Nature Biomedical Engineering

Automated interpretation of neuroimaging studies requires simultaneous assessment of multiple imaging evidence variables, each tied to distinct anatomical structures. Vision-language models (VLMs) offer a unified framework for multi-task analysis, but adapting pre-trained VLMs remains challenging. Full fine-tuning is computationally prohibitive, and joint multi-task training requires simultaneous access to all task data, which is often infeasible in clinical settings. Although model merging enables multi-task composition without joint re-training, existing methods focus on post-hoc algorithms with limited extension to VLMs and minimal application to neuroimaging. Here, we present GRadient-guided Adapter Merging (GRAM), a layer-selective low-rank adaptation (LoRA)-based fine-tuning and merging framework for multi-task neuroimaging visual question-answering (VQA). GRAM uses a gradient ratio that contrasts class-specific gradients to identify task-discriminative layers, and applies subspace-constrained projected gradient descent to restrict LoRA updates to directions consistent with the geometry of the pre-trained model. We leveraged a structured VQA benchmark, developed from the National Alzheimer's Coordinating Center (NACC) dataset, that pairs multi-sequence brain MRI studies with question-answer pairs across clinically relevant imaging evidence variables. Experiments on the VQA benchmark showed that GRAM outperformed or matched all-layer LoRA fine-tuning and a standard merging baseline while reducing inter-task interference during merging, and approached or surpassed the performance of joint multi-task training without joint re-training.

Biobank-scale genomic analyses remain computationally expensive, CPU-bound workflows, particularly when adjusting for confounding. Here, we present CuGen, a GPU-accelerated framework for large-scale genomics. CuGen uses UltraLasso, a novel hierarchical application of univariate-guided sparse regression (uniLasso), to select a compact, phenotype-informed active set of fewer than 30,000 variants. This achieves robust leave-one-chromosome-out (LOCO) confounding control, enabling both downstream GWAS and in-sample fine-mapping. Additionally, we introduce the .cugen file format, a genotype representation designed for memory-optimized, high-throughput streaming and random access on GPU hardware. Building on this substrate, we provide a general GPU-accelerated genomics toolkit handling polygenic prediction, data manipulation, quality control, analysis, and visualization. We demonstrate CuGen's efficacy in the UK Biobank with up to 408,624 individuals, where the full GWAS pipeline and fine-mapping against 6.8 million imputed variants completes in approximately 10 minutes on a single high-throughput GPU with 80 GB of memory. The pipeline scales efficiently to massive phenome-wide analyses with sublinear resource consumption.

Minimally invasive endovascular procedures offer reduced surgical trauma, shorter recovery times, and improved outcomes, but rely on 2D fluoroscopic X-ray imaging, which provides limited depth perception and exposes patients and clinicians to ionizing radiation. Here we present an augmented reality (AR) system that fuses intravascular ultrasound (IVUS) and electromagnetic (EM) position tracking with preoperative computed tomography (CT) to produce an anatomically accurate, deformation-corrected navigational reference. A robotic device performs ECG-gated pullback of the IVUS probe, capturing 4D aortic motion across the cardiac cycle. We introduce a deep learning architecture for extracting vascular lumen boundaries and side-branch orifices from artifact-prone IVUS streams, and a semantically driven non-rigid CT-IVUS fusion pipeline robust to false positive landmarks. We evaluate the platform with trained surgeons in benchtop phantom studies and in-vivo ovine models, and demonstrate its application to fenestrated endovascular aneurysm repair (FEVAR). Compared to fluoroscopy alone, AR guidance significantly reduces cannulation time, radiation exposure, and cognitive workload, while improving procedural efficiency and safety. Our IVUS-EM and CT aortic datasets are released open source.

