276

Papers

10/10

Roadblocks Active

0

Connections

The paper trains a generative model exclusively on healthy brain signals to learn what 'normal' EEG looks like, then uses how much a new patient's signal deviates from that normal as a guide for depression classification — no synthetic data needed. This directly attacks the small-sample problem that has plagued EEG-based psychiatry: most clinical datasets have fewer than 100 subjects. The Cross-Channel Spatial Adaptation module also addresses a practical barrier — different hospitals use different EEG hardware configurations, and this approach allows models trained at one site to work at another.

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

ReadSaveConnections

Rather than building another depression classifier, this paper asks which brain regions the classifier is actually using — and whether different explanation methods agree. The finding that gradient-based and perturbation-based methods converge on frontal and right-hemisphere temporal regions is meaningful because it aligns with known neurobiological theories of depression, lending modest credibility to the underlying model. For clinical adoption, explainability is not optional — clinicians won't trust a black box, so this kind of method auditing is a necessary step before deployment.

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

ReadSaveConnections

Most digital health algorithms assume that whether a patient follows treatment recommendations is independent of the recommendations themselves — this paper shows that assumption is mathematically wrong and proposes a model where past adherence and current recommendations jointly shape future engagement. The UCB-BOLD algorithm provides theoretical guarantees on how quickly the system learns the right intervention policy even as adherence evolves. This matters because real-world digital therapeutics routinely fail due to drop-off, and algorithms that treat adherence as fixed are systematically miscalibrated from the start.

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

ReadSaveConnections

This paper applies reinforcement learning to teach a large language model to reason through mental health assessments the way a clinician does — moving from gathering evidence to forming a provisional judgment, rather than answering immediately. The 10.4 percentage point improvement in weighted F1 over other RL-trained baselines across eight datasets covering depression, anxiety, and suicide risk is notable for a training-free architecture. The stage-wise entropy mechanism — which forces the model to be uncertain early and confident late in its reasoning chain — is a novel alignment technique with potential applicability beyond psychiatry.

██████████ 0.9 computational-psychiatry Preprint

ReadSaveConnections

This paper tests whether voice characteristics — subtle irregularities in pitch and amplitude (jitter and shimmer), sentence complexity, and emotional tone — consistently track with symptom severity across five different datasets for depression, anxiety, and ADHD. Finding consistent relationships across multiple independent datasets is more meaningful than strong results on a single dataset, because it suggests the signal is real rather than a dataset artifact. The use of SHAP to explain which features drive predictions adds transparency that is increasingly required before clinical tools can be considered for deployment.

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

ReadSaveConnections

This is one of the few papers this cycle that combines a randomized design with real-world wearable sensing in a veteran mental health context — a population with high PTSD prevalence and historically poor engagement with conventional therapy. Veterans receiving the wearable-plus-digital intervention showed stabilized hyperarousal trajectories weeks after the cycling event, while the cycling-only group's gains faded; this suggests the digital layer adds durable benefit beyond exercise alone. The sample is very small (n=13 randomized), so these results are hypothesis-generating only, but the pre-registration and naturalistic deployment design are methodological strengths worth building on.

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

ReadSaveConnections

Using a publicly available dataset of 290 older women, this paper shows that a single mathematical property of EEG signals — the Hurst exponent, measuring how 'self-similar' brain activity is over time — can classify deep sleep with 87% accuracy using a simple Naive Bayes classifier. Counterintuitively, more complex models (neural networks, SVMs) performed worse, which suggests the underlying signal structure is probabilistic rather than linearly separable. Accurate automated deep sleep detection matters for mental health because N3 sleep deficits are implicated in depression, PTSD, and cognitive decline, and current detection requires expensive overnight lab stays.

██████████ 0.9 sleep-circadian-psychiatry Preprint

ReadSaveConnections

Speech-based mental health screening creates a privacy problem: voice recordings reveal gender, age, and other sensitive attributes that patients may not want exposed to a diagnostic system. InfoShield reduces the ability to infer gender from processed speech from 92.6% to near-chance (55.5%) while preserving depression classification performance, using an information-theoretic technique that strips demographic signal from the learned representation. The identification that standard mutual information estimators fail on speech data — because speech is sequential while demographic labels are static — is a methodological contribution with broader implications for privacy-aware health AI.

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

ReadSaveConnections

This paper applies multifractal analysis to sleep EEG to test whether brain signal dynamics during non-REM sleep can predict future cognitive decline — finding significant differences between those who later developed dementia and those who remained cognitively healthy, particularly during deep sleep stages N2 and N3. The finding that healthy brains operate closer to a 'critical state' (a specific mathematical regime linked to optimal information processing) provides a mechanistic interpretation for why the signal works. Sleep EEG is far cheaper and less invasive than amyloid PET or lumbar puncture, making this a practically important direction if the biomarker replicates in prospective cohorts.

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

ReadSaveConnections

This paper asks whether a model trained to recognize fear responses in people with spider phobia — a publicly available dataset — can be repurposed to measure PTSD severity in veterans using heart rate and skin conductance, without retraining from scratch. Achieving 86% classification accuracy and reasonable severity estimation (mean absolute error ~5.6 PCL-M points) in a 21-subject military sample is a proof-of-concept that fear-response physiology may generalize across anxiety disorders. The practical implication is that large labeled PTSD datasets are hard to collect, but this approach could bootstrap from easier-to-obtain phobia data — though the tiny sample demands replication before drawing firm conclusions.

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

ReadSaveConnections