282

Papers

10/10

Roadblocks Active

0

Connections

This paper trains a generative model exclusively on healthy-brain EEG to learn what 'normal' looks like, then uses deviations from that baseline as an explicit signal to guide a depression classifier — bypassing the need for data augmentation tricks that often introduce artifacts. It also includes a module that normalizes across hospitals using different EEG headset configurations, which is a practical barrier that has blocked multi-site depression studies. The approach matters because it separates the problem of 'what does a depressed brain look like?' from 'how do we get enough labeled data?', which is a real bottleneck in psychiatric neuroimaging.

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

ReadSaveConnections

Most digital therapy systems treat patient engagement as something that just happens, unaffected by the recommendations themselves — this paper argues that is wrong, and builds a mathematical model where adherence is shaped by past treatment history and current recommendations simultaneously. The UCB-BOLD algorithm derived from this model can learn in real time which interventions sustain engagement for a specific patient without burning out their willingness to participate. This matters practically because dropout is the leading cause of digital therapeutic failure, and this framework gives a principled way to avoid it.

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

ReadSaveConnections

This pilot study combines two non-invasive brain measurement techniques — EEG (electrical activity) and fNIRS (blood oxygenation) — simultaneously during a memory task, then applies a single end-to-end neural network to classify depressive states without manual feature engineering. The motivation is that clinical depression diagnosis still depends on self-report questionnaires, which are vulnerable to patient bias or minimization, particularly relevant in aging populations where depression and dementia commonly co-occur. The sample is very small (11 students) so the results are preliminary, but the methodology — multimodal passive sensing during cognitive tasks — points toward a plausible path to objective screening.

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

ReadSaveConnections

This system decomposes a clinical interview transcript into five structured themes aligned with how psychiatrists actually assess depression (matching the SCID diagnostic framework), then uses token-level uncertainty scores to weight each piece of evidence before making a diagnosis — all without any model fine-tuning. It outperforms zero-shot approaches across 21 different large language models and matches or beats supervised specialist models, suggesting the structure of the reasoning matters as much as the model size. A significant caveat: the paper compares against commercial LLMs with version identifiers that do not correspond to any known deployed systems, which undermines the benchmark credibility.

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

ReadSaveConnections

This study systematically maps which vocal and language features — things like voice shakiness (shimmer/jitter), word variety, and sentence complexity — consistently track with depression, anxiety, and ADHD severity scores across multiple independent datasets. The finding that shimmer and jitter are stable across datasets is important because it suggests these acoustic features may be genuinely capturing something biological, not just dataset-specific noise. The practical implication is that a smartphone microphone during a structured conversation could provide clinically meaningful signal without any specialist equipment.

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

ReadSaveConnections

This paper uses a measure of how 'at the edge of chaos' the brain's electrical activity is during sleep — a property called criticality, quantified via the Hurst exponent — to classify whether someone is in deep (N3) sleep with 87% accuracy using only four EEG channels. The key finding is that a simple probabilistic classifier outperforms more complex linear models, revealing that these criticality features exist on a curved, non-linear manifold that linear approaches miss. This matters for mental health because deep sleep quality is closely linked to mood regulation, and a lightweight passive sensor that reliably detects N3 sleep could enable real-world neurofeedback interventions.

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

ReadSaveConnections

This small pre-registered trial tested whether adding a smartwatch-based hyperarousal detection system to an endurance cycling program for veterans with PTSD provided benefits beyond exercise alone — it did, with the digital group showing more stable anxiety trajectories rather than late-study relapse. The system uses heart rate and movement data to flag hyperarousal events in real time, confirmed by the participant, giving the person actionable self-awareness rather than passive tracking. The study is honest about its limitations (n=7 in the digital arm) but establishes a methodology for testing wearable-augmented physical rehabilitation in trauma populations.

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

ReadSaveConnections

When five different explanation methods are applied to the same EEG-based depression classifier, they largely agree on which brain regions matter most — frontal, temporal, and right-hemisphere areas — which is reassuring because it suggests the model is learning something neurologically plausible rather than exploiting dataset artifacts. The convergence between gradient-based and perturbation-based methods is methodologically important: these approaches make very different assumptions, so when they agree it is stronger evidence of genuine signal. This kind of audit is necessary before any EEG classifier could responsibly be deployed in a clinical setting.

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

ReadSaveConnections

Using a large longitudinal cohort of older women, this study shows that how close a person's sleeping brain activity is to a mathematical 'critical state' predicts whether they will develop cognitive decline five years later — measured non-invasively from standard sleep EEG. The effect is strongest during N2 and N3 deep sleep stages, suggesting that overnight brain dynamics encode information about long-term neurodegeneration that waking assessments miss. Because the underlying dataset (SOF/NSRR) is publicly available, this result is unusually well-positioned for independent replication compared to most papers in this space.

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

ReadSaveConnections

This paper asks whether a model trained to recognize fear responses to spiders can be adapted to measure PTSD severity in soldiers — and finds it can, achieving 86% binary PTSD classification accuracy and a mean severity estimation error of 5.6 points on a standard scale. The insight is that fear is fear at the physiological level: heart rate and skin conductance patterns during a stressful simulation carry transferable structure regardless of whether the fear source is biological or combat-related. The major limitation is that the target dataset is restricted military data unlikely to be shared, making independent validation currently impossible.

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

ReadSaveConnections