Maximin Lange

Background: Unspecified psychosis-spectrum diagnoses (ICD code F29) are commonly used when the underlying disorder remains uncertain, particularly at first presentation. These diagnoses may persist or later be revised as the clinical picture evolves, yet little is known about their long-term trajectories or implications for treatment and service use in routine mental healthcare. Objectives: To describe the usage and persistence of unspecified non-organic psychosis (F29) and other non-specific diagnoses, and the time to transition to a specific diagnosis. Secondary objectives are to describe temporal trends in their use, identify demographic and clinical characteristics associated with diagnostic persistence, and characterise treatment patterns, service utilisation, mortality, and discharge across diagnostic groups. Methods: We will conduct a retrospective longitudinal cohort study using de-identified electronic health-record data from the South London and Maudsley NHS Foundation Trust (SLaM), accessed through the Clinical Record Interactive Search (CRIS) system, between 1 January 2007 and 31 December 2024. The study population will include patients with an ICD-10 psychosis-spectrum diagnosis (F2x.x), substance induced psychotic disorder (or F1x.5), bipolar disorder or affective disorder with documented psychotic symptoms(F32/3.3). Diagnoses will be classified as specific or non-specific according to their ICD-10 category. Primary analysis will examine time from first F29 diagnosis to first subsequent specific psychosis-spectrum or bipolar diagnosis using Cox proportional hazards regression, with discharge from SLaM, death and administrative end of follow-up treated as censoring events. Associations with diagnostic transition will be examined adjusting for age, sex, ethnicity, area-level deprivation, route of service entry, and calendar period. Secondary competing-risk analysis will estimate the cumulative incidence of diagnostic transition using the Aalen– Johansen estimator, treating discharge and death as competing events. Ethics and Dissemination: CRIS operates under ethical approval from the South Central–Oxford C Research Ethics Committee (23/SC/0257). Project-specific approval was received from the CRIS Oversight Committee as Project 23-095. The study will be reported in accordance with the RECORD extension of the STROBE guidelines. Findings will be disseminated through peer-reviewed publication and will provide evidence on the longitudinal use and refinement of non-specific psychosis-spectrum diagnoses in routine secondary mental healthcare.

Background: Antipsychotic medications are recommended for schizophrenia spectrum disorders, yet their long-term effects on functional recovery remain unclear, with conflicting evidence often derived from between-subject comparisons vulnerable to confounding by indication. Methods: We conducted a nationwide register-based cohort study of 65,630 individuals with incident schizophrenia spectrum disorders in Denmark (1998–2023). We modeled antipsychotic exposure against ‘productive engagement’ (employment or education). We used two analytical approaches: (1) within-subject stratified Cox models with time-varying covariates to eliminate time-invariant confounding; and (2) Fine–Gray competing risks models with baseline exposure, accounting for mortality and emigration. Results: Over 26.9 million person-weeks, the overall productive engagement rate was 48.2%. Integration of hospital pharmacy data revealed 6.1% exposure misclassification in studies relying solely on community records. The primary within-subject analysis revealed significant temporal heterogeneity: medication use was associated with reduced engagement rates in the acute (0–2 years; HR = 0.908) and consolidation phases (2–5 years; HR = 0.946), but reversed to a small positive association in the maintenance phase (5+ years; HR = 1.019). The between-subject Fine–Gray model, which estimates cumulative engagement probabilities, yielded an SHR of 1.002 (95% CI = 0.988–1.015), a population-level average that obscured these phase-specific dynamics. Conclusions: Antipsychotic pharmacotherapy exerts a time-dependent, biphasic impact on vocational recovery. We identified a window of vulnerability during the post-acute ‘consolidation’ phase (years 2–5) where treatment is associated with a transient reduction in productive engagement, before becoming protective after 5 years. These findings challenge the assumption that symptomatic stability automatically facilitates functional reintegration.