We present FootNet, a 453-image multi-view smartphone foot dataset for binary foot segmentation, with expertannotated masks across six anatomical views (dorsal, medial, and plantar, both left and right). We benchmark four segmentation models under a controlled protocol: U-Net with a MobileNetV2 encoder achieves the best performance (IoU 0.9268, Dice 0.9608, 95 % CI [0.9209, 0.9320]); DeepLabV3 with MobileNetV3-Large scores IoU 0.8984 (Dice 0.9449); UNet++ with MobileNetV2 scores IoU 0.8913 (Dice 0.9391); and SAM ViT-B with oracle boundingbox prompt scores IoU 0.9219 on the matched 191-image subset. Bonferroni-corrected Wilcoxon signed-rank tests (k = 6 comparisons) show U-Net significantly outperforms DeepLab (p < 0.001, r = 0.638) and SAM ViT-B with oracle boundingbox (p = 0.005, r = 0.202); UNet++ does not significantly differ from DeepLab (p = 0.062). Connected-component postprocessing yields negligible benefit (mean {triangleup}IoU = +0.0003, 12 of 453 images improved). The extended dataset is available upon request

Multimorbidity is highly prevalent in ageing populations, yet its shared molecular basis remains poorly defined, limiting the development of therapies that target multiple conditions. We systematically integrated measurements of 1,954 circulating proteins from 54,219 individuals in discovery and 35,559 in replication, focusing on ten common age-related diseases: coronary artery disease, chronic kidney disease, chronic obstructive pulmonary disease, dementia, heart failure, major depressive disorder, osteoarthritis, Parkinson's disease, stroke, and type 2 diabetes. Coronary artery disease emerged as a central condition in the multimorbidity network, sharing circulating protein signatures with seven other diseases. Through genetic causal-inference analyses, we identified 40 circulating proteins with cross-disease relevance, of which four were further supported by colocalization of genetic variant associations. Among these, complement C1r, encoded by C1R, emerged as a key link between coronary artery disease and dementia, supported by independent colocalization evidence (PP.H4 = 0.86). Phenome-wide association analyses of C1R variants suggested that this signal was not driven by widespread unrelated genetic effects, but instead may reflect a more specific contribution to coronary artery disease-dementia pathogenesis. In vitro experiments further suggested that fibroblast-derived C1R promotes endothelial inflammation and neuronal apoptosis, providing mechanistic plausibility. Together, these findings position C1R as a biologically plausible and therapeutically relevant molecular link between coronary artery disease and dementia.

We conducted a scaling evaluation of unlabeled pretraining for electrocardiogram foundation model performance. One-dimensional vision transformer masked autoencoders were pretrained across increasing ECG volumes and fine-tuned for rhythm, morphology, diagnostic, and structural heart disease tasks. Models pretrained below 400,000 ECGs failed to consistently exceed controls without self-supervised pre-training, whereas 600,000 to 800,000 ECGs improved AUROC across tasks, suggesting a minimum threshold for effective ECG representation learning.

As machine learning and artificial intelligence systems are increasingly used in healthcare, rigorous evaluation of their classification performance has become critical. The F1 and F{beta} scores are widely adopted metrics for assessing performance in imbalanced biomedical data. Recently, we introduced psF1, a unified statistical framework for inference and study design for single and comparative F1 and F{beta} scores under the assumption of independent classifiers. In practice, however, benchmarking two classifiers on the same dataset creates a correlated paired setting. Ignoring this intrinsic dependency leads to overestimation of the standard error and a substantial loss of statistical power. To address this, we develop psF1pair, an advanced framework for statistical inference and power analysis that explicitly accounts for correlations between classifier pairs. Extensive simulation studies demonstrate the performance of psF1pair, and its utility is further illustrated through application to a real-world imaging classification system. As expected, higher correlation between classifiers yields narrower confidence intervals and enhanced statistical power. A freely available R package is provided to facilitate implementation, supporting accurate evaluation and study design for predictive and classification models in biomedical research.

