285

Papers

11/11

Roadblocks Active

1

Connections

Researchers built a Bayesian model that fuses EEG, eye tracking, and behavioral reaction-time data collected during implicit association tasks to classify people by suicidality or psychosis risk, reaching AUC 0.73–0.79 versus 0.50–0.53 for the standard scoring method. The key mechanism is a hierarchical structure that absorbs the large natural variation between individuals — a persistent obstacle in psychiatric biomarker work — while sparse regularization prevents overfitting to small clinical samples. This is one of the few papers today offering a statistically grounded, multi-modal approach to a hard classification problem in psychiatry, even if the sample sizes (n=34–39) are still too small for clinical deployment.

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

ReadSaveConnections

In 20 healthy adults, EEG recorded during a face-discrimination task showed that processing sad faces drives the brain into a significantly more 'stable' network state (lower Hopfield energy) than happy faces, with the strongest effect in the alpha frequency band (Cohen's d = 0.83). The Hopfield energy metric essentially measures how attracted the brain is to a particular activity pattern, and more negative energy means the network is locked into a tighter configuration — potentially a quantitative signature of the rumination-prone states seen in depression. This approach offers a physics-grounded, single-number readout of emotional brain states that could eventually serve as an objective depression biomarker, though replication in clinical samples is the obvious next step.

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

ReadSaveConnections

SynSym uses large language models to generate realistic synthetic training examples for identifying psychiatric symptoms (PHQ-9/DSM-5 depressive criteria) by first breaking each symptom into sub-concepts, then generating both clinical and colloquial phrasings, and finally simulating realistic co-occurrence patterns across symptoms. Models trained exclusively on this synthetic data matched the performance of models trained on real patient language, and fine-tuning on even small amounts of real data improved results further. This matters because real psychiatric text data is scarce and heavily restricted; a credible synthetic generation pipeline could unlock development of NLP-based screening tools without requiring large collections of sensitive patient records.

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

ReadSaveConnections

Across three widely used depression detection datasets, AI models trained on clinical interview recordings were found to use the interviewer's scripted prompts — not the patient's responses — as a primary signal for classifying depression. The fixed wording and timing of interviewer turns are predictable enough that a model can learn spurious patterns from them rather than genuine patient language. This finding is a significant methodological warning for the field: automated depression screening tools trained on these corpora may be detecting interviewer consistency rather than patient pathology, inflating apparent accuracy.

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

ReadSaveConnections

The DRIFT framework addresses a core problem in personalizing antidepressant treatment: how to estimate which individual patients will respond to a drug when outcomes are measured across many different symptom domains simultaneously. It combines statistical factor models (which compress symptom items into latent constructs like 'mood' or 'anhedonia') with an adversarial optimization step that ensures predictions remain valid even for outcome domains not seen during training. Applied to the EMBARC sertraline trial for major depressive disorder, DRIFT outperformed existing individualized treatment effect methods on held-out outcomes including side effects and self-reported symptoms.

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

ReadSaveConnections

Audio-based depression detection systems inadvertently encode speaker identity information alongside depression-related acoustic features, creating a privacy risk when voice data is processed by third-party diagnostic services. TAAC addresses this by using adversarial training to separate a 'depression features' subspace from an 'identity features' subspace in the audio representation, enabling depression detection while actively stripping out information that could be used to re-identify the speaker. The framework reportedly outperforms standard encryption approaches on both detection accuracy and identity protection, which matters for any deployment of voice-based mental health screening at scale.

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

ReadSaveConnections

A randomized experiment found that exposure to upward social comparison content on social media (seeing others who appear more attractive or socially successful) produced a trend toward reduced appearance self-esteem, while lateral comparisons (seeing peers at a similar level) actually increased self-esteem. Religiosity did not buffer against these effects, contrary to the hypothesis. The finding is relevant to youth mental health because it provides controlled experimental evidence — rare in this literature — that the type of social comparison encountered online, not just time spent online, drives self-esteem effects.

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

Read

A deep learning model trained to automatically stage sleep from EEG in healthy individuals failed to generalize to patients with ischemic stroke, a population with disrupted sleep architecture. Gradient-based attention maps revealed the model was focusing on EEG regions that carry physiologically irrelevant information in the patient group — essentially, it learned shortcuts specific to healthy brain signals. This matters for psychiatric applications because sleep disruption is a core feature of depression, bipolar disorder, and PTSD; automated sleep staging tools trained on healthy data may produce systematically wrong outputs in the very patient populations that need them most.

██████████ 0.7 sleep-circadian-psychiatry Preprint

ReadSaveConnections

This paper proposes a graph neural network framework that simultaneously models how the brain's physical wiring (structural connectivity) and functional activity patterns (functional connectivity) relate across multiple spatial scales, from individual brain regions up to large-scale networks. The method outperforms existing approaches on predicting brain age, cognitive test scores, and disease classification. For mental health, the practical value is a richer brain fingerprint that could distinguish depression subtypes or track treatment response by capturing structure-function relationships that single-scale or single-modality analyses miss.

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

ReadSaveConnections

Using resting-state fMRI from 400 healthy participants, this paper applies a branch of mathematics called Hodge theory to decompose directed brain communication into components that dissipate energy versus components that persist stably in circuits. The key addition is a 'counterfactual' layer: the framework can simulate how brain network organization would change under a hypothetical perturbation — for instance, mimicking a drug intervention or a lesion — by treating these as mathematical changes to the energy landscape of network flows. For treatment-resistant depression and psychedelic research, this could provide a principled way to predict network-level effects of interventions before administering them, though the framework currently lacks inferential statistics.

██████████ 0.7 treatment-resistant-depression Preprint

ReadSaveConnections