Biostatistics

Causal mediation analysis is widely used to identify biological pathways linking exposures to outcomes, but most methods assume homogeneous mediation effects across individuals. In high-dimensional omics settings, this assumption can mask important heterogeneity driven by demographic, genetic, or environmental factors. We propose the M-high-learner, a flexible framework for detecting heterogeneous mediation effects with high-dimensional mediators. The method identifies mediators with subgroup-specific indirect effects while distinguishing them from null or homogeneous signals and controlling the type I error rate. It is computationally efficient, scalable, and yields interpretable sub-types. Simulation studies show that the proposed approach achieves high power while maintaining accurate error control. Applications to the Framingham Heart Study and the Multi-Ethnic Study of Atherosclerosis reveal that the mediation role of gene expression in sexs effect on high-density lipoprotein varies across subgroups defined by body mass index and age. Our framework provides a practical tool for uncovering heterogeneous biological mechanisms in high-dimensional genomic studies. Author SummaryBiological processes linking risk factors to disease often differ across individuals, but many existing methods assume these processes are the same for everyone. This can hide important differences between groups. We developed a powerful method to identify when these pathways vary across subgroups using large-scale molecular data. Our approach detects differences in how intermediate biological factors contribute to outcomes in populations defined by characteristics such as age and body mass index. Applying our method to population studies, we found that some biological pathways operate differently across groups, suggesting that key mechanisms may be missed when differences are ignored. Our work provides a tool to better understand how disease-related processes vary across individuals, which may support more targeted and personalized approaches to health research.

Background. Patient stratification from multi-omics and clinical data is essential for uncovering disease heterogeneity and moving toward more personalized treatment strategies. However, integrating heterogeneous data layers while identifying robust patient strata remains challenging. Methods. We introduce Reduced Fusion of Multi-Omics Stratification (RedFuMOS), a novel three-step approach for patient stratification based on mixed-type multi-omics data. RedFuMOS extends Similarity Network Fusion to accommodate mixed-type data layers and layer-specific similarity measures for data integration, includes a dimensionality reduction step to mitigate the curse of dimensionality, and performs patient stratification using density-based hierarchical clustering with HDBSCAN. It also implemented an automated optimization procedure to identify the best set of hyperparameters, minimizing the need for manual tuning. Results. RedFuMOS outperformed six state-of-the-art tools for multi-omics patient stratification in a comprehensive simulated benchmarking study, which also confirmed that, although computationally expensive, the dimensionality reduction step is crucial for achieving good stratification performance. Additionally, RedFuMOS identified two clinically relevant patient strata in a small real-world cohort of patients with Philadelphia chromosome-positive chronic myeloid leukaemia. Conclusion. RedFuMOS provides a flexible framework for integrating heterogeneous multi-omics and clinical data. RedFuMOS is available as an R package at http://github.com/delucasara/RedFuMOS.

Standard polygenic risk scores (PRSs) are constructed based on additive genome-wide association study (GWAS) summary statistics. Nonlinear machine learning methods have been increasingly applied to construct PRSs directly from individual-level data, with the aim of improving predictive performance over standard PRSs through their ability to model non-additive genetic effects. However, their superiority across studies has been inconsistent, and the conditions under which they provide meaningful improvements remain unclear. We combined theoretical analysis, simulations and a real-world application to investigate when two widely used nonlinear machine learning methods, random forest and XGBoost, outperform standard PRSs. Theoretical analysis showed that standard PRSs can implicitly capture part of the genetic variance attributable to nonadditive genetic effects through their contributions to marginal SNP effects, thereby losing less information than commonly assumed. Although nonlinear models have a higher theoretical potential, their greater flexibility incurs a bias-variance trade-off that can limit predictive gains at finite sample sizes. Simulations showed that XGBoost outperformed the standard PRS only when the genetic architecture involves a sufficiently large proportion of interaction genetic variance concentrated across relatively few interaction effects and large training samples were available. Random forest consistently underperformed the standard PRS. In an application to ischemic heart disease prediction using UK Biobank data, XGBoost showed no meaningful improvement in predictive performance over the standard PRS, whereas random forest again performed worse. Together, these findings suggest that nonlinear machine learning do not uniformly outperform standard PRSs; rather, their relative performance depends jointly on genetic architecture and training sample size. Our study helps to reconcile the inconsistent results reported across previous studies and provides a framework for identifying settings in which more complex PRS models are likely to be beneficial.