Offline reinforcement learning (RL) has emerged as a promising framework for clinical decision support in sepsis, yet most existing studies focus exclusively on adult populations, leaving pediatric care largely unexplored despite important physiological and treatment differences. In this work, we develop offline RL policies for pediatric sepsis management in the Pediatric Intensive Care Unit (PICU) using a retrospective cohort of 2,229 episodes from Great Ormond Street Hospital (GOSH), formalized as finite horizon Markov Decision Process (MDP) with joint intravenous fluid and vasopressor actions. To better capture pediatric organ dysfunction dynamics, we incorporate Phoenix 8, a recently proposed pediatric sepsis severity score, as an intermediate reward shaping signal in addition to terminal 90 day mortality. We systematically vary the time step size (4, 8, and 12 hours) and reward structure (terminal 90 day mortality, with and without Phoenix 8 based intermediate shaping), and compare Double Deep Q Networks (DDQN), Conservative Q Learning (CQL), and a behavior cloning (BC) model of clinician practice. CQL consistently exhibits stable learning dynamics and favorable Fitted Q Evaluation estimates, while DDQN is prone to overestimation and instability, particularly at finer temporal resolutions and with dense rewards. CQL policies achieve high action-level agreement with historical clinician decisions for both fluids and vasopressors and reproduce clinically plausible escalation patterns across sepsis severity strata, whereas DDQN policies diverge more frequently toward implausible dosing. Temporal aggregation emerges as a key regularizer: moving from 4 hour to 8 hour bins shortens horizons, smooths reward noise, and improves stability without erasing clinically meaningful dynamics, with 8 hour binning providing the best trade off between policy performance and granularity. Our findings highlight time step size as a core design choice in offline RL for healthcare and provide empirical evidence that alternatives beyond the conventional 4 hour setup can enhance stability and safety while preserving clinical interpretability.

The Cancer Genome Atlas (TCGA) holds clinical data for over 11,000 patients across 33 cancer types, but access is hard because of complex file structures, heterogeneous formats, and the need for programming. We present an agentic system for natural language querying and statistical analysis of TCGA clinical data. The system uses a large language model as an autonomous ReAct agent that selects from eight computational tools, including data extraction, descriptive statistics, Kaplan-Meier survival analysis with log-rank tests, hypothesis testing, and verification against the curated TCGA Pan-Cancer Clinical Data Resource (CDR). The agent reasons about intermediate results, adapts its approach, and returns clinically contextualized responses with source attribution and auditable traces. We introduce TCGA-Agent-Bench, 440 queries across five difficulty tiers with ground truth from the independently curated TCGA-CDR, evaluated with dual metrics of numerical accuracy and clinical completeness. The system achieves 93.4% overall accuracy (100% single-patient lookups, 99.1% cohort statistics, 92.8% comparative analyses), outperforming a fixed rule-based pipeline (87.1%), a single-pass LLM (81.8%), and retrieval-augmented generation (66.9% on a subset). Most of the benchmark is answerable from the CDR alone, so we locate the extraction layer's value in fields the CDR lacks (drug treatments, TNM components, biomarkers, biospecimen metadata): on 26 queries targeting these, the full system answers 100% versus 3.8% for CDR-only. Ablations show the reasoning loop is most impactful (+9.1% accuracy, +22.0 completeness points). A tool-based agentic architecture enables accurate, auditable analysis of clinical repositories, with value driven by tool design and recovered fields rather than model scale.

Background The 2022-2023 global mpox outbreak highlighted the critical need for robust genomic surveillance capabilities to track mpox virus (MPXV) evolution and transmission dynamics. Methods Building upon our established SARS-CoV-2 sequencing infrastructure, we implemented a Molecular Loop probe-based long-read sequencing approach using Pacific Biosciences Sequel II technology for comprehensive MPXV genomic surveillance across the United States (US). From August 2024 to June 2025, we generated 326 high-quality whole genome sequences from residual mpox-positive clinical specimens collected by Labcorp across all 10 US Department of Health and Human Services regions. Results Our analysis identified two samples containing clade Ib MPXV in January and June 2025 and captured shifting trends in clade IIb diversity, with 13 distinct lineages observed. We also identified multiple instances of large (~1.6-17.6kb) deletions proximal to the inverted terminal repeats in clade IIb genomes. APOBEC3 mutation analysis indicated substantial evidence of human-to-human transmission among both clades. Further, we observed significantly higher APOBEC3-associated SNPs per kilobase (P<0.001) in clade IIb genomic variable regions relative to their central conserved region. Our assay exhibited strong reproducibility across biological replicates from individual patients and accuracy was confirmed via parallel sequencing of select specimens by US Centers for Disease Control and Prevention (CDC) using metagenomic sequencing. We also demonstrated via custom simulation that our assay discriminates all known MPXV clades and lineages, including those we have not observed in the US. Conclusions Our integrated nationwide surveillance system facilitates real-time genomic tracking of outbreak evolution, with demonstrated capacity across SARS-CoV-2 and MPXV, positioning this platform for rapid deployment during future pathogen emergence.

