Early detection of mental health risk in non-clinical, non-help-seeking populations is a persistent challenge. We investigate large language models (LLMs) for identifying individuals at high risk of mental illness using standardized speech tasks. LLMs perform poorly on spontaneous, non-declarative speech due to limited self-disclosure and the lack of suitable training data. To address this, we propose the Latent Emotion Proxy for Mental Health Detection with AI-integrated Decision-making (LEMon-AID), a multi-step prompting framework that leverages LLMs’ emotion profiling capabilities, integrated with output from machine learning models. This enables LLMs to link emotional features with mental health labels, significantly improving AUC from 0.55 to 0.67. While more improvement is needed, this novel hybrid design introduces layered reasoning that enhances performance while improving explainability and protecting data privacy. Our results demonstrate that emotion-aware prompting, augmented with cross-model integration, improves early risk detection and LLM interpretability.
Psychiatric trials typically define success as symptom reduction, yet patients prioritize meaning, vitality, and functioning. This imbalance undervalues experiential benefits, distorting clinical, policy, and FDA decisions. Psychedelic trials illustrate this, where robust well-being gains sometimes accompany modest symptom effects. We argue well-being should be elevated to co-primary endpoint status, propose a Delphi-developed consensus well-being instrument with rigorous validation, and outline guardrails preventing approval of therapies that worsen the underlying condition.
This scoping review maps literature for use of ketamine or esketamine with psychotherapy for depression or depressive symptoms. Forty-seven studies were identified including 4/47 (9%) randomized controlled trials, 4/47 (9%) open label trials, 27/47 (57%) observational or retrospective studies, and 12/47 (26%) individual case reports. Studies were categorized as Ketamine/Esketamine Combined with Psychotherapy (KCP/ES-KCP), where an altered state is not necessary for psychotherapy, or Ketamine/Esketamine Assisted Psychotherapy (KAP/ES-KAP), where an altered state and/or integration of the experience is central to the psychotherapy. The 47 included 23/47(49%) KAP, 1/47 (2%) ES-KAP, 7/47 (15%) KCP, 3/47 (6%) ES-KCP, 8/47 (17%) dual KAP/KCP and 1/47 (2%) dual ES-KAP/ES-KCP. Four studies could not be categorized. Separation into these categories was useful to more clearly report current state of the literature. Though previous reviews have suggested KAP is of benefit for depression, KAP studies were limited by lack of RCTs, lack of comparison groups without psychotherapy, and a large number of individual case reports. To date, fewer studies, but higher quality data support a KCP model, with 3 studies suggesting benefit to combining ketamine with cognitive behavioral therapy (CBT). Further, high quality research including accurate reporting and description of psychotherapy model is required.
Suicide is a major public health issue, and machine learning offers promising tools to identify individuals at heightened risk. However, conventional models often fail to optimize performance across population subgroups, which can lead to inconsistent detection and outcomes. Using 2013–2023 data from the U.S. National Survey on Drug Use and Health (NSDUH), this study introduces the Functionally Adaptive Interaction Regularization-Penalized Logistic Regression (FAIR-PLR) framework: a logistic extension of the linear FAIR framework that combines group-covariate interaction terms with subgroup-size-weighted elastic-net regularization. Benchmarked against standard logistic regression, pooled elastic net with uniform weights, a decision tree, and subgroup-specific PLR across single subgroups (age, sex, race, BMI, insurance, rurality) and cross-sectional strata of psychological distress and treatment, FAIR-PLR matches or improves upon every baseline in overall performance while narrowing subgroup-level gaps in AUC, true positive rate, and precision at top-risk thresholds. FAIR-PLR gives a subgroup-balanced tool for national-scale suicide risk stratification.