We describe LRSPAT (low-rank spatial toolkit), a fast and memory-efficient framework for approximating measures of spatial association for high-dimensional data. While LRSPAT can be applied to any multivariate spatial dataset, development was motivated by the computational challenge of identifying spatially variable genes in high-resolution spatial transcriptomics (ST) data generated by technologies such as 10x Visium HD, Xenium and Atera. LRSPAT leverages a truncated SVD of the expression data and a thresholded spatial weights matrix to perform reduced-rank reconstruction of spatial statistics in the quadratic form family, including global and local versions of Moran's I, Geary's C, and Getis-Ord G. A regularization approach is leveraged to account for the inflated null distribution of spatial statistics computed on latent variables. By performing key operations on the low-dimensional embeddings, LRSPAT is orders of magnitude faster than standard implementations with significantly lower memory requirements. Because the low-rank approach denoises and desparsifies ST data, LRSPAT is also more accurate than standard techniques at identifying genes with true spatial expression patterns. The dramatic improvements in execution time and memory consumption enable the genome-wide analysis of spatially variable genes (SVGs) and exploration of the full range of hyperparameters including spatial scale, distance metric, and embedding rank. This preprint outlines the background and mathematical details of the approach with limited preliminary results and a short conclusion.

Constructing microbial association networks is a common strategy for exploring relationships among taxa in microbiome studies. Although marginal correlation methods are easy to implement and allow formal inference, they can produce spurious edges driven by indirect associations through other taxa. Conditional graphical-modeling methods aim to recover direct associations, but many rely on Gaussian or linear assumptions and often provide limited uncertainty quantification. We propose a conditional, nonparametric approach based on the scaled expected conditional covariance (SEcov). SEcov measures population-level conditional association by residualizing each taxon with respect to the remaining taxa and scaling the resulting expected conditional covariance. The resulting estimator can incorporate flexible machine-learning methods for conditional-mean estimation and admits asymptotic normal inference, enabling p-values and confidence intervals for taxon-pair associations. We demonstrate through simulation studies that our proposed approach improves network recovery relative to other methods, and we illustrate the new method via construction of a co-occurrence network for the vaginal microbiome during pregnancy. IMPORTANCEHigh-throughput sequencing has made it possible to characterize microbial communities at large scale, and network analysis is widely used to summarize relationships among taxa. However, networks based on marginal correlations may include indirect associations, whereas many conditional graphical models rely on assumptions that may be difficult to justify for sparse, zero-inflated, compositional microbiome data. SEcov offers a practical alternative by estimating conditional associations nonparametrically and attaching inferential uncertainty to individual edges. This allows investigators to construct microbiome networks using statistically interpretable evidence for taxon-pair associations, rather than relying solely on arbitrary correlation cutoffs or regularization tuning parameters.