Asthma is a heritable complex disease that disproportionately burdens minority and admixed populations in the US. However, the causal genes and regulatory mechanisms governing inherited risk remain largely unresolved. We performed a European-ancestry meta-analysis of 141,894 cases and 1,361,846 controls drawn from the Trans-national Asthma Genetic Consortium (TAGC) and Global Biobank Meta-analysis Initiative (GBMI), yielding an estimated h2SNP of 0.056 (SE = 0.0038) and 275 independently associated loci. To enhance mechanistic inference beyond variant-level associations, we developed a multimodal framework to predict asthma risk integrating GWAS summary statistics, bulk tissue expression quantitative trait loci (eQTL) data from the Genotype-Tissue Expression (GTEx) project, and single-cell gene eQTL data from the OneK1K Project. We performed transcriptome-wide association studies (TWAS) and subsequently applied probabilistic fine-mapping with FOCUS to prioritize putative causal genes expressed in bulk tissues and higher resolution immune cell populations. Fine-mapping asthma-associated genes implicated barrier-immune and metabolic-endocrine tissues alongside adaptive T-cell subsets as the primary mediators of asthma genetic risk, resolving canonical CD4+ Th2 effector genes including IL1RL1, TSLP, STAT6, and GATA3. Using these prioritized genes, we constructed a polygenic transcriptome risk score (PTRS) using random forest to integrate gene-level effects across critical tissues and cell types. Evaluated in two ancestrally distinct pediatric asthma cohorts, the Childhood Asthma Management Program (CAMP) and the Genetics of Asthma in Costa Rica Study (GACRS), our PTRS demonstrated improved transferability over the standard variant-level and gene-level baseline models. While modest common variant heritability limits the discriminative power of our models, we estimated a theoretical maximum achievable area under the receiver operating characteristic (AUROC) curve of 0.64. Our integrative nonlinear model of PRS-CSx and cross-modal (bulk tissue and single cell) FOCUS PTRS resulted in the best cross-cohort performance (CAMP AUC = 0.632, sd = 0.04, 3.55 case/control odds ratio in top vs. bottom quartiles), representing an increase of +0.118 AUC over PRS-CSx, +0.067 AUC over tissue-specific TWAS pruning and thresholding, and +0.041 AUC over cell-type-specific FOCUS PTRS. Our results demonstrate that modeling nonlinear interactions between variant- and gene-level effects across both bulk tissue and single cell eQTL data improves our ability to determine high-risk individuals and to explain the likely mechanisms driving genetic susceptibility of childhood-onset asthma.

Rising costs, slow accrual and molecular substratification of cancers necessitate novel clinical trial designs. We demonstrate that artificial intelligence-generated synthetic patients can replace real controls to reproduce results of the SORAML trial. Using external multimodal data from 1,377 acute myeloid leukemia (AML) patients from previous trials and a real-world registry, we fine-tuned a tabular foundation model to generate synthetic patients, reproducing clinical and genetic features and outcome associations. Synthetic patients were then matched to the original SORAML intervention group using Cox risk scores, replacing the original control and reproducing the original trial result with near-identical median event-free survival (EFS) and treatment effect (original hazard ratio [HR] 0.64, 95%-confidence interval [CI] 0.47-0.87, p=0.004; with synthetic control HR 0.66, 95%-CI 0.48-0.90, p=0.009). Our findings demonstrate that AI-generated synthetic patients can serve as statistically rigorous controls supporting novel trial designs.

