283

Papers

9/9

Roadblocks Active

0

Connections

Instead of generating fake EEG data to balance training sets, this paper trains an AI exclusively on healthy brains to learn what 'normal' looks like, then flags deviations as a depression signal. The key insight is that measuring how abnormal a brain signal is (the 'pathological prior') provides better classification cues than augmented data. This matters because synthetic augmentation can introduce artifacts that make models look good on benchmarks but fail in real clinical settings, and the paper also handles the messy reality that EEG equipment varies across hospitals.

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

ReadSaveConnections

Most digital health algorithms assume patients either follow a treatment plan or they don't — as if adherence is random luck. This paper models adherence as something the treatment itself shapes over time: a patient who is pushed too hard drops out, shrinking their capacity to engage in future sessions. The UCB-BOLD algorithm derived here can select personalized interventions online while learning each patient's engagement trajectory, with mathematical guarantees on how quickly it improves — a meaningful step toward adaptive mental health apps that don't inadvertently burn users out.

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

ReadSaveConnections

Voice-based depression screening is clinically promising but leaks sensitive demographic details: a model that hears your speech can infer your gender with 92% accuracy even when you only want it to assess depression risk. InfoShield strips out those demographic signals mathematically — reducing gender inference to near-chance (55%) while preserving depression classification — using a new mutual-information estimator designed specifically for the temporal structure of speech. This is important because privacy concerns are a real barrier to deploying voice screening tools in practice.

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

ReadSaveConnections

This paper systematically asks which speech features — vocal irregularities, sentence complexity, emotional tone — reliably correlate with depression, anxiety, and ADHD across multiple independent datasets, not just one. Finding that shimmer, jitter, and semantic coherence hold up across datasets is meaningful because most prior work optimizes for a single benchmark and then fails to replicate. The use of interpretable, perceptually grounded features (rather than black-box deep features) also means clinicians can understand what the model is responding to.

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

ReadSaveConnections

When a deep learning model says someone's EEG looks depressed, which brain regions drove that decision? This paper applies five different explanation techniques to the same model and checks whether they agree. The finding that gradient-based and perturbation-based methods converge on frontal and right-hemisphere regions is reassuring — it suggests those regions are genuinely informative rather than artifacts of one explanation method. Agreement across methods is a basic validity check that is rarely done but essential before clinical deployment.

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

ReadSaveConnections

Getting labeled PTSD physiological data from military veterans is hard, so this paper asks whether a fear-response model trained on spider phobia data can transfer to PTSD severity estimation via heart rate and skin conductance. With only 21 participants it achieves 86% binary classification accuracy and a 17% error rate on continuous severity scores — intriguing but fragile at this sample size. The transfer-learning angle is the novel contribution: if fear-circuit physiology generalizes across phobia types, it could reduce the data burden for rare or stigmatized conditions.

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

ReadSaveConnections

Most NLP tools for depression assessment are designed for a single conversation snapshot, but real therapy involves many sessions with trajectories that matter. EmoTrack uses a language model to extract clinical signals from therapy transcripts and adds a memory mechanism that activates only when prior sessions exist, achieving a 13.5% improvement over the best single-session benchmark. The ability to track symptom change across sessions — not just classify a snapshot — is what clinicians actually need to monitor treatment response.

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

ReadSaveConnections

This paper explores whether tracking where students look in classrooms can give teachers objective signals about mental health and engagement — potentially replacing purely subjective observation. Eye-tracking in schools is notable because it targets the environment where youth mental health problems often first become visible and where early intervention is most feasible. The approach is early-stage, but connecting gaze behavior to teacher training could extend mental health support into school settings without requiring clinical staff.

██████████ 0.8 youth-mental-health-crisis Peer-reviewed

Read

This study stress-tests five large language models on psychiatric screening using 555 real clinical interviews, finding accuracy ranges from 0.49 to 0.86 depending on the diagnosis and model — meaning some models perform barely above chance for certain conditions. A secondary finding that depression classification was more accurate for male than female participants raises a fairness concern that needs to be resolved before LLM screening tools reach clinical use. The use of structured clinical interviews (SCID) as ground truth rather than self-report questionnaires makes this a more rigorous evaluation than most prior work.

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

ReadSaveConnections

EEG foundation models — large pretrained neural networks meant to generalize across patients — turn out to encode who a person is far more than what disorder they have, with subject-identity variance 13 to 89 times larger than chance across all tested models. Crucially, this problem gets worse after fine-tuning, not better. The paper also shows that mathematically erasing the identity signal improves disorder classification by up to 27 percentage points, which is a practical fix — but it also means current benchmarks are likely overstating the clinical utility of EEG foundation models.

██████████ 0.7 computational-psychiatry Preprint

ReadSaveConnections