The causal role of established risk factors for premenstrual disorders (PMDs) remains unclear. We used Mendelian randomization (MR) to assess causality for eight known-risk factors identified through a literature review. Summary statistics for these risk factors were from genome-wide association studies (GWAS) with sample size ranging from 129,017 to 1.2 million, and for PMDs from a GWAS of 72,297 participants. Findings were validated using one-sample MR in LifeGene cohort (n = 5674-5937). In two-sample MR, genetic liability to smoking initiation (OR = 1.25 (1.10-1.43)), earlier menarche (OR = 0.94 (0.89-0.98) per year), and higher BMI (OR = 1.19 (1.05-1.34) per kg/m2) were associated with PMD risk. No causal association was indicated for anemia, childhood abuse, childhood asthma, diabetes, and endometriosis. In one-sample MR, point estimates for BMI (OR = 1.10 (0.88-1.38) per kg/m²) and earlier menarche (OR = 0.98 (0.83-1.15) per year) were directionally consistent with the two-sample MR findings, but the confidence intervals included null effects. However, a null association was observed for smoking (OR = 0.94 (0.77-1.16)). The two-sample MR supports causal role of earlier menarche, higher BMI, and smoking in risk of PMDs. The lack of replication in one-sample MR highlights the need for triangulating evidence from future well-powered and methodologically comparable studies to strengthen these findings.
Postpartum depression (PPD) is among the most common complications of childbirth, and identifying novel treatments is vital. We aimed to identify potential drug targets for PPD by integrating the plasma proteome, transcriptome and epigenome. We designed a comprehensive analysis pipeline involving two-sample Mendelian randomisation (MR) (for proteins), colocalisation (for coding genes), and summary-based MR (SMR) (for mRNA and DNA methylation) to identify potential therapeutic targets for PPD. Genetic data on the plasma proteome were obtained from 4907 aptamers in 35,559 Icelanders and 7596 proteins in 828 FinnGen participants. The PPD genome-wide association study data were sourced from the Psychiatric Genomics Consortium (PGC) (Ncase = 17,339, Ncontrol = 53,426). A two-step MR approach was used to assess whether brain imaging-derived phenotypes (IDPs) and metabolites from blood, brain and cerebrospinal fluid mediated the observed effects. Across the two proteome datasets, the genetically predicted levels of 18 plasma proteins were nominally significantly associated with PPD, and the expression of steroid receptor RNA activator 1 (SRA1), a regulator of steroid hormone signalling, was significantly associated with PPD. SRA1, angiotensinogen (AGT, a key mediator of the renin-angiotensin and stress-response system), and glycerol-3-phosphate phosphatase (PGP, involved in lipid metabolism and cellular stress) showed increased colocalisation. The methylation of SRA1 at cg02434007 in the brain was associated with increased expression of SRA1 and a high risk of PPD, which aligns with the positive effect of SRA1 gene expression on PPD risk. Isoleucine (mediation proportion: 5.8%, p = 0.042) from blood metabolites and the IDP ICA100 edge 442 (mediation proportion: 7.6%, p = 0.044) may play mediating roles. This study reveals that SRA1 is a novel therapeutic target for PPD, which enhances the understanding of its molecular aetiology and the development of therapeutic strategies.
Prompt engineering has the potential to enhance large language models’ (LLM) ability to solve tasks through improved in-context learning. In clinical research, the use of LLMs has shown expert-level performance for a variety of tasks ranging from pathology slide classification to identifying suicidality. We introduce clickBrick, a modular prompt-engineering framework, and rigorously test its effectiveness. Here, we explore the effects of increasingly structuring prompts with the clickBrick framework for a comprehensive psychopathological assessment of 100 index patients from psychiatric electronic health records. We compare the performance of two locally-run LLMs against an expert-labeled ground truth for a variety of successively built-up prompts for the extraction of 12 transdiagnostic psychopathological criteria. Potential clinical value was explored by training linear support vector machines on outputs from the strongest and weakest prompts to predict discharge ICD-10 main diagnoses for a historical sample of 1692 patients. We could reliably extract information across 12 distinct psychopathological classification tasks from unstructured clinical text with balanced accuracies spanning 71% to 94%. Across tasks, we observed a substantially improved extraction accuracy (between +19% and +36%) using clickBrick for the most reactive model. The comparison unveiled great variations between prompts with a reasoning prompt performing best in 7 out of 12 domains. Clinical value and internal validity were approximated by downstream classification of eventual psychiatric diagnoses for 1,692 patients. Here, clickBrick led to an improvement in overall classification accuracy from 71% to 76%. ClickBrick prompt engineering, i.e., iterative, expert-led design and testing, is critical for unlocking LLMs’ clinical potential. The framework offers a reproducible, explainable pathway for deploying trustworthy generative AI across mental health and other clinical fields.