The aim of this work is to validate patient-specific EEG baseline establishment using the e-norms method as a screening and retrospective-review tool for seizure detection in pediatric epilepsy. The method was applied to 247 seizure-free EEG recordings (263.92 hours) from 10 patients in the CHB-MIT Scalp EEG Database (ages 3-18). A composite stability metric combining first-derivative dynamics, spectral entropy, variance, and line length was computed per 2-second epoch across 23 channels. Patient-specific detection thresholds were derived from each patient's seizure-free baseline using a weighted statistical procedure. Performance was validated against 72 expert-annotated seizures (2,705 epochs) across 62 seizure files, with durations spanning 6 to 264 seconds (44-fold range). The results show that detection achieved 94.4% event-level sensitivity (68 of 72 seizures; 95% CI 86.6-97.8%) and 81.5% epoch-level sensitivity (2,204 of 2,705 epochs; 95% CI 80.0-82.9%). Eight of ten patients achieved 100% event-level sensitivity with epoch-level sensitivity ranging from 58.7% to 100.0%. Two patients showed partial event-level failures (CHB-15: 17 of 20; CHB-18: 5 of 6), with the four missed events attributable to two characterizable failure modes. Patient-specific thresholds ranged from 4.06 to 4.81 (mean 4.51 +/- 0.25); threshold variation did not correlate reliably with age or sex, confirming that no universal threshold could achieve comparable performance. Detection margins ranged from 0.88 to 1.24 times. Patient-specific e-norms achieves 94.4% event-level sensitivity for pediatric EEG seizure detection without requiring labeled seizure training data, exceeding published human expert inter-rater agreement (50-76%) and recent automated approaches in adult cohorts using behind-the-ear EEG and wearable ECG. Two characterizable failure modes account for the four missed events and inform appropriate clinical use. As a high-sensitivity screening tool complementary to real-time alarm systems, the method is ready for adult validation, prospective deployment, and head-to-head benchmarking.

Pneumonia is a leading cause of infectious disease mortality worldwide, accounting for approximately 2.5 million deaths annually and 15% of deaths in children under five. Chest X-ray imaging remains the primary diagnostic tool, but accurate interpretation requires radiological expertise that is disproportionately concentrated in high-income settings, creating a diagnostic gap where disease burden is highest. Automated deep learning offers a scalable complement to specialist-dependent diagnosis, yet clinical adoption requires both high accuracy and transparent, interpretable reasoning. Convolutional neural networks (CNNs) have shown strong potential for pneumonia detection from chest X-rays, but two barriers impede clinical translation: the interpretability of black-box models and the computational feasibility of large architectures in resource-constrained settings. Explainable AI (XAI) methods such as Grad-CAM, Grad-CAM++, and Score-CAM address the interpretability barrier, yet systematic quantitative comparisons across multiple CNN architectures remain scarce. Furthermore, CNN architectures widely used for medical image classification carry high parameter counts that limit feasibility in resource-constrained settings, motivating architectures that achieve competitive accuracy with substantially fewer parameters. Here we propose a parameter-efficient deep learning framework for pneumonia detection based on transfer learning, evaluated across three CNN architectures representing distinct architectural families: EfficientNet-B0 with fine-tuning (proposed method), ResNet50, and DenseNet121, trained under identical conditions on the Kaggle chest X-ray dataset (5,863 images). Our method achieved 90% classification accuracy, outperforming both baselines while requiring 4.8x fewer parameters than ResNet50. To evaluate explainability, Grad-CAM, Grad-CAM++, and Score-CAM were applied across all three architectures and compared quantitatively using Intersection over Union against manually annotated lung segmentation masks, Insertion score, and Deletion score, with pairwise statistical validation via Wilcoxon signed-rank tests and Bonferroni correction. Findings show that classification accuracy and XAI explanation quality must be evaluated independently, and that the proposed parameter-efficient architecture offers a favorable trade-off for resource-constrained clinical deployment.

