251

Papers

10/10

Roadblocks Active

1

Connections

This framework combines two types of brain MRI — structural anatomy and resting-state functional activity — and uses bidirectional cross-attention so each modality can inform interpretation of the other during depression classification. It consistently outperforms simpler approaches that just concatenate features from the two scans, demonstrating that the spatial relationship between structural and functional brain abnormalities is itself diagnostically meaningful. For mental health researchers this matters because it shows that modeling how brain structure and function interact — not just measuring each separately — captures additional signal relevant to MDD diagnosis.

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

ReadSaveConnections

A three-stage AI pipeline uses GPT-o3 to generate progressively detailed clinical summaries that guide a multimodal model combining text, audio, and video to detect depression — outperforming current state-of-the-art on the E-DAIC and CMDC benchmarks across binary screening, five-level severity classification, and continuous score regression. The LLM-generated summaries act as an intermediate reasoning layer, making model decisions more auditable than typical black-box fusion approaches. The system's full dependence on a closed, paid API (GPT-o3) whose outputs are non-deterministic is a real reproducibility concern that must be resolved before clinical translation.

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

ReadSaveConnections

A virtual brain model pretrained on 2,707 subjects was fine-tuned on Parkinson's patients and used to predict treatment response to transcranial stimulation (AUPR 0.853) and deep brain stimulation (AUPR 0.915) by quantifying how far each patient's brain dynamics deviate from a simulated healthy norm. The approach generates individualized biomarkers non-invasively from resting-state fMRI and was validated in external and prospective cohorts. The counterfactual modeling strategy — estimating the gap between a patient's pathological state and a healthy simulated state — is directly transferable to treatment-resistant depression, where biomarker-guided intervention selection is a major unmet need.

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

ReadSaveConnections

A hierarchical Bayesian model combining neural signals with behavioral reaction-time data from Implicit Association Tests doubled classification accuracy for suicide-related ideation (AUC 0.73) and psychosis risk (AUC 0.76) compared to the standard D-score method, with further improvement to AUC 0.79 when restricted to MDD participants. The IAT probes implicit associations that patients may not consciously disclose, making this approach complementary to self-report. This matters for mental health because high-risk states like suicidal ideation are frequently concealed — objective multimodal detection methods could fill a critical clinical gap.

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

ReadSaveConnections

Models trained exclusively on what the interviewer says — not the patient — achieve surprisingly high depression classification scores on three widely-used benchmark datasets (ANDROIDS, DAIC-WOZ, E-DAIC), because structured interviews have fixed, predictable prompts that inadvertently encode diagnostic signal. The effect is architecture-agnostic, appearing in both Longformer and graph-based models, suggesting it is a structural property of the datasets rather than a modeling artifact. This is a serious methodological warning: published accuracy numbers for many depression detection systems may reflect exploitation of interviewer scripting rather than genuine capture of patient psychopathology.

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

ReadSaveConnections

TAAC separates depression-related speech features from identity-revealing features in audio using adversarial subspace decomposition and noise encryption, addressing the privacy risk that depression detection tools simultaneously expose who a patient is. The framework is designed to enable accurate detection while protecting patient identity, tackling a dual objective that single-purpose systems ignore. For clinical AI deployment, demonstrating both diagnostic efficacy and identity privacy protection is increasingly a regulatory and ethical prerequisite — this work provides an architectural approach to meeting both requirements.

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

ReadSaveConnections

The DRIFT framework estimates personalized treatment effects from high-dimensional psychiatric symptom questionnaires by extracting latent constructs via factor analysis and optimizing for worst-case performance across unmeasured symptom domains, validated on the EMBARC sertraline-versus-placebo randomized trial for major depression. Unlike approaches that average across patients or rely on observed summary scores, DRIFT generates predictions robust to the fact that standard rating scales never capture the full breadth of what matters to an individual patient. This directly addresses a core challenge in treatment-resistant depression: selecting which intervention is most likely to benefit a specific person rather than a statistical average.

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

ReadSaveConnections

Misty Forest uses real-time ADHD attention patterns as input mechanics that drive collaborative gameplay in a VR environment, reframing attentional variability as a functional asset that creates shared momentum between youth rather than a deficit to be suppressed. This strength-based design philosophy contrasts sharply with conventional ADHD interventions that focus on reducing inattention, and may improve engagement and acceptability in this population. As a digital therapeutic concept it represents a proof-of-concept that leveraging neurodivergent cognitive styles in designed environments can enable collaboration rather than requiring normalization.

██████████ 0.8 digital-therapeutics 🔗 1 cited Peer-reviewed

Read

SynSym uses LLMs to generate synthetic psychiatric symptom descriptions in diverse linguistic styles through a four-stage pipeline that expands sub-concepts, generates dual-style expressions, models symptom co-occurrence, and quality-filters outputs — with models trained only on synthetic data matching the performance of real-data-trained models on three benchmark depression datasets. Fine-tuning the synthetic-trained model on even small amounts of real clinical text improves performance further. This matters practically because real patient data is scarce, sensitive, and hard to share; a validated synthetic training pipeline substantially lowers the barrier to building and evaluating depression detection tools without compromising patient privacy.

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

ReadSaveConnections

Using EEG from 20 healthy adults, the study found that processing sad faces produces significantly more stable (lower energy) brain network states than happy faces, with the strongest effect in the alpha frequency band (Cohen's d = 0.83), and that these more stable states correspond to higher network connectivity efficiency (r = -0.72). The Hopfield energy framework translates complex network dynamics into a single interpretable scalar per emotional state, grounding emotional valence in dynamical systems theory. For depression research this provides a computationally tractable biomarker framework — pathologically stable negative-valence brain states could be a measurable signature of depressive rumination — though the current sample of 20 healthy participants limits direct clinical inference.

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

ReadSaveConnections