Introduction Smartphone ownership among UK adolescents is near universal, and teachers report phones increasingly being involved in classroom disruption, with misuse during school hours among the more common serious behavioural issues in secondary schools. Evidence on whether restrictive policies improve behaviour, attainment, or wellbeing remains limited. The primary objective is to assess the impact of a lockable smartphone pouch on educational attainment and behaviour. Secondary objectives are to assess impacts on general functioning, psychological wellbeing, and school-level indicators such as exclusions, and to examine whether effects differ for pupils who may be most at risk. Methods and analysis We will conduct a mixed methods cohort study in secondary schools across Northern Ireland and England during the 2025 to 2026 academic year. The quantitative component uses a serial cross-sectional design, with students completing an online questionnaire at 0, 4, and 8 weeks covering homework completion, classroom disruption, participation in PE and extracurricular activities, peer interaction, and smartphone use. Measures include the Strengths and Difficulties Questionnaire, the Revised Child Anxiety and Depression Scale, the short form of the Smartphone Addiction Scale, and the Bergen Social Media Addiction Scale. Schools will also supply half-termly aggregate data on exclusions, detentions, CAMHS referrals, counsellor visits, and parent visits from September 2023 to May 2026. Assuming 90% power, a two-sided type 1 error of 0.05, an intracluster correlation of 0.02, and 25% loss to follow up, we aim to recruit a minimum of 3,200 students from six or more schools to detect a small effect (Cohen's d = 0.2) on SDQ hyperactivity score. Continuous outcomes will be analysed with linear regression and binary outcomes with logistic regression. Aggregate school data will be analysed using an interrupted time series design. Prespecified subgroup analyses cover SEN or neurodivergent status, area-level deprivation, and existing school phone policy. Qualitative data from focus groups with students and staff and semi-structured interviews with school leads will be analysed thematically using Braun and Clarke's six-phase approach. Ethics and dissemination The study has been approved by the King's College London Research Ethics Committee. A Data Protection Impact Assessment has been agreed with the Northern Ireland Department of Education. Findings will be disseminated through a final report to the Department of Education, peer-reviewed publications, conference presentations, and accessible summaries for participating schools, pupils, parents, and policy makers. Keywords: Smartphone; Schools, Secondary; Adolescent Behavior; Mental Health; Academic Performance

We present BODHI (Balanced, Open-minded, Diagnostic, Humble, and Inquisitive), an engineering framework for curiosity driven and humble clinical decision support artificial intelligence (AI) systems. Despite growing capabilities, large language models (LLMs) often express inappropriate confidence, conflating statistical pattern recognition with genuine medical understanding. BODHI addresses this through a dual reflective architecture that: (1) decomposes epistemic uncertainty into task specific dimensions, and (2) constrains model responses using virtue based stance rules derived from a Virtue Activation Matrix. We validate the framework through controlled evaluation on 200 clinical vignettes from HealthBench Hard, assessing GPT-4o-mini and GPT-4.1-mini across 5 random seeds (2000 total observations). Statistical analysis included bootstrap resampling, paired t tests, and effect size computation. BODHI improved overall clinical response quality (GPT-4.1-mini: +16.6 pp, p<0.0001, Cohen's d=11.56; GPT-4o-mini: +2.2 pp, p<0.0001, Cohen's d=1.56) and achieved very large effect sizes on curiosity (context seeking rate: Cohen's d=16.38 and 19.54) and humility (hedging: d=5.80 for GPT-4.1-mini) metrics. Crucially, 97.3% of GPT-4.1-mini responses and 73.5% of GPT-4o-mini responses included appropriate clarifying questions, compared with 7.8% and 0.0% at baseline, demonstrating the framework's effectiveness in eliciting information gathering behaviour. Findings suggest LLMs can be reliably constrained to operate within epistemic boundaries when provided with structured uncertainty decomposition and virtue aligned response rules, offering a pathway towards safer clinical AI deployment

Background: Statistical Analysis Plans (SAPs) are essential for trial transparency and credibility but are resource-intensive to produce. While Large Language Models (LLMs) have shown promise in drafting protocols, their ability to generate high-quality, protocol-compliant SAPs remains untested against current content guidance. This study developed and validated an LLM-based pipeline for drafting SAPs from clinical trial protocols. Methods: We developed a structured, section-by-section prompting pipeline aligned with standard SAP guidance. We applied this pipeline to nine clinical trial protocols using three leading LLMs: OpenAI GPT-5, Anthropic Claude Sonnet 4, and Google Gemini 2.5 Pro. The resulting 27 SAPs were evaluated against a 46-item quality checklist derived from the published SAP guidelines. Items were double-scored by independent trial statisticians on a 0–3 scale for accuracy. We compared performance across LLMs and between item types (descriptive vs. statistical reasoning) using mixed-effects logistic regression. Results: Across 9 trials, the models produced SAP drafts with high overall accuracy (77%–78%), with no difference in performance between the three LLMs (p=0.79) but varied by content type (p < 0.001). All models performed well on descriptive items (e.g., administrative details, trial design), with lower accuracy for items requiring statistical reasoning (e.g., modelling strategies, sensitivity analyses). Accuracy for statistical items ranged from 67% to 72%, whereas descriptive items achieved 81% to 83% accuracy. Ǫualitatively, models were prone to specific failure modes in complex sections, such as omitting necessary details for secondary outcome models or hallucinating sensitivity analyses. Discussion: Current LLMs can effectively draft portions of SAPs, offering the potential for substantial time savings in trial documentation. However, a human-in-the-loop approach remains mandatory; while models demonstrate strong capability in producing descriptive content, their independent application to complex statistical methodology design still requires further methodological development and training. Future work should explore advanced prompt engineering, such as retrieval-augmented generation or agentic workflows, to improve reasoning capabilities.

