286

Papers

10/10

Roadblocks Active

1

Connections

This paper introduces MSFL, a framework that combines two usually-separate aspects of brain activity measured by fMRI — how strongly regions co-activate (amplitude) and how synchronised their timing is (phase) — into a single model for diagnosing autism and major depressive disorder. Both information channels independently contributed predictive signal according to interpretability analysis, and combining them outperformed existing approaches. This matters because most clinical brain imaging models discard phase information entirely, so the result suggests current biomarker pipelines may be leaving diagnostically meaningful signal on the table.

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

ReadSaveConnections

Using EEG recordings from 20 healthy adults during an emotional face task, this study quantifies how 'stable' the brain network state is when processing sad versus happy stimuli, finding that sad processing locks the brain into a lower-energy, more stable configuration — especially in the alpha band (Cohen's d = 0.83). Hyperconnected networks corresponded to more stable states, with a strong negative correlation between network stability and efficiency (r = -0.72). If this finding replicates in depressed patients, it offers a quantitative EEG-based marker of the rumination-like attractor states thought to maintain depressive episodes.

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

ReadSaveConnections

This paper tests whether combining EEG, eye-tracking, and facial action data during implicit association tests can predict suicidality and psychosis risk better than the standard D-score — a simple response-time measure. The multimodal Bayesian model achieved AUCs of 0.73–0.79 versus 0.50–0.53 for D-score alone, a meaningful improvement, though confidence intervals were wide (±0.18) and significance was marginal after correction. The study is important because implicit measures aim to bypass self-report bias in high-risk populations, and this is one of the first attempts to ground them in neural data with a principled statistical model.

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

ReadSaveConnections

Across three widely-used depression detection datasets (ANDROIDS, DAIC-WOZ, E-DAIC), this study shows that classifiers can distinguish depressed from control participants using only the interviewer's speech — not the patient's — because interviewers use predictably different prompts with the two groups. High classification performance was achieved without any participant language at all. This exposes a systemic data contamination problem: many published depression AI results may reflect learned interviewer habits rather than genuine patient biomarkers, which would make those models clinically useless.

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

ReadSaveConnections

A deep learning model trained on healthy and typical sleepers to classify sleep stages performed poorly when applied to EEG recordings from ischemic stroke patients with comorbid sleep disorders, and Grad-CAM visualisations showed the model was attending to physiologically uninformative EEG regions in patient data. This is a concrete demonstration that sleep-staging AI — which could support psychiatric diagnoses given the tight link between sleep and mental health — breaks down precisely in the clinical populations it needs to work in. The finding argues for training and validating sleep AI on clinical rather than convenience populations from the outset.

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

ReadSaveConnections

SynSym uses large language models to generate synthetic training examples for identifying psychiatric symptoms in text, with two key design choices: expanding each symptom into sub-concepts to capture how people actually describe it, and generating multi-symptom combinations guided by real clinical co-occurrence patterns. Models trained on synthetic-only data performed comparably to those trained on real data across three benchmarks. Because labelled psychiatric text data is scarce and ethically sensitive to share, this pipeline could meaningfully lower the barrier to training symptom detection tools without requiring patient data.

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

ReadSaveConnections

FEMBA pre-trains a bidirectional Mamba sequence model on over 21,000 hours of EEG data using a physiologically-informed objective (reconstructing low-frequency brain signals), then compresses it to run in real time on a microcontroller chip that draws very little power. Adding the physiological pre-training objective improved abnormal EEG detection AUROC from 0.863 to 0.893, and 2-bit quantization preserved accuracy where standard compression degraded it by 30%. Running a foundation EEG model on ultra-low-power hardware is a prerequisite for wearable psychiatric monitoring devices — this paper shows it is technically feasible with minimal performance cost.

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

ReadSaveConnections

DRIFT addresses a real clinical frustration: depression trials measure dozens of symptoms across multiple scales, but most statistical methods collapse these into a single score and miss patients who improve on some dimensions but not others. The framework uses factor analysis to discover latent dimensions of improvement and then finds treatment rules that are robust to uncertainty about which dimensions matter most — validated on the EMBARC MDD randomised trial. Identifying which depressed patients benefit from which antidepressants based on multi-dimensional response profiles is a central goal of precision psychiatry, and this provides a statistically rigorous approach to it.

██████████ 0.8 treatment-resistant-depression Preprint

ReadSaveConnections

This paper builds personalised computational brain models for individual Parkinson's patients by fine-tuning a foundation model pre-trained on over 5,600 brain scans, then asks the model counterfactual questions — 'what would this brain look like if it were healthy?' — to generate biomarkers that predict whether deep brain stimulation will work (AUPR = 0.915). The framework transfers to predicting response to a non-invasive alternative (temporal interference stimulation) without retraining. While the target is Parkinson's, the methodology is directly applicable to treatment-resistant depression, where predicting DBS or TMS response before surgery is an unsolved problem.

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

ReadSaveConnections

TAAC identifies that audio-based depression detection systems leak speaker identity information — meaning a diagnosis model inadvertently creates a re-identifiable voice profile of the patient — and proposes a framework that separates depression-related acoustic features from identity-related ones using adversarial training. The system outperforms encryption-based approaches on both depression detection accuracy and identity protection across three standard datasets. Privacy is a real deployment barrier for voice-based mental health apps, and a technical solution that doesn't sacrifice diagnostic accuracy could matter for regulatory approval.

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

ReadSaveConnections