Nature Machine Intelligence

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.

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.

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.

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.

External validation of clinical AI emphasizes discrimination, although deployment requires the endpoint, probability estimates and operating policy to transport. Here we show that these layers diverged in retrospective bidirectional evaluation of five model families across eICU and MIMIC-IV. Coarse-label AUROC fell from 0.87-0.92 internally to 0.66-0.83 during source-only transfer. For assessment-conditioned repeated monitoring of persistence or recurrence, external AUROC reached 0.76-0.94, but removing assessment history reduced it by 0.16-0.32; broader features did not help consistently. Transported scores concentrated future-positive ICU stays 2.4-6.9-fold in the top risk decile. Development-selected cutoffs alerted 0.3-2.0% of prediction rows and captured 9.2-11.0% of future-positive rows; after deduplication, 4.9-12.2% of stays were alerted, capturing 43.9-49.4% of future-positive stays. Thus, ranking can persist while probability and policy transport remain site dependent. Layered validation is a prerequisite for prospective evaluation, not evidence of clinical benefit.

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.

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.

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}| [≥] {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.

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

Causal effect estimates can often be biased in clinical and epidemiological studies as patient cohorts frequently exhibit substantial covariate imbalances between treated and control groups, often amplified in multicentre studies due to heterogeneous recruitment, clinical practice, and case mix. Covariate balancing methods are therefore essential for valid causal inference. However, their application becomes challenging when data are distributed across cohorts and cannot be pooled because of privacy, legal, or institutional constraints, leaving a gap in practical methods for causal effect estimation in federated and imbalanced clinical data settings. We develop a privacy-preserving framework for covariate balancing and causal effect estimation across distributed data providers, combining federated aggregation with differential privacy to enable propensity score subclassification and matching without sharing individual-level records. Matching relies on non-disclosive quantities and differentially private distance evaluation, and the resulting matched subsets remain local to each server. Balance can be assessed through federated diagnostics and privacy-preserving visualisations, and we provide secure estimators for average treatment effects with associated uncertainty quantification. We implement this framework in the DataSHIELD federated analysis platform via 2 R packages. In simulations, we demonstrate agreement between federated and centralised analyses in the absence of privacy noise and quantify the bias--variance trade-offs induced by differential privacy. We illustrate applicability in two multinational settings-a Long COVID cohort and very preterm birth cohorts-showing that the approach enables practical causal analyses under real-world data protection constraints. The DataSHIELD packages are available on Github. Additional methodological details are provided in the Supplementary Material.

INTRODUCTION: The biomarker-based amyloid/ tau/ neurodegeneration (A/T/N) framework has become a popular staging method for Alzheimer's disease (AD) research. Previous studies use the framework either as a rule-based or data-driven approach but typically sacrifice either adaptivity or interpretability. METHODS: We present an interpretable, hybrid method, called Neurosymodal Data Fusion, for predicting incident AD in the ADNI dataset. Specifically, we encode the A/T/N framework as a logic program, where the input biomarker features are extracted by one or more neural networks. RESULTS: Our pipeline predicted four-year incident AD with a sensitivity of up to 0.84. Additionally, our models learned scores for each A/T/N profile, denoting relative importances to model predictions. These scores also indicated that empirically-derived cut-off values for the A and T criteria might be uninformative for the ADNI data. DISCUSSION: Our pipeline provides a novel way to use the A/T/N framework that could potentially improve early AD screening years before clinical manifestations.

The incidence rate ratio R is the standard measure for comparing event rates in clinical trials and epidemiology. In vaccine trials, the vaccine efficacy is VE = 1 - R. When events are rare, the two arm counts are Poisson. The estimator of R is heteroskedastic: its sampling variance changes with the data. So no fixed-width interval covers correctly everywhere. The usual log-Wald interval is undefined at zero events and covers poorly at small counts. Early vaccine and drug-safety readouts fall in exactly this regime. We show that a single reparameterization collapses this bivariate problem to an effective one-parameter family with a quadratic variance function, whose variance-stabilizing transformation is 2 arcsinh(sqrt(R)). The reduction yields a closed-form confidence interval for R. Its two leading errors, a curvature bias and the variability of the estimated scale, each admit a closed-form correction with no tuning constants. In a Monte Carlo study of our seven arcsinh variants and five competitors, the +Curve+Stu variant covers within 0.002 of the nominal 0.95 for about 50 control and 5 treatment events. Its width is on par with the best competitor. It avoids the conservatism and zero-count breakdown of log-Wald and MOVER. For moderate counts, we recommend this interval; for sparser data, our Bar-Lev and Enis count-shift variant is more robust. The result is a ready-to-use, closed-form interval for the low-count regime. We illustrate it on early Covid-19 vaccine-efficacy readouts and provide reference implementations in R and Python.

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.

