiatric diagnoses, achieving accuracy between 49% and 86% depending on diagnosis and model — far from clinical-grade reliability. The study reveals that models systematically miss cases by under-weighting functional impairment cues relative to symptom descriptions, and that accuracy is higher for male than female participants, raising equity concerns. This matters because LLMs are already being deployed in mental health triage contexts, and this work provides the first systematic evidence of where and how they fail.

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

ReadSaveConnections

This paper proposes training a generative model exclusively on healthy EEG data to learn what 'normal' brain activity looks like, then flagging depression by measuring how far a patient's signal deviates from that norm — without any data augmentation or synthetic sample generation. A cross-channel spatial adaptation module allows the same model to work across EEG devices with different electrode configurations, addressing a key barrier to multi-site clinical deployment. If the approach holds up under broader validation, it could reduce the data quantity needed to build usable depression detection systems.

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

ReadSaveConnections

Vocal irregularities — specifically shimmer (amplitude variation) and jitter (pitch instability) — show stable correlations with depression, anxiety, and ADHD symptom severity across five separate datasets, suggesting these features reflect something real about mental state rather than dataset-specific noise. The study also finds that sentence-level grammatical patterns and affective tone carry independent diagnostic signal beyond acoustics alone. The use of publicly available tools (Praat, spaCy, HuBERT) means clinical teams could in principle replicate the feature pipeline, though a proprietary dataset limits full reproducibility.

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

ReadSaveConnections

This paper addresses a practical problem that undermines most digital therapeutics trials: patients who don't follow recommendations become less likely to follow future ones, making adherence itself a moving target that depends on the interventions being tested. The authors model this feedback loop as a linear dynamical system and derive an algorithm (UCB-BOLD) that provably learns optimal treatment sequences while accounting for this dependency. The theoretical guarantees are validated on a synthetic cohort calibrated to real micro-randomized trial data, making this relevant to any app-based mental health intervention where dropout is a design variable.

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

ReadSaveConnections

TimeSRL converts raw passive sensing data (phone usage, movement, etc.) into natural language descriptions before making mental health predictions, allowing a reinforcement-learned LLM to reason about behavior in human-interpretable terms rather than raw numbers. Tested under a leave-one-dataset-out protocol — the hardest generalization test, where the model never sees data from the target population — it reduces anxiety prediction error by 3–44% over baselines. The key insight is that semantic abstraction acts as a regularizer, preventing the model from memorizing dataset-specific quirks.

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

ReadSaveConnections

EmoTrack extracts structured clinical signals from therapy session transcripts using an LLM, then combines these with semantic sentence embeddings to predict PHQ-8 depression severity scores, achieving a 13.5% error reduction over the best existing single-session benchmark. The system also handles multi-session therapy by maintaining a compact memory of the previous session, mimicking how a clinician tracks change over time. A notable limitation is that the multi-session evaluation used a fully synthetic benchmark generated by another LLM, so real-world longitudinal performance remains unconfirmed.

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

ReadSaveConnections

Five different explainability methods — covering gradient-based, perturbation-based, and Shapley value approaches — were applied to the same EEG depression classifier to see whether they agree on which brain signals matter most. They largely do: frontal and right-hemisphere temporal regions consistently emerge as the most diagnostically informative, aligning with established neuroscience of depression. This cross-method agreement is meaningful because it reduces the risk that any single explainability technique is just reflecting its own mathematical assumptions rather than true brain signal.

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

ReadSaveConnections

Cancer survivors often feel worst precisely when they stop logging their symptoms — a 'diary paradox' that breaks standard prediction pipelines. PULSE equips an LLM with eight specialized tools to actively query smartphone sensor data (location, activity, sleep patterns) and achieves 74.3% balanced accuracy predicting when survivors want emotional support, outperforming fixed feature pipelines. The agentic architecture — where the model decides which data to examine and in what order — is the key innovation, and the approach is transferable to any mental health population where self-report is unreliable under distress.

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

ReadSaveConnections

This design research paper examines how AI companion systems can be structured to provide meaningful emotional support for young adults rather than superficial engagement, proposing a framework organized around dimensions of psychological benefit. The work addresses a real deployment gap: AI companions are increasingly popular among young adults, but most are not designed with evidence-based emotional support principles. The findings are relevant to product teams building mental health apps and to researchers studying parasocial human-AI relationships in vulnerable populations.

██████████ 0.8 digital-therapeutics Peer-reviewed

Read

Rather than requiring a large labeled PTSD dataset, this study trains a physiological fear-response model on publicly available arachnophobia data and transfers it to predict PTSD severity in 21 military veterans using heart rate and skin conductance signals, achieving 86% classification accuracy and a mean absolute error of 5.6 PCL points. The transfer learning strategy is pragmatically motivated: trauma datasets from military populations are extremely hard to obtain at scale, so borrowing signal structure from a more accessible fear condition is a reasonable workaround. The small sample size (21 participants) means these results need replication before clinical interpretation.

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

ReadSaveConnections