The goal of this study was to determine whether the number of connective tissue features in hypermobility is associated with the level of neurodivergent characteristics and establish whether autonomic reactivity may be an explanatory factor in the relationship between variant connective tissue and neurodivergent characteristics. 99 adult participants were assessed for joint hypermobility syndrome/hypermobile Ehlers-Danlos-Syndrome and filled out screening questionnaires for autism and ADHD. 99% of participants met criteria for generalised joint hypermobility, and 57% for hypermobile Ehlers-Danlos-Syndrome. 47% of participants scored above the screening threshold for autism, and 20% for ADHD. All measures were significantly correlated. Level of autonomic reactivity (as measured by the Body Perception Questionnaire) mediated the relationship between the number of connective tissue features and neurodivergence, even after controlling for anxiety level. This shows that autonomic reactivity has a potential mechanistic role in the established link between variant connective tissue and neurodivergence, opening novel pathways for research and clinical care.
Anorexia nervosa (AN) involves extreme food restriction and body image disturbances, partly sustained by altered responses to food cues, including implicit and behavioral avoidance, especially toward high-calorie foods. We tested whether altering body ownership in virtual reality could modulate such biases. Female participants with restricting-subtype AN (AN-R; n = 29) and healthy controls (HC; n = 31) completed three sessions: a baseline and two full-body illusion (FBI) sessions in which they embodied slimmer or larger avatars than their own body size. Across sessions, they performed a binocular rivalry task and a food-specific approach-avoidance task with high- and low-calorie foods, together with body-size estimation and symptom measures. At baseline, AN-R participants showed greater perceptual salience and stronger avoidance of high-calorie foods than HC. Relative to the baseline, embodying a larger avatar attenuated avoidance in AN-R, whereas HC showed the opposite pattern. These findings suggest that the FBI can transiently modify implicit food-related avoidance in AN-R.
Healthcare worker resilience is essential to building effective, functional, and crisis-ready health systems. This systematic map aimed to provide a comprehensive overview of the global evidence and identify gaps in resilience interventions for healthcare workers. A framework of interventions and outcomes was prepared with the help of advisory group members. The protocol was registered with the Open Science Framework. The search records were deduplicated and subjected to two levels of screening. We mapped 587 studies in the intervention-outcome framework. Most studies were conducted in high-income countries (473, 80.58%) and among nurses (284 studies, 48.38%), followed by other categories of healthcare workers. The most frequently included interventions were resilience training (240, 40.89%) and training and mentorship in the workplace (236, 40.20%). Artificial Intelligence-based or systems-thinking interventions (31, 5.28%) and herbal/traditional or conventional pharmacological interventions (4, 0.68%) were the least studied. Psychological outcomes were the domain of outcomes assessed most frequently (521, 88.76%). Patient safety (7, 1.19%), systems/policy level outcomes (16, 2.73%), and performance/ presenteeism (20, 3.41%) were the least measured outcomes. Overall, this systematic map provides evidence and gaps in resilience interventions for healthcare workers and can be used for planning and conducting future primary studies or systematic reviews.
Anxiety, depression, and stress are common in individuals with diabetes and cancer and are associated with poor self-management and well-being. Cognitive behavioral therapy (CBT) and mindfulness-based interventions (MBIs) are widely used in nonpharmacological treatments, but their comparative effectiveness remains unclear. We conducted a stratified subgroup meta-analysis to indirectly compare CBT and MBIs in adults with diabetes or cancer, examining moderators including condition, delivery modality, and dose. Following PRISMA guidelines, we searched Scopus, Web of Science, EBSCOhost, and ScienceDirect through October 2025. Randomized trials (N = 107; n = 23,585) reporting validated post-intervention outcomes were pooled using random-effects models, with subgroup and meta-regression analyses performed in RStudio. Overall, CBT and MBIs significantly improved outcomes (SMD = -0.78). CBT showed larger effects, particularly for depression (SMD = -0.95) and diabetes (SMD = -1.21), although comparisons were indirect and heterogeneity was substantial. Stronger effects were observed in group-based interventions lasting ≥8 weeks, providing ≥8 contact hours, and including homework or booster components; higher session frequency was also associated with greater improvement (β = -0.086; p < 0.01). MBIs may offer scalable benefits, particularly in cancer care. These findings support condition-tailored psychosocial strategies, although results should be interpreted cautiously.
