285

Papers

11/11

Roadblocks Active

0

Connections

A multi-agent AI system autonomously analyzed wearable sensor data from over 9,000 participant-observations across two depression cohorts and identified 41 candidate digital biomarkers, with sleep timing variability (circadian instability) emerging consistently in both independent samples. The system combines large language models for hypothesis generation with rigorous statistical checks including permutation testing and bootstrap confidence intervals, going beyond typical automated ML pipelines. While two of three datasets are proprietary and independent replication is currently blocked, the convergent finding on circadian instability across cohorts gives this more credibility than a single-dataset result.

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

ReadSaveConnections

This study treats a depressed person's voice during conversation not as a bag of acoustic features but as a trajectory through a high-dimensional state space, then measures how often that trajectory revisits the same regions — a technique called recurrence quantification analysis. Applied to 142 participants in the DAIC-WoZ clinical interview dataset, this approach achieved an AUC of 0.689, outperforming static acoustic descriptors, entropy measures, and several other nonlinear baselines. The mechanism matters because it captures the loss of vocal flexibility over time rather than average pitch or energy, which may better reflect the psychomotor rigidity characteristic of depression.

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

ReadSaveConnections

Using the same DAIC-WoZ dataset as the companion recurrence paper, this study finds that Shannon entropy of acoustic trajectories over time achieves AUC 0.646 for depression detection, statistically significantly outperforming static acoustic pooling (AUC 0.593, p=0.017) and also outperforming recurrence, sample entropy, and fractal complexity measures on this dataset. The core insight is that depressed speech is less informationally complex over time — the voice loses its moment-to-moment unpredictability. Together with the recurrence paper above, today's evidence strongly supports temporal vocal dynamics as a reproducible passive biomarker target, pending validation in non-clinical settings.

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

ReadSaveConnections

This paper stress-tests whether frontier AI models (GPT-4o-mini, DeepSeek-V3, Gemini, GLM-4) can accurately simulate population-level mental health distributions when asked to generate synthetic patient profiles across 120 demographic intersections. The finding is a clear failure mode: LLMs produce internally coherent-sounding individuals but systematically compress population variance by 14–62%, erasing the clinically important tails of severity distributions, and 37% of cases flip diagnostic category between two identical test runs. This matters because synthetic data generation for mental health AI training is increasingly common, and these distortions would silently bias any model trained on such data.

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

ReadSaveConnections

Two vision-language models evaluated for depression detection showed performance swinging from 33.9% to 80.4% accuracy depending on whether the data came from a controlled lab or a naturalistic clinical setting — a warning that benchmark performance may not transfer to real-world deployment. Both models also showed systematic biases: one over-predicted depression broadly, and the two models divided their demographic errors differently, with one showing higher gender disparity and the other higher racial bias. The paper proposes chain-of-thought fairness prompting and a counterfactual fairness loss as partial mitigations, though these are not yet fully validated.

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

ReadSaveConnections

K-SENSE improves automated detection of depression and stress in social media text by augmenting a language model with external commonsense knowledge about mental states — specifically using the COMET model to infer what someone might be feeling, wanting, or fearing based on their posts. This is combined with contrastive learning to make the model's internal representations more robust, yielding F1 scores of 86.1% on stress detection and 94.3% on depression detection, representing modest but consistent improvements over prior baselines. The approach is meaningful because it moves beyond pattern-matching on surface language toward reasoning about psychological context, which should generalize better across writing styles.

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

ReadSaveConnections

PsyGAT models a clinical interview as a graph where each utterance is a node and edges represent psychological state transitions, encoding clinical evidence at the utterance level using constructs grounded in diagnostic criteria. This architecture achieves 89.99 Macro F1 on DAIC-WoZ and 71.37 on E-DAIC, surpassing both graph-based baselines and closed-source large language models, while also offering an interpretability module that highlights which conversational moments drove the prediction. The interpretability component is practically important because clinician-facing depression detection tools require transparent reasoning, not just accuracy.

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

ReadSaveConnections

EviDep addresses a practical gap in multimodal depression assessment: most models give a severity score but no indication of how confident they are, making it hard to know when to trust the prediction. By using evidential learning with a Normal-Inverse-Gamma distribution, the model simultaneously outputs a depression severity estimate and two types of uncertainty — aleatory (irreducible noise in the data) and epistemic (model uncertainty from insufficient information). The disentangled fusion strategy also reduces redundancy when combining audio, video, and text, achieving state-of-the-art results across four benchmark datasets (AVEC 2013, 2014, DAIC-WoZ, E-DAIC).

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

ReadSaveConnections

A persistent problem in multi-site neuroimaging studies is that AI models trained on data from one hospital forget what they learned when updated with data from a new site — a phenomenon called catastrophic forgetting. This paper addresses it by training a structure-aware variational autoencoder to synthesize realistic functional connectivity matrices from previously seen sites, then replaying those synthetic examples during training on new data. A hierarchical bandit scheme prioritizes which past cases to replay, and multi-level knowledge distillation keeps the model's representations aligned across sites, enabling sequential learning without performance collapse on earlier datasets.

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

ReadSaveConnections

This system uses GPT-o3 to generate progressively detailed clinical summaries from interview data — first capturing emotions, then PHQ-8 symptom dimensions, then potential causes — and uses these structured summaries to guide fusion of audio, video, and text modalities via report-gated cross-attention. The coarse-to-fine pipeline (binary screening → five-class severity) is intended to mimic how a clinician builds an impression incrementally. The main limitation for adoption is hard dependence on proprietary GPT-o3 for the core summarization step, which creates reproducibility and cost barriers and exposes the system to API version drift.

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

ReadSaveConnections