Objectives. Radiomics pipelines extract hundreds of quantitative features that are widely known to be redundant, but the structure of this redundancy is usually treated as a per-dataset nuisance to be pruned away. We tested the alternative hypothesis that a substantial number of feature-feature correlations are universal: they persist across patients and across anatomically distinct structures because they reflect shared mathematical and image-statistical properties of how the image is summarised, rather than properties of the tissue being imaged. Materials and Methods. We re-analysed the publicly available Radiomics Atlas Dataset of normal Abdominal and Pelvic CT (RADAPT), restricting the analysis to the 526 non-contrast-enhanced examinations of the 531-subject atlas and to the 107 original (non-filtered) PyRadiomics features. The 53 segmented structures were grouped into four broad anatomical categories -- bones, muscles, vessels, and parenchymal organs. RADAPT is distributed as one Excel file per structure, with patients as rows and features as columns. Within each structure file we z-score-normalised every feature across patients, computed the absolute Spearman correlation matrix, and retained edges with |{rho}| [&ge;] {tau} for {tau} in {0.70, 0.80, 0.90}. We then intersected the edge sets across all structure files to obtain a "universal" correlation graph, in which an edge survives only if it exceeds the threshold in every structure (each estimated across the full patient sample). Stable feature communities were defined as the maximal cliques of this graph. Robustness to patient sampling was tested by repeating the entire pipeline on five independent random splits of each file into two patient halves (10 sub-cohorts per threshold), and the implementation was independently reproduced in R. Results. Despite the strictness of the global-intersection criterion, 34, 24, and 14 stable feature communities survived at {tau} = 0.70, 0.80, and 0.90 respectively, with the largest cliques containing six members at {tau} = 0.70 and {tau} = 0.80 and five members at {tau} = 0.90. The community structure was clearly interpretable: separate cliques captured (i) variance-like intensity dispersion, (ii) long-run / low-frequency (coarse) texture, (iii) high gray-level texture, (iv) low gray-level texture, (v) volume and surface shape, and (vi) local-homogeneity and energy/entropy duals. On random-half resampling the exact-match recovery rate of these communities was 81.5 %, 86.7 %, and 80.7 % across the three thresholds; departures from exact recovery were almost always a single boundary feature added or dropped, consistent with finite-sample fluctuation of near-threshold edges rather than structural instability. The R re-implementation reproduced the Python results exactly. Conclusion. A substantial portion of radiomics feature collinearity is universal across patients and tissues. We distinguish two layers within it: trivial near-algebraic duals that are universal by construction, and non-trivial cross-matrix-family communities that are the genuine empirical finding. Together they provide an interpretable, definition-grounded basis for aggressive dimensionality reduction, for retrospectively reconciling apparently different feature selections in the literature, and for moving radiomics pipelines toward organ-agnostic, more reproducible models. Clinical relevance statement. Selecting a single representative feature from each universal community shrinks the original-feature space by roughly an order of magnitude without sacrificing biologically distinct information. For example, the five variance-family members (first-order Variance, GLCM SumSquares, GLCM ClusterTendency, GLDM and GLRLM GrayLevelVariance) can be replaced by a single representative, removing redundant degrees of freedom that would otherwise inflate model variance; and labelling each retained feature by its community lets two studies that selected different variance-family names be recognised as having found the same signal, simplifying model development and improving cross-cohort generalisability in clinical CT workflows.