Bayesian borrowing of external information can improve trial efficiency, particularly in pediatric and rare disease settings where patient populations are limited, but may introduce bias and inflate the Type~I error rate when the trial differs from external studies. Recent U.S. Food and Drug Administration (FDA) draft Bayesian guidance emphasizes careful evaluation of external information, prior specification, and assessment of operating characteristics. This paper compares three meta-analytic-predictive (MAP)-based methods for Bayesian borrowing: the MAP prior, robust MAP (RMAP) prior, and self-adapting mixture (SAM) prior. An adaptive platform trial design in Rett syndrome is used as a case study. Simulation studies evaluate frequentist operating characteristics under varying prior--data conflict, between-study heterogeneity, treatment effects, and clinically significant differences (CSDs) for the SAM prior. The MAP prior achieved the greatest efficiency when external and current data were compatible but exhibited the largest bias under substantial prior--data conflict. The RMAP priors improved robustness through fixed robust-component weights, whereas the SAM prior adaptively adjusted borrowing and was less sensitive to prior--data conflict while retaining efficiency gains when the data were compatible. Although the CSD influenced the degree of adaptive borrowing, as reflected by effective sample size, it had only a modest impact on frequentist operating characteristics. Sensitivity analyses using a skeptical robust component yielded similar qualitative conclusions, while accentuating the differences between the MAP and RMAP priors. These findings provide guidance for evaluating and selecting MAP-based borrowing strategies before trial implementation, particularly in rare disease settings, consistent with current FDA recommendations.

We propose two new tests to detect drug effects (DE) in trials of one to very few patients followed during two periods (before and after initiation of a treatment). Both methods use longitudinal natural history data to inform the estimation of each patient's DE. The first method uses a non linear mixed effect model (NLMEM) reflecting an expected natural history with a hypothetical drug effect, to estimate the Conditional Distribution of the Drug Effect (CDDE). The second method trains a Pareto Depth Analysis (PDA) algorithm, a machine learning based approach based on outlier detection, that we implement using data simulated under the NLMEM. We evaluated the two tests with a simulation study. We used data from the PROSPAX study in Autosomal Recessive Cerebellar Ataxias (ARCAs, to derive a NLMEM for the Scale for the Assessment and Rating of Ataxia score. The CDDE method provided controlled type I error and, in some scenarios, adequate corrected power, though sensitivity analyses showed vulnerability to misspecification. The PDA method demonstrated lower statistical power except with high score precision. These results highlight different strategies for quantifying treatment effects in ultra rare, patient' specific trials. They can inform methodological design for future ARCA precision therapies.

Acute kidney injury (AKI) is a frequent complication within intensive care units, with its sudden onset often missed. This is especially important because a timely window for intervention is required as delayed detection leads to progressively worse outcomes. Existing machine learning and deep learning models have contributed to closing this gap, but their complexity, requiring hundreds to thousands of features, and lack of generalisation pose a limitation that prevents them from being integrated into clinical workflows across different electronic health-record ecosystems. This study presents a 37-feature XGBoost model trained on the MIMIC-IV dataset with 5.4% positive cases, with hyperparameters optimised via Optuna and probabilities calibrated using isotonic regression, designed for transportability across clinical settings. Validation was conducted internally using a temporal patient-level split simulating prospective deployment, training on 2008-2016 data and testing on 2017-2022 data"External validation was performed on the eICU Collaborative Research Database, a multi-centre dataset spanning 208 US hospitals, using the trained model without retraining. SHAP TreeExplainer was used to provide feature-level explainability for individual predictions. Internal testing yielded an AUROC score of 0.794 for predicting AKI onset within a 12-24 hour window. External validation produced a 0.750 AUROC without retraining. Equitable discrimination was observed across gender, age, chronic kidney disease presence, race, and AKI stages on both datasets, with a 95% internal CI of 0.789-0.799 confirming the model's estimate stability. These results suggest that clinically useful prediction systems are achievable with substantially fewer features than current models require.

