Med

Background Clinical data analysis typically requires statistical programming skills, whereas cloud-based artificial intelligence (AI) agents risk exposing sensitive patient records. We developed and functionally validated a privacy-preserving, zero-code conversational statistical analysis framework that translates natural-language clinical research requests into executable R workflows while strictly retaining raw patient data within local computing environments. Methods Orchestrated by the n8n engine, the system integrates the DeepSeek-Reasoner model with a Pinecone vector database for retrieval-augmented generation (RAG), grounding statistical selection in curated biostatistical guidance and R templates. Core functionalities include data schema perception, interactive data cleaning, requirements refinement, and local R code execution via a controlled command-line interface. System performance was evaluated by replicating a published prognostic model study on metabolic dysfunction-associated steatotic liver disease (MASLD). Findings All core analytical workflows, including data cleaning, multivariable Cox proportional hazards modeling, model diagnostics, and publication-ready tables and figures (e.g., baseline characteristics, Schoenfeld residuals, receiver operating characteristic curves, and forest plots), were executed solely through natural-language dialogues without manual coding. The external large language model actively clarified analytical prompts while receiving zero row-level patient data. Interpretation Decoupling remote cloud reasoning from local code execution lowers the technical threshold for clinicians conducting data-driven research while safeguarding data privacy. This architecture provides a practical, scalable, and reproducible framework for converting natural-language clinical questions into executable statistical workflows. Funding National Natural Science Foundation of China (82473291), Shaanxi Province "Three Qin Scholars" Innovation Team Project (2023001), and Fundamental Research Funds for the Central Universities (xtr062023003).

Clinical metagenomic sequencing detects hundreds of organisms whose relevance to infection requires clinicians' manual interpretation. We retrospectively reviewed 4,855 specimens from 19 sampling sites, linked 1.33 million detected organisms to clinical culprit-pathogen determinations, and fine-tuned a large language model on these labels. The model recovered 92.44% of reference culprit calls in the held-out test set and 84.49% in a later cohort. In a blinded re-adjudication of 215 cases, the model matched 90.5% of culprit calls on which both expert reviewers agreed. Adjudication accuracy varied with organism identity and specimen types. mNGS-derived metrics such as read counts or genome coverage alone were insufficient to identify culprit pathogens, and organisms classified as common colonizers could still be judged causative. Certain organisms were frequently co-reported as causative in the same specimen. These findings support further evaluation of human and AI-assisted report review, particularly regarding missed pathogens, excess calls, and differences across specimen types.

Medical evidence synthesis increasingly requires published studies, real-world clinical data and structured biomedical knowledge, yet most automated systems remain centered on literature retrieval and summarization. Here we present MES, a multi-agent framework for source-grounded medical evidence synthesis. MES coordinates six specialized agents to decompose clinical questions, retrieve literature and trial evidence, generate question-specific RWE by constructing and analyzing real-world cohorts, query biomedical knowledge graphs, integrate quantitative findings and screen generated claims against their cited evidence. The framework uses evidence-based medicine taxonomies to clarify underspecified questions, adapts evidence use when sources are absent or discordant and reports unresolved gaps rather than forcing consensus. MES also provides claim-level provenance tracking and cross-agent consistency checking, allowing final reports to distinguish trial evidence, real-world associations and knowledge-graph support. We evaluated MES in two clinical use cases and across 144 clinical queries spanning six evidence-based medicine categories. MES generated structured reports that preserved source traceability, identified cross-source disagreement and communicated uncertainty across diverse clinical question types. MES provides an auditable framework for organizing, testing and contextualizing heterogeneous clinical evidence while keeping the evidentiary basis of each conclusion explicit.