An epidemic's expected course is determined by the magnitude and timing of a typical person's infectiousness --- captured, in turn, by the basic reproduction number and the generation-time distribution. These fundamental, population-average quantities can mask individual-level variation that shapes how an epidemic actually unfolds: for example, individual variation in the magnitude of infectiousness (overdispersion) creates superspreading, a key feature of the SARS-CoV-1 and SARS-CoV-2 epidemics. However, the impact of individual variation in infectiousness timing is less well understood. Here, we demonstrate that individual infectiousness timing varies substantially and to different degrees across pathogens. For some common pathogens, including influenza, measles, and SARS-CoV-2, infectiousness is "bursty", or highly concentrated and variably-timed across individuals: for example, the window of appreciable infectiousness for SARS-CoV-2 may last for roughly a day, vs. the 9--12 days usually quoted. We show that bursty infectiousness creates superspreading without inherent superspreaders, makes epidemic timing more variable, amplifies the time-sensitivity of common interventions, and complicates inference of key epidemiological parameters. Together with the reproduction number, the generation-time distribution, and overdispersion, burstiness completes a family of basic parameters that govern how epidemics unfold.

Purpose: To compare the cross-site generalization of radiomic features and deep learning embeddings for MRI prediction of substantial lymphovascular space invasion (LVSI) in endometrial cancer. Materials and Methods: This retrospective two-center study included 206 women (mean age, 59.8 years) with endometrial cancer who underwent preoperative 3-T MRI from March 2016 to March 2023. Hospital A (n = 130) was used for development and Hospital B (n = 76) for strict external testing. T2-weighted, reduced field-of-view diffusion-weighted, and apparent diffusion coefficient images were manually segmented. Radiomic features and seed-pooled embeddings from 3D ResNet18, DenseNet121, and U-NEXtractor were modeled with elastic-net logistic regression or XGBoost. Out-of-fold Platt calibration and sensitivity-targeted thresholds were estimated using development data only. AUCs were summarized with 95% bootstrap confidence intervals. Results: External radiomics with elastic-net achieved an AUC of 0.609 (95% CI: 0.464, 0.740) and sensitivity of 0 of 12 (0%). DenseNet121 with elastic-net had the highest external AUC (0.685; 95% CI: 0.538, 0.822) but sensitivity of 3 of 12 (25%). U-NEXtractor with elastic-net detected 10 of 12 positive cases (83.3%) with specificity of 32 of 64 (50.0%) and balanced accuracy of 0.667. XGBoost showed higher apparent development performance but weaker external operating behavior. Conclusion: Under real-world cross-site MRI acquisition shift, DenseNet121 and U-NEXtractor embeddings showed better external generalization than handcrafted radiomic features for substantial LVSI prediction.

CRISPR-Cas12a nucleases are classically activated through CRISPR RNA (crRNA) guided and PAM-dependent target recognition, which together establish a canonical heteroduplex associated with nuclease activation. Here we identify a crRNA- and PAM-independent activation pathway for Cas12a that reveals previously unrecognized conformational plasticity within its nucleic acid recognition interface. We show that short RNAs can directly occupy the canonical crRNA-binding channel and trigger a catalytically competent trans cleavage state in the absence of PAM recognition or canonical R-loop formation. Biochemical assays indicate that short RNAs bind the crRNA-binding channel and are competitively displaced by cognate crRNA, consistent with binding at a conserved nucleic acid-binding interface. Cryo-electron microscopy (cryo-EM) further reveals that Cas12a maintains its global catalytic architecture while exhibiting loss of canonical PAM-dependent stabilization and increased flexibility of the RuvC lid, alongside accommodation of a noncanonical RNA-DNA hybrid with inverted polarity relative to the crRNA-target duplex. This crRNA-independent activation pathway enables programmable, amplification-free detection of DNA and RNA targets independent of canonical guide-mediated recognition. Together, these findings define an alternative activation geometry for Cas12a and expand models of Class 2 CRISPR-Cas effector activation beyond crRNA- and PAM-directed recognition.

Stochastic epidemic models are a cornerstone of infectious disease epidemiology and are often used to study intervention scenarios. However, large run-to-run variability can make intervention effects difficult to estimate precisely. We revisit the epidemic Sellke construction, which assigns each individual an infection threshold for the cumulative infection hazard such that, conditional on the thresholds, the epidemic trajectory becomes deterministic. This enables coupling of simulations with and without an intervention, yielding low-variance effect estimates even when outcomes such as final size or peak incidence vary widely between runs. We develop an exact, event-driven implementation that maintains infection and recovery events in priority queues. Cumulative infection-hazard updates require O(log N) time per event, yielding overall complexity O(Elog N) for E events in a population of size N. The implementation achieves computational performance comparable to the classical Gillespie algorithm while naturally accommodating non-Markovian infectious periods and complex infectiousness profiles. We illustrate the approach using distance-dependent spread of avian influenza between poultry farms in the Netherlands and a multilayer population with households, schools, and workplaces. In both examples, coupling enables efficient within-run comparisons of intervention scenarios across stochastic realisations.