A dental AI system and a dentist reading the same radiographs form a paired comparison. Published comparative studies often report the two arms separately against a reference standard, leaving the joint pattern of correctness between them unavailable for secondary paired inference. We show what that omission costs. The accuracy difference remains exactly identified; its sampling variance does not, so the report contains the estimate and not its uncertainty. On a study of 282 units, two published accuracies are consistent with 38 distinct joint tables whose confidence intervals differ in width by a factor of 2.5. The consequence is a three-zone decision map rather than a single threshold: differences at or below 1.06 points are non-significant under every compatible table, differences at or above 6.03 points are significant under every compatible table, and in between the published numbers cannot decide. We then show the omission is repairable at negligible cost. One additional integer, the number of units both arms classify correctly, identifies the joint table exactly and restores standard paired inference. For a panel of readers the pairwise dependences must arise from one joint distribution, a constraint that binds once three readers are present; publishing each reader's joint-correct count against a single reference reader cannot widen and may tighten every pairwise bound, and in a 7-arm experiment reduced them by a median of 37% even for pairs excluding that reference. Where the integer was never published we give DentalPair-Cert, an interval with finite-sample coverage uniformly over every admissible within-unit AI-dentist dependence under the independent-sampling-unit model, certified in both the nuisance maximization and the inversion. Across 4,200,000 simulated comparisons an independence analysis falls to 74.5% coverage with 12.2% type-I error; in a purposive sample of 9 recent comparative studies, 1 reported a paired test on discordant units.

Unmeasured confounding threatens causal inference and replicability in observational multi-omic studies across variable environments. Genetic instrumental variables (Mendelian randomization) and negative-control calibration each address complementary sources of unmeasured confounding, yet no existing framework unifies them for omics-scale mediation analysis. We introduce ICONIC, an R package that embeds genetic instruments and negative controls within a proximal causal inference framework for total-effect and mediation analysis. ICONIC implements eight estimators spanning five confounding-control strategies, supports continuous, binary, and time-to-event outcomes, and provides extensive diagnostics including sensitivity analyses that map estimator performance across plausible assumptions. Ground-truth benchmarks are calibrated to real-omics covariance structures via a hybrid generative model (GAN + feature-level Gaussian copula) rather than parametric simulation, and a companion planning tool predicts performance gains from collecting additional omic data. We demonstrate ICONIC in two case studies: identifying placental transcriptomic mediators of gestational diabetes on birth weight (n = 164), and tumor-expression mediators of smoking intensity on lung cancer survival (n = 494). Notably, ICONIC's diagnostics recommended different estimation strategies across the two scenarios, reflecting differences in the likely influence of unmeasured confounding. ICONIC is freely available at https://github.com/sbresnahan/iconic/.

Improvements in data availability, sharing, and integration, together with the development of explainable artificial intelligence (XAI) techniques, are advancing precision medicine for pediatric cancer by facilitating diagnosis, biomarker discovery, and drug development. Data sharing commons and initiatives like the Childhood Cancer Data Initiative (CCDI) provide access to pediatric-specific genomic and clinical data cohorts and improve data availability for pediatric cancer research. Based on CCDI, a scalable AI platform, Graph Artificial Intelligence for Pediatric Oncology (GAIPO), integrates various data modalities from bulk and single-cell omics data to clinical information. Such multi-modal data facilitates the training and development of advanced XAI models for pediatric cancers. We then developed an end-to-end multi-modality framework, PCGS, for pediatric cancer by incorporating omics-specific representation learning via GNN models with cross-attention fusion and multi-objective learning for downstream tasks such as classification, clustering, and survival analysis. This framework outperforms previous supervised multi-omics integration baseline approaches based on glioma and Wilms tumor cohorts and enables GNN model explainability via Shapley value-based feature attribution approaches to explain the contributions of gene-level features across various biomedical tasks, including classification and survival. Given specific background samples (e.g., age groups, sex, grades) as baselines, this explainable GNN model estimates and ranks the importance scores for input features from each omics modality. It identifies background-specific key features for biomarker discovery, risk group identification, and survival analysis in glioma and Wilms tumor, with potential applicability to other pediatric cancers.