Background: We present BODHI (Balanced, Open-minded, Diagnostic, Humble, and Inquisitive), an engineering framework for curiosity-driven and humble clinical decision support AI. Despite growing capabilities, large language models (LLMs) often express inappropriate confidence, conflating statistical pattern recognition with genuine medical understanding. BODHI addresses this through a dual-reflective architecture that: (1) decomposes epistemic uncertainty into task-specific dimensions, and (2) constrains model responses using virtue-based stance rules derived from a Virtue Activation Matrix. Methods: We validate the framework through controlled evaluation on 200 clinical vignettes from HealthBench Hard, assessing GPT-4o-mini and GPT-4.1-mini across 5 random seeds (1,800 total observations). Statistical analysis included bootstrap resampling, paired t-tests, and effect size computation (Supplementary Materials S3) Findings: BODHI significantly improved overall clinical response quality (GPT-4.1-mini: +17.3pp, p < 0.0001, Cohen’s d = 0.50; GPT-4o-mini: +7.4pp, p < 0.0001, Cohen’s d = 0.22) while achieving very large effect sizes on curiosity (context-seeking rate: Cohen’s d = 16.38 and 19.54) and humility (hedging: d = 5.80 for GPT-4.1-mini) metrics. Crucially, 97.3% of GPT-4.1-mini responses and 73.5% of GPT-4o-mini responses included appropriate clarifying questions, compared to 7.8% and 0.0% at baseline, demonstrating the framework’s effectiveness in eliciting information-gathering behavior. Interpretation: These findings suggest LLMs can be reliably constrained to operate within epistemic boundaries when provided with structured uncertainty decomposition and virtue-aligned response rules, offering a pathway toward safer clinical AI deployment.

Background: Antipsychotic medications are recommended for schizophrenia spectrum disorders, yet their long-term effects on functional recovery remain unclear, with conflicting evidence often derived from between-subject comparisons vulnerable to confounding by indication. Methods: We conducted a nationwide register-based cohort study of 65,630 individuals with incident schizophrenia spectrum disorders in Denmark (1998–2023). We modeled antipsychotic exposure against ‘productive engagement’ (employment or education). We used two analytical approaches: (1) within-subject stratified Cox models with time-varying covariates to eliminate time-invariant confounding; and (2) Fine–Gray competing risks models with baseline exposure, accounting for mortality and emigration. Results: Over 26.9 million person-weeks, the overall productive engagement rate was 48.2%. Integration of hospital pharmacy data revealed 6.1% exposure misclassification in studies relying solely on community records. The primary within-subject analysis revealed significant temporal heterogeneity: medication use was associated with reduced engagement rates in the acute (0–2 years; HR = 0.908) and consolidation phases (2–5 years; HR = 0.946), but reversed to a small positive association in the maintenance phase (5+ years; HR = 1.019). The between-subject Fine–Gray model, which estimates cumulative engagement probabilities, yielded an SHR of (95% CI = 0.988–1.015), a population-level average that obscured these phase-specific dynamics. Conclusions: Antipsychotic pharmacotherapy exerts a time-dependent, biphasic impact on vocational recovery. We identified a window of vulnerability during the post-acute ‘consolidation’ phase (years 2–5) where treatment is associated with a transient reduction in productive engagement, before becoming protective after 5 years. These findings challenge the assumption that symptomatic stability automatically facilitates functional reintegration. Danish Lay Summary: I denne undersøgelse har vi set på, hvordan personer kommer tilbage i arbejde eller uddannelse, efter de er startet i medicinsk behandling for skizofreni. Formålet er at forstå, om antipsykotisk medicin øger eller mindsker deres mulighed for at vende tilbage til arbejde eller uddannelse. Resultaterne peger på et udviklingsforløb, som kan forklares i tre faser: De første 2 år har en person i medicinsk behandling 9 % lavere sandsynlighed for at være i arbejde eller uddannelse, sammenlignet med perioder hvor personen er uden behandling. Efter 2–5 år bliver forskellen mindre (omkring 5 %). Medicinen forsinker stadig tilbagevenden til arbejdsmarkedet eller uddannelse, men i mindre grad end før. Efter 5–10 år vender billedet. Her ses en tendens til, at flere som får medicinsk behandling er i arbejde eller uddannelse. Effekten er lille, men den går i en positiv retning. På kort sigt kan antipsykotisk medicin begrænse muligheden for arbejde eller uddannelse. På længere sigt kan medicinen give den stabilitet, der gør det lettere at fastholde eller vende tilbage til arbejde eller uddannelse.