Mental health mobile applications are increasingly used to address psychological distress and promote well-being, yet evidence on how and when these tools exert their effects over time and what constitutes meaningful engagement remains mixed. This study examined whether use of the Social Activity Guardian and Intervention Project (SAGIP) mental health mobile application predicts differential growth trajectories in psychological well-being and psychological distress among members of an academic community, and whether app engagement explains individual differences in these trajectories. Using an experimental repeated-measures design, participants from a university academic community were randomly assigned to a SAGIP app group or a waitlist control group and assessed at baseline (Month 0) and monthly for three months (Months 1 to 3). Psychological well-being and psychological distress were analyzed using latent growth curve modeling and mixed-effects piecewise growth models to capture nonlinear change. Compared to waitlist controls, SAGIP users demonstrated significantly greater improvements in psychological well-being and larger reductions in psychological distress during the randomized phase. Piecewise models indicated phase-specific change, with distress reduction strongest early in the intervention and well-being gains emerging more gradually and strengthening later. Frequency and duration of app use were associated with baseline mental health status but did not predict growth trajectories. Supplementary analyses provided additional support for the robustness of the findings: descriptive post-randomized analyses suggested that the experimental group generally maintained gains in psychological well-being and stabilized psychological distress below baseline levels through Month 6, whereas the control group showed modest improvements in well-being and distress after crossover to app access. Sensitivity analyses restricted to student participants replicated the primary pattern of intervention effects, while exploratory moderation analyses found no clear evidence that intervention effects differed between students and employees. These findings suggest that the therapeutic impact of self-guided digital interventions may depend more on the timing and quality of engagement than on the quantity of use and highlight the importance of modeling nonlinear change when evaluating digital mental health interventions.
Many patients do not improve during their first round of psychotherapy, partly due to variability in the quality of the patient-provider relationship. Matching patients to providers based on providers' historical performance with similar patients offers a proactive, data-driven solution. In a real-world retrospective cohort study, 24,303 participants using a mental health benefit (Spring Health) from 2021-2024 chose either matched therapists (i.e., based on providers' continuity of care, their historical patient-improvement percentile, and clinical alignment) or self-selected a therapist from a general list. Patients with depression using matched rather than self-selected therapists improved 8.5% faster (b = 0.12-point reduction on PHQ-9 per log-day, [95% CI: -0.15 to -0.09], p < 0.001) and were more likely to achieve reliable change (OR = 1.09 [1.05-1.14], p < 0.001), with similar results for anxiety and PTSD. Higher patient-provider fit scores predicted stronger therapeutic alliance (b = 0.05 [0.04-0.06], p < 0.001). Matching also reduced the total cost of care by 12.9% per reliable change and 10.8% per recovery. Thus, a data-driven patient-provider matching algorithm that is informed by providers' historical records with similar patients can provide modest but consistent acceleration of symptom improvement, increase clinical outcomes, and yield cost savings.
Psychiatric disorders are characterized by familiarity, which could also be implicated in the process leading to suicide attempts. This study aimed to evaluate the associations between sociodemographic, clinical, and psychopharmacological factors and previous suicide attempts in individuals with psychiatric disorders and multi-affected families. Separate networks for the sociodemographic, clinical, and psychopharmacological domains were estimated using Gaussian graphical models. The most influential variables were then combined to estimate an overall network. Results indicated that previous suicide attempts were primarily associated with alcohol abuse, female sex, and a younger age of onset of psychiatric disorders. The sociodemographic network revealed that age, sex, and marital status were directly related to a previous suicide attempt. In the clinical network, alcohol abuse and peripartum complications were central, while mood stabilizer therapy duration emerged as a significant factor within the psychopharmacological network. These findings underscore the importance of a multidimensional approach in suicide risk assessment to address the complex determinants of suicidal behavior effectively.
