278

Papers

9/9

Roadblocks Active

0

Connections

This study tested whether measurable properties of a person's voice — how their pitch wavers, how irregular their vocal folds vibrate, and how they structure sentences — can track the severity of depression, anxiety, and ADHD across multiple independent datasets. Using standard acoustic tools and an interpretable machine learning model (XGBoost with SHAP explanations), the authors found consistent associations between specific vocal irregularities and validated symptom scores. The value here is replication across datasets rather than a single dataset result: if the same vocal markers appear across different populations and recording conditions, they become more credible candidates for passive monitoring tools.

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

ReadSaveConnections

Most EEG depression classifiers rely on data augmentation to compensate for small clinical datasets, which risks introducing artificial patterns. This paper instead trains a generative model exclusively on healthy-brain EEG to learn what 'normal' looks like, then scores patient EEG by how far it deviates from that normal — the deviation score becomes the diagnostic signal. An additional module handles the practical problem that different EEG headsets have different numbers and placements of electrodes, making models trained on one device useless on another. If the approach holds up under peer review, it addresses two real deployment barriers simultaneously: small sample sizes and hardware heterogeneity.

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

ReadSaveConnections

A persistent failure mode in digital therapeutics is that patients disengage — and most optimization algorithms treat this dropout as random background noise rather than as something caused by the intervention itself. This paper proposes a mathematical model where a patient's capacity for engagement fluctuates over time and is directly shaped by what the system recommends, then derives an algorithm (UCB-BOLD) that accounts for this feedback loop when selecting treatment actions. The theoretical guarantees are solid, but the only empirical test uses synthetic patients generated from a sedentary-behavior trial, so clinical validation remains the critical next step.

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

ReadSaveConnections

After ibogaine administration, brain electrical activity shifts toward slower oscillations for weeks — reduced fast-frequency power, lower signal complexity, and changes in inhibitory processing that persist well beyond the acute drug experience. This paper proposes that this extended post-acute window represents a distinct neurobiological state ('neural attunement') where the brain is more receptive to learning and behavioral change, which could explain why ibogaine shows clinical promise for addiction and mood disorders. The framework is entirely theoretical with no new data, so its main contribution is structuring testable predictions for future longitudinal EEG studies.

██████████ 0.9 psychedelic-mechanisms Peer-reviewed

Read

Cancer survivors experience high rates of depression and anxiety, but they tend not to fill out mood diaries at the moments when their distress is worst — creating a blind spot precisely when intervention would help most. PULSE addresses this by using smartphone sensors (movement, sleep, screen use, location) to infer emotional state without requiring self-report, with an AI agent that autonomously queries and interprets the sensor data. On 50 participants, the agentic approach outperformed a fixed-pipeline approach, reaching 74.3% balanced accuracy for predicting when a person wants emotional regulation support — a meaningful threshold for triaging proactive outreach.

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

ReadSaveConnections

EmoTrack extracts structured clinical signals from therapy session transcripts using a large language model, then combines these with sentence-level embeddings to predict PHQ-8 depression severity scores. Across the DAIC-WOZ benchmark (real virtual-agent interviews), it achieves a 13.5% reduction in prediction error compared to the previous best method. A notable design choice — keeping the turn-level embeddings frozen rather than fine-tuning them — preserves contextual nuance and is likely why performance improves; this is a reusable architectural insight for other transcript-based assessment tools.

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

ReadSaveConnections

Collecting large labeled datasets from PTSD patients is expensive and ethically complex, but phobia datasets are more accessible and share physiological fear-response signatures. This study trains a model on heart rate and skin conductance data from people watching spider videos, then transfers that fear-response model to estimate PTSD severity in 21 military personnel during a combat simulation. The transfer approach achieved 86% classification accuracy and a mean absolute percentage error of 17% on clinical severity scales — promising results, but the tiny sample size (21 participants) means these numbers should be treated as proof-of-concept rather than validated performance.

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

ReadSaveConnections

When a deep learning model says someone has depression based on their EEG, clinicians need to know which brain regions and frequency bands drove that decision. This paper applies five different explanation methods to the same EEG classifier and asks whether they agree — and largely they do, with frontal and right-hemisphere temporal regions consistently flagged, which aligns with established neuroscience. The practical implication is that EEG depression models may be interpretable enough to support clinical trust, though the absence of a named dataset in the available text limits how much weight to place on the specific results.

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

ReadSaveConnections

Five LLMs were tested on 555 semi-structured clinical interviews to see whether they could classify anxiety disorder, depression, PTSD, and any mental health disorder — accuracy ranged widely from 0.49 to 0.86, and Matthews correlation coefficients were low across the board (0.16–0.38), meaning current LLMs are unreliable for psychiatric screening. Crucially, false negatives often occurred not because symptoms were absent from the text, but because the LLM downweighted symptoms when coping or social support was also mentioned — mirroring a well-known bias in human clinicians. This is a useful empirical calibration for anyone considering LLMs as triage tools: the models make systematically predictable errors, not random ones.

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

ReadSaveConnections

Computational psychiatry is split between two types of brain models: biologically realistic simulations that match brain anatomy but cannot perform cognitive tasks, and task-performing neural networks that are functionally capable but anatomically disconnected from real brain structure. This position paper argues both traditions are incomplete and proposes a roadmap for 'functional whole-brain models' that would satisfy both biological and functional criteria simultaneously. No new data or models are presented — this is a conceptual proposal — but if the field takes it seriously, it would change how computational tools for psychiatric diagnosis are designed and validated.

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

ReadSaveConnections