Employment is a crucial part of recovery for individuals with severe mental illness. Individual placement and support (IPS) is the gold standard for vocational rehabilitation, yet IPS reaches only a fraction of who could benefit. Large language models (LLMs) have been proposed as potential tools for vocational guidance, but their utility for vulnerable populations is unknown. We conducted an analysis of LLM‐generated job recommendations for individuals with schizophrenia spectrum disorders, and for a matched control cohort without psychiatric diagnoses. We used discharge summaries from 450 patients with a primary diagnosis of schizophrenia spectrum disorder and 50 control cases in the MIMIC‐IV database, fitting three independent job recommendations per case with Gemini 2.0 Flash and Claude Sonnet 4. Recommendations were summarised as a frequency and LLM‐automated content analysis was used to analyse reasoning patterns, workplace accommodations, and alignment with supported employment principles. Both, Gemini and Claude, showed little diversity and strong bias toward entry‐level roles. In the schizophrenia cohort, Gemini mostly recommended data entry and other clerical jobs while Claude produced a similarly narrow pattern with the majority suggesting library‐related. The controls revealed comparable clustering, with Gemini defaulting to clerical work and medical secretary roles, and Claude to customer service. There was limited diversity in the role settings, which almost uniformly suggested flexible schedules and minimal social interaction. Nor was there diversity in how roles were tailored to patient strengths, qualifications, or prior experience; instead, demographic stereotypes such as age‐based framing, gendered role allocation, and assumptions about language skills often shaped the recommendations. Based on our data and procedures, preliminary evidence does not support immediate deployment of LLMs for job recommendations for the tested population; further evaluation is needed after integrating human oversight and bias‐mitigation steps.

Background: No drug recommender system exists to guide antipsychotic medication selection for hospital in-patients with schizophrenia. This study developed and validated a prototype personalised medication recommender system for in-hospital patients with schizophrenia. Methods: This prognostic study analysed data from the Medical Information Mart for Intensive Care IV database on patients prescribed antipsychotic medications with hospital admission diagnoses of schizophrenic disorders between 2008 and 2022. Five similarity-based algorithms (two collaborative filtering and three distance-based representation methods) were evaluated using nested patient-wise cross-validation. The best performing model underwent external validation across geographic and temporal contexts using (a) Northwestern Intensive Care Unit and (b) Medical Information Mart for Intensive Care III datasets. Treatment success was measured through an affinity score (range 0–1) incorporating time between admissions, medication switches, and length of stay. Algorithm performance was assessed using root-mean-square error and mean average precision at position three, with visit-specific analysis across first, second, and third admissions. Results: Among 1,152 patients (660 first visits, 251 s visits, 241 third visits), cosine similarity-based collaborative filtering with K = 7 neighbours achieved optimal performance. Internal cross-validation demonstrated excellent prediction accuracy (root-mean-square error 0.116, standard deviation 0.007) and moderate recommendation quality (mean average precision at position three 0.490, standard deviation 0.029), improving to 0.697 (standard deviation 0.019) for first visits only. Geographic validation on 32 patients maintained recommendation quality (mean average precision at position three 0.432) despite increased prediction uncertainty (root-mean-square error 0.498). Temporal validation on 66 patients showed similar patterns (mean average precision at position three 0.359, root-mean-square error 0.578). Second-visit recommendations demonstrated exceptional transferability across validation contexts, achieving mean average precision at position three scores up to 0.813. In 163 cases where the model disagreed with attending clinicians, clinicians’ top-3 recommendations achieved significantly superior treatment outcomes (mean difference − 0.122 affinity points, 95% confidence interval: -0.145 to -0.098, p < 0.001, Cohen’s d=-0.81). Conclusions: This proof-of-concept study demonstrates feasibility of developing a machine learning-based medication recommender system for hospital in-patients with schizophrenia, that appear to maintain recommendation quality across different healthcare institutions and time periods. Prospective clinical trials are needed to establish real-world effectiveness and safety.