The concept of Sabr (patient endurance) is central to Islamic spirituality yet remains significantly under-theorized in mental health research. In this Perspective, I argue that Sabr offers a culturally grounded framework for understanding resilience among Palestinian Muslims facing collective trauma in contexts of occupation, displacement, and recurrent violence. Drawing upon Islamic theological sources, trauma psychology, and Palestinian studies, I propose a three-dimensional model of Sabr: (1) Sabr al-‘ibadah (endurance in devotion), (2) Sabr ‘an al-ma’siyah (restraint from prohibited responses), and (3) Sabr ‘ala al-musibah (endurance through affliction). I suggest that this framework addresses critical gaps in culturally responsive mental health care by honoring indigenous knowledge while engaging clinical science. The model offers implications for spiritually integrated assessment, intervention design, and provider training when working with Muslim populations affected by collective trauma. By centering Islamic conceptualizations of suffering and resilience, this Perspective contributes to broader efforts to incorporate diverse spiritual epistemologies into global mental health research and aligns with global health priorities such as SDG 3: Good Health and Well-Being for conflict-affected populations.
Ghrelin, a hunger-related gut hormone, may contribute to higher risk of depressive symptoms in obesity. Despite animal studies suggesting antidepressant effects of circulating ghrelin, human studies remain inconclusive. Therefore, we aimed to explore the association between obesity, ghrelin serum levels, and depressive symptoms in a large population-based cohort. Assessments of the LIFE-Adult cohort (n = 6037, 18-82 years) included questionnaires to evaluate depressive symptoms (CES-D and IDS-SR), anthropometric measurements for BMI, fasting ghrelin serum levels via radioimmunoassay (n = 1089), and 3 T MRI for hippocampal volume (n = 1080). Statistical analyses were pre-registered ( https://osf.io/y7sbx ). Higher BMI predicted more frequent depressive symptoms (β = 2.033, p < 0.001) and lower fasting serum ghrelin (β = -0.622, p < 0.001). Ghrelin did not correlate with depressive symptoms in obesity (n = 263, β = 0.123, p = 0.918). Exploratory analyses revealed links between ghrelin and eating-related depressive symptoms, and that higher BMI was more strongly associated with depressive symptoms in females than males. In this large, well-characterized sample, ghrelin was not associated with overall severity of depressive symptoms in participants with obesity. Future studies using more specific ghrelin assessments and clinical samples could help clarify this relationship.
Treatment‑resistant depression (TRD) is one of the toughest clinical challenges in psychiatry, characterized by high recurrence, heavy disease burden, and elevated suicide risk. Neuroimaging studies have mainly focused on single‑modality data, overlooking interactions between brain structure and function. This cross‑sectional study integrated multimodal MRI to examine alterations of structure-function coupling (SFC) and their associations with symptoms and diagnostic potential in TRD and non‑treatment‑resistant depression (nTRD). A total of 72 TRD patients, 152 nTRD patients, and 84 healthy controls were recruited. SFC was computed for each brain region from whole‑brain structural and functional data, and group differences, symptom correlations, and diagnostic classification were analyzed. TRD patients showed marked SFC decoupling in the right middle frontal gyrus, left inferior parietal lobule, left precentral gyrus, and right superior temporal gyrus. In nTRD, higher hippocampal SFC correlated with suicidal ideation and despair. Machine‑learning models based on SFC achieved high accuracy in distinguishing TRD from nTRD, outperforming previous unimodal approaches. These findings indicate that altered structure-function coordination represents a specific neural phenotype of TRD, linking network‑level decoupling with clinical symptoms and supporting its potential as an imaging‑based biomarker for individualized treatment. Trial Registration: ChiCTR2200055320, https://www.chictr.org.cn/showproj.aspx?proj=132558 . Registration date: January 1, 2022.
Abstract Smartphone-delivered interventions offer a scalable solution for university students experiencing depression, yet outcomes remain inconsistent. This study examined predictors of depression remission and response in the AI-enhanced Vibe Up adaptive trial of brief smartphone-based interventions in 1282 Australian university students (mean age 23.52 years; 78.39% women) with elevated distress (Kessler-10 ≥ 20). After a two-week monitoring period, participants were randomised to two-week self-guided smartphone-based interventions targeting sleep hygiene, mindfulness, physical activity, or an ecological momentary assessment control intervention. Predictors of depression remission (DASS-21 in normal range) and response (≥50% reduction) were examined using hierarchical logistic regression. At post-intervention, 40.87% achieved remission and 29.88% showed response. Baseline depression severity, quality of life, general practitioner visits, and pre-intervention credibility predicted remission and/or response across all intervention arms. Higher baseline anxiety specifically predicted poorer remission in the sleep hygiene arm. Individual factors modestly predicted outcomes. Trajectory-based predictors may be needed to improve outcome prediction.
