287

Papers

10/10

Roadblocks Active

0

Connections

Instead of generating synthetic EEG data to balance training sets — a common but noisy practice — this paper trains an unsupervised generative model on healthy brains only, then uses how anomalous a new EEG looks to that model as a diagnostic signal for depression. The approach also includes a module to handle mismatched electrode configurations across hospital sites, a practical barrier that has long blocked multi-center EEG studies. If the results hold under rigorous cross-validation, this could reduce the data-collection burden for deploying EEG depression screening in new clinical settings.

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

ReadSaveConnections

This pilot study simultaneously records two complementary brain signals — EEG (electrical activity) and fNIRS (blood-flow proxies in the frontal cortex) — during a memory task, then feeds both into a single machine learning pipeline to classify depressive states. Combining these signals matters because EEG has high time resolution but poor spatial specificity, while fNIRS provides localized frontal information; together they may catch signatures either alone misses. The study is very small (11 healthy students) and should be treated as a feasibility demonstration rather than a validated tool, but the multimodal pipeline design is worth watching for replication.

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

ReadSaveConnections

Speech-based depression screening works by detecting acoustic patterns linked to mood, but those same patterns inadvertently reveal the speaker's gender and age — a privacy risk that could prevent clinical deployment. InfoShield uses an information-theoretic method to mathematically squeeze out demographic identifiers from the audio representation while keeping the depression-relevant signal, reducing gender inference accuracy from 93% to near-chance (56%) with minimal loss in depression classification. The key technical contribution is a new estimator (TimeAwareMINE) that correctly handles the sequential, time-varying nature of speech, where standard privacy-stripping methods had previously failed.

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

ReadSaveConnections

Black-box neural networks can classify depression from EEG with reasonable accuracy, but clinicians need to know which brain regions and time windows drove that classification before they can trust — or act on — the output. This paper runs five different explanation methods on the same EEG depression model and finds they largely agree: frontal and right-hemisphere temporal regions are consistently highlighted, which aligns with established neuroscience of mood disorders. Convergence across methods strengthens confidence that the model is picking up real biology rather than data artifacts, though the absence of a ground-truth attribution benchmark limits how far that confidence can be pushed.

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

ReadSaveConnections

Measuring PTSD severity objectively — rather than relying solely on self-report questionnaires — has been a persistent challenge, especially for military populations. This paper transfers a physiological fear-response model originally trained on spider-phobia data (heart rate and skin conductance during VR exposure) to predict PTSD severity in veterans, achieving 86% classification accuracy and a mean severity estimation error of about 5.6 PCL-M points. The cross-disorder transfer is conceptually interesting because it treats PTSD as a variant of a conditioned fear response, but the very small retained sample (21 of 41 participants) and restricted dataset make generalization claims premature.

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

ReadSaveConnections

This small randomized trial tested whether adding a smartwatch-based digital intervention to a veteran cycling program produced better mental health outcomes than cycling alone, tracking anxiety, depression, and PTSD symptoms weekly alongside continuous physiological sensing. Veterans in the digital intervention arm showed more stable hyperarousal trajectories over time, while the cycling-only group showed late-study symptom escalation — suggesting the digital layer may help veterans consolidate acute benefits from physical activity. With only 10 randomized participants, this is explicitly a pilot, but it provides useful signal about the added value of passive physiological monitoring layered on top of existing wellness programs.

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

ReadSaveConnections

Large language models can classify mental health conditions from text, but they tend to reason inconsistently — jumping to conclusions or hedging indefinitely. This paper introduces a reinforcement learning training scheme (CRPO) that teaches an LLM to move through structured cognitive stages (appraisal, reaction, mental state) before reaching a diagnostic label, mimicking how a clinician builds a formulation. The approach improves weighted F1-score by 10.4 percentage points over the best RL baseline across eight mental health classification tasks, without requiring labeled clinical notes for fine-tuning — a meaningful practical advantage given how scarce annotated psychiatric data are.

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

ReadSaveConnections

A systematic audit of 12 pretrained EEG model–dataset combinations finds that all of them encode who the subject is far more strongly than what mental or cognitive state they are in — subject identity variance was 13 to 89 times higher than chance across every tested combination. This is a serious methodological problem for the depression-biomarker field: if a model is mostly learning individual brain fingerprints rather than disorder-related patterns, accuracy metrics computed at the dataset level will look good but the model will fail on new patients. The paper shows that simply removing the subject-identity axis from representations boosts label decoding by 6–27 percentage points, pointing toward a concrete fix.

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

ReadSaveConnections

Deep (N3) sleep is the most restorative sleep stage and its disruption is strongly linked to depression and cognitive decline, but reliably detecting it from EEG in real time has required complex models. This paper shows that a single feature derived from fractal scaling analysis (Detrended Fluctuation Analysis) can identify N3 sleep with 87% balanced accuracy using a simple Naive Bayes classifier — far outperforming linear approaches that score as low as 51%, which the authors explain by showing the feature lives on a curved, non-linear manifold. The result is validated on nearly 350,000 epochs from 290 older women, making it one of the largest sleep-EEG validation efforts for this approach and a credible candidate for lightweight wearable sleep tracking.

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

ReadSaveConnections

Most LLM-based depression screening tools require expensive fine-tuning on clinical data; this paper shows that carefully structuring the prompt around five clinical dimensions (cognitive patterns, emotional state, behavioral changes, somatic symptoms, social functioning) and then quantifying how confident the model is in each rationale via token-level entropy can match or exceed fine-tuned competitors. The approach was benchmarked across 21 different foundation models, which is unusually broad coverage and increases confidence that the gains reflect the prompting strategy rather than a lucky model-dataset pairing. A training-free method that generalizes across model families would lower the barrier for deploying depression screening in resource-limited settings.

██████████ 0.7 depression-biomarkers Preprint

ReadSaveConnections