Introduction Parkinson's disease (PD) is a multifactorial disorder, affecting multiple neurotransmitter systems, including the cholinergic system. Cholinergic denervation is heterogeneous across patients and difficult to predict based on clinical presentation. In this study, we assessed the sensitivity of structural MRI (sMRI) and functional MRI (fMRI) to cholinergic degeneration related to PD and to cognitive functioning in PD. We compared our results to results from previously reported [18F]Fluoroethoxybenzovesamicol ([18F]FEOBV) PET imaging, which is considered the gold standard for cholinergic imaging. Methods 34 PD patients and 10 healthy controls underwent structural T1-weighted MRI. A subset of 14 patients and 9 controls also underwent resting-state fMRI. We extracted the bilateral volumes of the nucleus basalis of Meynert (NBM) from the sMRI images. Functional connectivity (FC) from the NBM to the cortex (NBM-FC) was determined using fMRI data. Principal component analysis (PCA) was applied to reduce the dimensionality of the NBM-FC images. We assessed performances for NBM-FC in distinguishing patients from controls using stepwise logistic regression. Similarly, NBM volume was used using logistic regression. Furthermore, the relation between these measures and cognitive function in several domains was investigated with (stepwise) linear regression. Leave-one-out cross validation (LOOCV) and bootstrapping was performed to assess robustness of the results. Results NBM-FC was well able to discriminate patients from controls with an AUC of 0.84 (95% CI: 0.62-1). NBM volume showed lower performance, but was still better than chance: AUC: 0.75 (95% CI: 0.57-0.93). Significant correlations were found between 1) cognition in the attentional domain and NBM-FC (r=0.63; p=.015) and 2) global cognition and NBM volume (r=0.55, p=.001). These results were inferior to those previously reported using [18F]FEOBV tracer uptake (see Chapter 6). Bootstrapping revealed that NBM volume of only the left hemisphere was stably related to PD diagnosis and global cognition in PD patients. We found that a lower NBM-FC in specific brain areas, including the fusiform gyrus, supramarginal gyrus and dorsolateral prefrontal cortex, was related to PD diagnosis. Bootstrapping revealed no stable NBM-FC pattern related to attention. Conclusion Although MRI results were slightly inferior to [18F]FEOBV PET data, MRI may provide a cheaper and more widely available alternative for cholinergic imaging. We recommend testing the utility of MRI as predictor and monitor of cholinergic treatment effect in a longitudinal study.

Background: Actigraphy-derived rest-activity rhythm (RAR) features are widely used to characterize clinical states in bipolar disorder (BD). Both mean levels and temporal variability of these features have been associated with mood episodes; however, variability measures are often statistically coupled with the mean, particularly in skewed distributions. This raises a question as to whether variability reflects a separate characteristic of the data or whether the observed association arises from statistical properties of the data. Objective: In this study, we aim to determine whether temporal variability of actigraphy-derived RAR features provides standalone information on mood episodes in BD beyond mean activity levels after accounting for mean-variance dependence. Methods: We analyzed actigraphy data from a subset of 72 participants with BD drawn from a larger longitudinal study, extracting 22 daily RAR features aggregated weekly as sample mean (MEAN) and within-week temporal variability computed as sample standard deviation (VAR). Variance-stabilizing transformations (Box-Cox or Yeo-Johnson) were applied to the entire study cohort to reduce mean-variance dependence. Associations with mood episodes and remission (mania: n=34; depression: n=58 annotated participants) were evaluated using generalized linear mixed-effects models with a logistic link function, including univariate (MEAN or VAR) and multivariate (MEAN+VAR) specifications, assessed by likelihood-based metrics and the area under the receiver operating characteristic curve (AUC). Results: Transformations reduced mean-absolute correlations from 0.43 to below 0.06. Temporal variability remained significantly associated with clinical state for 11/22 RAR features in mania and 16/22 features in depression, with all significant associations remaining after false discovery rate correction (p<0.05). Joint models showed modest incremental gains (AUC 3%-4% overall; up to 12% in mania, 7% in depression), with absolute performance remaining limited (AUC 0.50-0.66). In both mania and depression, nearly all significant variability-based regressors contributed incremental information beyond mean-based models. Only sleep duration and activity changes around wake time (+-1 hour), did not improve discrimination between mania and remission. Conclusions: Temporal variability in RAR features can be considered a standalone state marker of mood episodes not captured by mean activity. We found it to be more consistently associated with depression than mania. Its incremental discriminative contribution is modest, suggesting greater utility within multivariate or multimodal frameworks.