We evaluated aggregated, privacy-preserving smartphone-detected nocturnal cough activity from the Sleep Cycle application as a potential syndromic surveillance signal in England. Weekly cough metrics from January 2023 to January 2026 were compared with UK Health Security Agency indicators: NHS 111 acute respiratory infection (ARI) triage calls, influenza and COVID-19 PCR positivity, and hospital admission rates for influenza, COVID-19, and respiratory syncytial virus. We evaluated total cough counts alongside two population-normalised metrics, coughs per user and coughs per hour of sleep, and assessed temporal relationships nationally and regionally using cross-correlation with prewhitening. The strongest and most consistent associations were observed for NHS 111 ARI triage calls, where population-normalised cough metrics showed raw national correlations of approximately 0.95 and retained prewhitened correlations above 0.55 at lag 0. This indicates that nocturnal cough activity closely tracks short-term variation in an established syndromic surveillance indicator, beyond shared seasonality, long-term trends, and autocorrelation. Similar near-contemporaneous patterns were observed across regions. Population-normalised cough metrics also showed epidemiologically plausible leading associations with pathogen-specific indicators: coughs per hour of sleep peaked one week before influenza PCR positivity, while both coughs per user and coughs per hour of sleep peaked one week before COVID-19 PCR positivity. Hospital-based indicators showed weaker and more heterogeneous relationships, but the normalised cough metrics still showed plausible temporal alignment with influenza and COVID-19 admissions, including contemporaneous associations with influenza admissions and short leading associations with COVID-19 admissions. In contrast, unnormalised total cough counts produced less stable and often non-interpretable lag structures, consistent with sensitivity to variation in observation volume. These findings suggest that passive, near-real-time nocturnal cough monitoring can provide a population-level signal of respiratory symptom burden, with greatest utility as a broad syndromic indicator that complements surveillance sources affected by healthcare-seeking behaviour, laboratory turnaround times, backfilling, and reporting delays.

Respiratory diseases cause considerable morbidity in autumn and winter and are a priority in public health monitoring. In Germany, they are subject to a number of surveillance systems, including both pathogen-specific and syndromic indicators. In this paper we present a collaborative multi-target and multi-model real-time forecasting system rolled out during the 2024/25 season, and discuss differences to earlier efforts carried out during the COVID-19 pandemic. A total of nine models were run to generate forecasts of general practitioner consultations for acute respiratory infections (ARI), hospitalizations for severe acute respiratory infections (SARI) and confirmed cases of seasonal influenza and RSV. As all indicators were subject to retrospective revisions, forecasting models were combined with a nowcasting step. Whenever multiple models were available for the same indicator, we combined them into an ensemble. Nowcasts showed convincing performance, even though for some models Christmas break effects led to an upward bias in early January. Forecasts were overall well-calibrated and most models outperformed simple benchmark models. These improvements were generally more substantial for age-stratified than pooled targets, and concentrated at lead times of two to three weeks. Anticipating the peak timing and magnitude proved to be challenging, with many models predicting too flat curves with a too early turnaround (e.g. already in late January rather than mid-February for SARI). The combined ensemble forecast was among the best-performing approaches, but unlike in previous related projects did not consistently outperform individual models. We conclude by discussing learnings on the organization of collaborative forecasting projects in post-COVID-19 times and the potential of AI-supported modelling.

This study presents ReCo (Research Cosmos), a self-configuring and self-extending agentic research framework for the biomedical domain. ReCo is orchestrated by a large language model that interacts with native computing tools, bundled Model Context Protocol (MCP) servers, structured skills, persistent project memory, and a desktop interface. Its bundled MCP servers provide biomedical analysis capabilities while serving as implementation paradigms for integrating new computational and AI frameworks. Structured skills encode procedures for environment configuration and framework ingestion, enabling ReCo to inspect repositories, manuscripts, or local codebases; identify dependencies and execution patterns; create isolated runtime environments; design and implement MCP interfaces. Self-extension was evaluated using five heterogeneous systems: the Merlin computed tomography foundation model, MAISI-v2 medical image synthesis framework, asari liquid chromatography-mass spectrometry workflow, DosimeTron agentic radiation-dosimetry platform, and Orthanc DICOM server. ReCo successfully operationalized all five systems and completed predefined functional evaluations. Re-hosted DosimeTron outputs demonstrated near-perfect agreement with the reference pipeline across 651 organ observations (Pearson correlation and Lin concordance correlation coefficient, 0.99999; mean absolute percentage difference, 0.37%). Notably, ReCo configured Orthanc as a PACS-like coordination layer, integrated it with DosimeTron, Merlin, and TotalSegmentator, and orchestrated data retrieval, analysis, and return of valid DICOM RTSTRUCT, RTDOSE, and Structured Report. ReCo provides a unified environment for configuring, documenting, and operationalizing heterogeneous biomedical frameworks, reducing technical barriers to the adoption and integration of emerging computational and AI methods. The official open-source ReCo GitHub repository is available at: https://github.com/eltzanis/ReCo