Background: Traditional real-world studies comparing SGLT2 and DPP4 inhibitors on renal outcomes rely on propensity score matching, which causes high-dimensional data loss. We used causal machine learning (Causal ML) to unmask heterogeneous treatment effects in diabetic kidney disease (DKD). Methods: Using data from 4,588 patients within the Japanese J-CKD-DB-Ex registry, we implemented a doubly robust (DR) learning framework (Linear DR-learner with XGBoost) to compare SGLT2 and DPP4 inhibitors. Outcomes included the chronic eGFR slope and a composite renal endpoint ([≥] 50% eGFR decline or end-stage kidney disease). Heterogeneity was explored via causal SHAP and decision trees. Results: At the population level, SGLT2 inhibitors modestly slowed chronic eGFR decline (average treatment effect [ATE] = 0.14 [95% CI: -0.86, 1.15] mL/min/1.73m^2/year) and reduced composite endpoint risk by 9% (ATE: -0.09 [-0.11, -0.08]) versus DPP4 inhibitors. However, individual-level counterfactual analysis suggested that for the chronic eGFR slope, non-glinide users with stable pre-treatment trajectories who were also taking ACE inhibitors had a greater benefit from SGLT2 inhibitors (ATE: 2.95 [-0.68, 6.58]). Conversely, glinide users with steep pre-treatment decline had a greater benefit from DPP4 inhibitors (ATE: -8.98 [-16.11, -1.85]). For composite renal events, SGLT2 inhibitors had a 28% absolute risk reduction within the algorithmically identified high-risk subgroup (eGFR [≤] 28.1 mL/min/1.73 m^2 and positive proteinuria; ATE: -0.28 [-0.33, -0.23]). Even non-proteinuric decliners demonstrated a 8% risk reduction with SGLT2 inhibitors (ATE: -0.08 [-0.10, -0.06]). Conclusion: Causal ML advances precision medicine in DKD, shifting from uniform prescribing to individualized, data-driven therapy targeting distinct intrarenal pathways.

Wearable (accelerometer) data and machine-learning allow objective assessment of the amount of daily physical activity. However, wearable-derived human activity is subject to measurement error. No studies have corrected the dose-response association between physical activity and survival time to chronic diseases, including cardiovascular disease (CVD). The objective is to estimate the measurement error-corrected association between CVD events and multiple measures of daily duration of light and total physical activity, derived from machine-learning and conventional accelerometer-processing methods. Our method combined an accelerated failure time model, spline, and simulation-extrapolation (SIMEX). The method recovered the true dose-response non-linear association in simulated data, while the naive model failed to capture it due to substantial attenuation. Application to the UK Biobank accelerometer cohort also showed an increased protective association of total physical activity after SIMEX correction (Time Ratio [TR] = 1.56, 95% CI: 1.28-1.82 vs. TR = 1.38, 95% CI: 1.24-1.54 for SIMEX-corrected vs. uncorrected dose-response association between the 95th and 5th percentiles of total activity), with a similar increase for light physical activity. Sensitivity analysis indicates that the female population experiences a substantially larger protective association after SIMEX correction than males. Dose-response survival analysis is a widely used analytical method in physical activity epidemiology and benefits from measurement error correction.

Agentic workflows can coordinate modelling, but balancing predictive performance, measurement burden and reproducibility is unclear. We developed DXA Agent, an agentic workflow for dual-energy X-ray absorptiometry (DXA) outcomes integrating planning, feature-model refinement, tools, provenance and hypothesis-generating interpretation. Models were independently developed and tested in UK Biobank (5,318 participants) and the National Health and Nutrition Examination Survey (NHANES; 3,777 participants), using cost-efficient and no-limit strategies. Across 20 UK Biobank and three NHANES bone mineral density sites, cost-efficient models achieved lower RMSE and higher R2 than the best conventional comparator, with median relative RMSE reductions of 10.9% and 9.9%, respectively. Classification was task dependent: UK Biobank osteoporosis averaged AUROC 0.839 and PR-AUC 0.182, whereas NHANES performance was comparable with conventional models. Higher-burden features did not consistently improve prediction. These retrospective, cohort-internal findings position DXA Agent as an inspectable, measurement-burden-aware research workflow requiring independent prospective validation.