Background: Large language models (LLMs) struggle with dynamic, longitudinal clinical reasoning. We developed a Multi-Stage Iterative Clinical Reasoning Agent framework to address this gap and systematically decouple the clinical efficacy of static retrieval-augmented generation (RAG) from dynamic self-refinement. Methods: Ten complex longitudinal nephrology cases, rigorously selected via a modified Delphi consensus technique, were blindly evaluated by four board-certified nephrologists and a multi-model AI panel. We compared three architectures across nine cognitive steps: (Model A) a baseline frontier LLM, (Model B) an LLM augmented with static guideline-based RAG, and (Model C) our proposed multi-agent framework featuring RAG integrated with iterative self-critique and refinement. Results: In human evaluations (20-point scale), Model C (mean 17.2, SD 1.2) significantly outperformed both Model A (16.1, 1.3) and Model B (16.2, 1.2) (P < 0.001). Implementing static RAG (Model B) yielded no significant improvement over the baseline. Automated AI evaluations (15-point scale) corroborated these findings: Model C (14.7, 0.6) outscored Model A (14.2, 0.9, P < 0.001) and Model B (14.3, 0.9, P = 0.01). While monolithic models exhibited severe score degradations in planning-heavy tasks such as dynamic differential diagnoses, the multi-agent framework effectively intercepted error cascades, achieving significantly higher diagnostic accuracy (mean 17.6, P = 0.019) and therapeutic management scores (17.3, P = 0.002). Conclusions: Static knowledge retrieval alone fails to enhance frontier LLM performance in longitudinal medical reasoning. Distributing clinical workflows into a multi-agent dynamic refinement pipeline significantly improves reasoning completeness, intercepts error cascades, and safely resolves planning bottlenecks in complex patient care.

Lithium has been the gold standard for the treatment and prevention of relapse in bipolar disorder for over 60 years. Guidance from the National Institute for Health and Clinical Excellence states explicitly to 'offer lithium as a first-line, long-term pharmacological treatment for bipolar disorder'. Yet, in the last two decades its use has been in decline with clinicians favouring anticonvulsants or antipsychotics when treating this condition. In this study, we have used three openly available datasets containing prescribing data from primary and secondary care to explore trends in the use of lithium in England, showing both regional and temporal variance between 2015-2024. We have shown that lithium use declined in primary care by 20.9% in the last ten years (2015-2024) and 10.9% overall in the last five years (2019 to 2025). We have also shown how there is some regional variation in the source of lithium for patients, although the vast majority is prescribed in primary care. Further research into clinical behaviour is needed to understand what is driving the decrease in lithium usage, and what barriers and enablers may influence its use across the country.

Objective: There is a bidirectional relationship between emotional distress and pain. However, this relationship is understudied in people with HIV in low-resource settings. This study sought to describe the temporal relationship between emotional distress and pain in people with HIV. Design: Longitudinal observational study. Methods: Participants with virally suppressed HIV, reporting either no pain or persistent pain at baseline, provided weekly remote ratings of distress, worst pain, and average pain using 0-10 visual analogue scales. Within-individual fluctuations in distress and pain were visualised over time. Group-level correlations were determined using Spearman's correlation tests. Cumulative link mixed models assessed whether distress and pain each predicted the other in the following week. Results: 72 participants provided responses over 49 weeks. The participants had a median (IQR) age of 43 (37-51) years, 63% (n=45) were unemployed and most were females (n=51;71%). Distress and pain fluctuated concurrently within individuals: distress was positively correlated with worst pain ({rho}=0.66, 95% CI= 0.60-0.72, p<0.001) and average pain ({rho}=0.70, 95% CI=0.64-0.75, p<0.001) intensity within the same week. Worst pain (OR=1.42, 95% CI=1.17-1.71, p<0.001) and average pain (OR=1.43, 95% CI=1.20-1.71, p<0.001) intensity both predicted distress in the next week. Distress predicted worst pain intensity (OR=1.25, 95% CI=1.07-1.46, p=0.023) but not average pain intensity (OR=1.19, 95% CI=1.01-1.40, p=0.152) in the next week. Conclusions: The temporal relationship between distress and worst pain intensity was bidirectional, whereas distress did not temporally predict average pain intensity. Both pain and emotional distress should receive attention from HIV research and clinical care in low-resource settings.