282

Papers

10/10

Roadblocks Active

0

Connections

A multi-agent AI system autonomously analyzed wearable sensor data from over 9,000 participant-observations and surfaced 41 candidate digital biomarkers for depression, with sleep timing variability (when you fall asleep and how consistently) appearing as a signal across two independent depression cohorts. This matters because it demonstrates that an AI pipeline can replicate the kind of exploratory biomarker work that normally takes a research team months, and it points toward circadian disruption as a robust, measurable signal. The key caveat is that the system relies on non-public datasets and non-deterministic LLM components, so external validation will be essential before clinical use.

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

ReadSaveConnections

This study tested whether the geometric patterns of how someone's voice moves through acoustic space during conversation — not just average pitch or volume — could flag depression, achieving an AUC of 0.689 on a standard benchmark. The mechanism is that depression may alter the temporal structure of vocal behavior in ways that static averages miss, and recurrence-rate features captured this better than entropy or instability measures. Confidence in the result is limited by a small imbalanced sample (42 depressed vs 100 controls), an unjustified threshold choice, and the absence of shared code.

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

ReadSaveConnections

Using the same DAIC-WOZ dataset, this paper argues that Shannon entropy of acoustic trajectories over time (how unpredictably someone's voice varies across a conversation) outperforms static pooling, recurrence, and fractal measures for depression classification, lifting AUC from 0.593 to 0.646. The finding suggests depression's vocal signature lies in the dynamics of conversation rather than in any single acoustic snapshot. Notably, the authors discovered and corrected a data corruption issue in a shared aggregate file, which is a meaningful transparency contribution, though the absence of released code limits reproducibility.

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

ReadSaveConnections

ADAPTS breaks long clinical interview transcripts into symptom-by-symptom reasoning tasks handled by specialized LLM sub-agents, achieving inter-rater reliability (ICC = 0.877) that matches or beats human raters on the most discrepant cases. This matters because consistent symptom rating is a chronic bottleneck in clinical research and trial design, and automating it could accelerate both assessment and data collection. However, the evaluation relies on proprietary datasets and undisclosed prompts, making independent replication currently impossible.

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

ReadSaveConnections

This system uses a GPT-generated clinical summary at each stage of a three-stage detection pipeline (screen, classify severity, regress score) to guide how audio, video, and text signals are combined, improving both accuracy and interpretability over prior multimodal baselines on two benchmark datasets. The key mechanism is that LLM-generated clinical language acts as a soft attention signal telling the fusion module which modalities to weight at each step. Results are promising but rest on a closed commercial model and very small test sets (56 samples for E-DAIC), limiting generalizability claims.

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

ReadSaveConnections

EviDep addresses a specific failure mode in multimodal AI diagnostics: when audio and video features share redundant information, naive fusion inflates the model's confidence in its own predictions — a dangerous property in clinical tools. The paper uses a statistical framework (Normal-Inverse-Gamma distribution) to separately track uncertainty from noise versus uncertainty from lack of knowledge, and disentangles shared from unique signal before fusing. This is a meaningful architectural contribution to trustworthy clinical AI, though code is not released and the method is evaluated only on benchmark datasets.

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

ReadSaveConnections

When asked to simulate patients across 120 demographic groups, frontier LLMs produced individuals that looked clinically plausible but compressed the real population distribution — erasing the extremes of illness that matter most clinically — and 37% of simulated cases crossed diagnostic thresholds between two runs. This coherence-without-fidelity problem means LLM-generated synthetic patients cannot safely substitute for real patient data in training or testing clinical AI systems. The finding is a direct warning for any researcher using LLM-simulated cohorts to develop or validate mental health tools.

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

ReadSaveConnections

This work shows that EEG recorded during a cognitive task (stop-signal task) classifies ADHD significantly better than resting-state EEG, with gamma-band activity over frontal and parietal regions driving the classification. Beyond diagnosis, white matter connectivity in motor-planning brain tracts correlated with individually modeled effort and reward sensitivity parameters, linking brain structure to motivational behavior. The translational value is that task-based biomarkers may capture motivational circuitry disruptions relevant not only to ADHD but to depression and treatment resistance.

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

ReadSaveConnections

Analyzing 830 therapeutic writing samples, this study found that how people structure a story (whether it has a clear beginning, complication, and resolution) predicts clinical mental health outcomes better than which words they use or how semantically coherent their sentences are. The mechanism is that narrative organization reflects cognitive and emotional processing capacities that word-level sentiment analysis misses. The practical implication is that therapy homework and journaling apps could extract richer clinical signal by evaluating story structure rather than just sentiment.

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

ReadSaveConnections

By routing content through a chain of specialized LLM agents with principled statistical stopping rules, this system reduced false positives in self-harm content detection by 40% compared to a single-agent baseline (false positive rate 0.095 vs 0.159) while maintaining sensitivity. The mechanism is that adaptive sampling — running additional agents only when early agents disagree — concentrates computational effort where uncertainty is highest. For platforms deploying automated crisis screening, reducing false positives matters because unnecessary interventions erode user trust and overwhelm support resources.

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

ReadSaveConnections