Active Feature Acquisition (AFA) adaptively selects which diagnostic test to order next and offers a route to reduce unnecessary laboratory testing in acute care. Existing clinical AFA evaluations, however, assume every feature can be retrieved on demand and split data at the visit level, both of which inflate apparent performance. We re-evaluate cost-aware AFA under constraints designed to reflect deployment. From MIMIC-IV we constructed a cohort of 64,766 acute admissions (39,884 patients; 21 conditions; 55 features in 30 test panels) with a patient-level split, a 12-hour decision cutoff, and a per-patient availability mask from what was actually measured, and priced panels using the 2026 Medicare fee schedule under panel-level billing. We evaluated EIG-Cost, which scores each panel by Monte-Carlo Expected Information Gain penalised by its dollar cost, against eight published methods across budgets \30--$60 over five patient-level resamples. At a $30 budget, EIG-Cost achieved the highest macro-F1 (0.188, 95% CI [0.185, 0.191]) at the lowest cost ($17.28), exceeding the strongest baseline in all five resamples (p<0.001; Cohen's d=4.0), and led at every budget. Three of the eight methods collapsed to a vitals-only baseline (macro-F1 approx 0.040), acquiring nothing even at higher budgets, a genuine failure to adapt to availability rather than a budget limitation. Despite modest absolute accuracy, EIG-Cost's probabilities were well-calibrated (expected calibration error $0.048$). Under realistic availability constraints, clinical AFA is substantially harder than full-availability benchmarks imply, several published methods fail outright, and cost-aware information-gain scoring is a robust choice in this harder setting.

Abstract Imaging-based spatially resolved transcriptomics (imSRT) technologies provide high-throughput molecular-resolution spatial characterization of genes within cells. Conventional analysis methods to identify cell-types and states in imSRT data rely on gene count matrices derived from tallying the number of mRNA molecules detected for each gene per segmented cell, thereby overlooking subcellular heterogeneity that can be useful in defining cell states. To take advantage of the molecular-resolution information in imSRT data and potentially identify cell-states based on subcellular heterogeneity, we developed STARIT (Spatial Transcriptomics As Rasterized Image Tensors). STARIT converts transcripts within segmented cells in imSRT data into an image-based tensor representation that can be combined with deep learning computer vision models for downstream analysis. Using simulated and real imSRT data, we demonstrate that STARIT distinguishes transcriptionally distinct cell-types and further separates cell states based on subcellular transcript localization, which conventional gene count analysis fails to capture. By providing a standardized framework to encode subcellular molecular information in imSRT data, STARIT will enable deeper insights into subcellular heterogeneity and enhance the identification and characterization of cell-types and states that are overlooked by gene count representations.

Inappropriate antibiotic use presents a major global health challenge, particularly in low-resource settings where access to quality care is limited but antibiotics remain relatively unrestricted. This study estimates the causal effect of frontline primary care quality on inappropriate community antibiotic use, combining detailed community-based data from approximately 100 rural villages in rural China with an instrumental variable (IV) approach embedded within a double/debiased machine learning (DML) framework. We linked objective measures of village doctor clinical practice quality, measured through unannounced standardized patient visits, to household-level antibiotic use data collected from the same villages. To identify the causal effect, we constructed multiple candidate instruments from extensive provider characteristics and used an ensemble of machine learning algorithms within a flexible DML-IV framework to approximate an optimal instrument, addressing a many-weak-instruments problem. We found that improving village provider clinical practice quality reduced both antibiotic receipt during healthcare encounters for common diseases and household antibiotic storage for future self-medication. Our findings suggest that strengthening frontline primary care quality can meaningfully reduce inappropriate community antibiotic use without restricting access to essential treatment. More broadly, this study illustrates how causal machine learning can strengthen conventional causal estimation in complex observational settings in global health economics research.

Establishing causality in nutrition research is challenging. While randomised controlled trials (RCTs) provide robust evidence, long-term dietary intervention studies with disease endpoints are often impractical. Short-term RCTs can instead identify intermediate traits that may lie on the causal pathway between diet and disease. Mendelian randomisation (MR) is an epidemiological approach that uses genetic variants as proxies for modifiable exposures to estimate the effects of lifelong differences in exposure on disease risk. However, the utility of MR is limited for complex dietary patterns because genetic variants typically reflect biological mechanisms rather than specific diets. We propose a two-step framework integrating dietary RCTs with MR to infer potential lifetime effects of dietary interventions. First, RCT data identify molecular traits altered by an intervention. Second, MR evaluates whether these traits are associated with long-term disease risk. We demonstrate this framework using the Diabetes Remission Clinical Trial (DiRECT), which measured circulating proteins and diabetes remission. Using protein data alone, 216 of 4,601 proteins changed following the intervention (step 1), and 10 were associated with diabetes risk using MR (step 2). We then compared these MR estimates with observed protein-remission associations from DiRECT. The broad agreement between the two (r{approx}-0.645, R2=0.416) supports this framework as a useful approach for estimating long-term effects of dietary interventions.

The dynamics of gene regulatory networks are governed by intrinsic noise, stemming from the random nature of biochemical reactions, and by extrinsic noise, arising from fluctuations in cellular components and environmental conditions. Together, these sources can compromise the reliability of predictive computational models if not properly accounted for, and capturing both effects within a single framework remains a non-trivial task in computational biology. In this work, we propose an uncertainty quantification framework that addresses these two contributions jointly: intrinsic stochasticity is described through a partial integro-differential equation (PIDE) for the protein probability density function, whereas extrinsic noise is represented as parametric uncertainty in the kinetic parameters. The propagation of the uncertainty is carried out via an intrusive polynomial chaos expansion (PCE), in which the PCE coefficients are obtained from a stochastic Galerkin projection of the PIDE, yielding a coupled deterministic system that is solved with standard numerical methods. We illustrate the approach on a positive autoregulatory gene network with one and two uncertain kinetic parameters. The proposed approach accurately reproduces the mean, variance, and full protein probability density function, including the bimodal distributions, at a substantially lower computational cost.

Objectives To evaluate whether access to a certified large language model (LLM)-based clinical decision support system improves physician diagnostic performance in rheumatology compared with conventional diagnostic resources alone. Methods In this multicentre, open-label, randomised controlled trial, 82 physicians from seven hospitals in two countries were randomised 1:1 to conventional diagnostic resources plus Prof. Valmed or conventional resources alone. Participants assessed three rheumatology vignettes before and after assistance. The primary outcome was top-1 diagnostic accuracy. Secondary outcomes included top-3 accuracy, diagnostic reasoning, confidence, case-processing time and perceived support quality. Results Top-1 accuracy increased from 22.2% to 33.3% in the intervention group and from 23.3% to 35.0% in the control group, with no between-group difference in improvement (adjusted OR 0.99, 95% CI 0.45 to 2.19; p=0.979). Differences in top-3 accuracy, diagnostic reasoning and confidence were also not significant. Assisted case-processing time was substantially shorter with LLM support (94 vs 206 s; adjusted mean difference -112 s, 95% CI -141 to -83; p<0.001). Information timeliness and perceived diagnostic support quality were rated significantly higher in the intervention group. Exploratory analyses showed persistent overconfidence and substantial AI over-reliance. Conclusions Certified LLM-based diagnostic support did not improve diagnostic accuracy compared with conventional resources, but substantially reduced case-processing time and improved perceived support quality. These findings suggest potential workflow benefits while highlighting overconfidence and over-reliance as important safety considerations.