The gut-brain axis has emerged as a critical modulator in neurological diseases, including stroke. While acute post-stroke dysbiosis is well established, it remains unclear whether these alterations persist long-term and can be therapeutically modulated. This study addresses this gap through integrated metagenomic and metabolomic profiling in two independent clinical stroke cohorts, complemented by experimental models. Experimental stroke induced persistent changes in gut microbiota composition and altered short-chain fatty acid profiles. In patients, microbiota composition remained distinct from controls up to three months after ischemic stroke or transient ischemic attack (TIA), accompanied by persistent reductions in propionic and valeric acid and reduced abundance of Prevotella-associated microbial signatures. We performed a randomized, double-blind interventional trial (PRISE) testing a multispecies probiotic versus placebo, using household members as additional controls. Probiotic intervention led to microbiome compositional changes with increased abundance of beneficial taxa and significantly improved functional recovery. The tested probiotic strains showed anti-inflammatory and antioxidant effects in vitro. These findings suggest that targeted modulation of the microbiome represents a promising therapeutic avenue to improve chronic stroke recovery, warranting validation in larger Phase III clinical trials.

Background: Host response and biomarker research to advance sepsis therapeutics requires well-characterized cohorts with longitudinal biospecimens across a range of hosts and infections. Automated screening based on SIRS criteria is sensitive but nonspecific and identifies more candidates than research teams can approach, sample, and follow longitudinally. We evaluated a structured enrichment review by non-physician study staff, done after automated screening in a hospital-wide prospective sepsis cohort. Methods: The CISID cohort is a prospective observational study of hospitalized patients with suspected sepsis, enrolling adults from the ED, inpatient wards, and ICUs at a tertiary academic medical center. Patients were identified through: (1) automated electronic health record screening for two or more SIRS criteria with a 24-hour lookback, documentation of microbiologic culture or PCR, and receipt of a new antibiotic; (2) a sepsis enrichment step of structured chart review by trained non-physician study staff; and (3) discussion with the primary clinical team. Enrolled patients underwent physician adjudication into four diagnostic categories. To evaluate the enrichment step, we analyzed all patients screened during a 9-week period, assigned a composite "sepsis-likely" variable (CDC Adult Sepsis Event criteria or ICD-10 code R65.20/R65.21 within 48 hours before or after screening), and assessed mortality across each stage of screening. We also assessed concordance of the composite with physician adjudication among enrolled patients. Results: From November 2023 through February 2026, 4,318 patients screened in through EHR-based automated screening; 1,220 (28%) were flagged for enrollment, and 288 were enrolled. Physician adjudication classified 209 (73%) as sepsis, 46 (16%) as infection without sepsis, 14 (5%) were classified as undifferentiated critical illness, and 19 (6.6%) as non-infectious sepsis mimics. Patients with sepsis were severely ill (median SOFA score, 9; ICU admission, 80%; shock, 48%), and 51% were immunocompromised; mortality was 14% in hospital and 27% at 90 days. In a 9-week sample of 907 screened encounters, 21% of those deferred and 70% of those flagged for enrollment met the sepsis-likely composite, a 3.3-fold increase in prevalence after structured study team review. Conclusions: A brief structured enrichment review after automated screening tripled the prevalence of EHR-based sepsis indicators among patients selected for enrollment, while concentrating enrollment efforts on roughly one-quarter of screened-in patients in this hospital-wide, all-source suspected-sepsis cohort. This approach may help prospective cohorts direct limited biospecimen resources toward patients most informative for host-response and biomarker research.

Human intestinal diseases, including inflammatory bowel disease (IBD), are difficult to model because existing animal and in vitro systems do not capture the integrated vascular, immune, and metabolic complexity of the intestine. Here, we establish a custom normothermic machine perfusion platform that maintains surgically resected human intestinal tissue for up to 72 hours under near-physiological conditions. Discarded specimens from patients with nonmalignant intestinal diseases, including IBD and diverticulitis, were continuously perfused at 37{degrees}C with real-time hemodynamic and metabolic monitoring. Perfusion preserved tissue viability, mucosal architecture, epithelial integrity, oxygen consumption, metabolic activity, and vascular patency. Daily perfusate sampling enabled longitudinal profiling and experimental modulation of inflammatory mediators, revealing patient-specific immune states. This long-duration platform bridges reductionist in vitro models and clinical disease, enabling mechanistic studies of intestinal inflammation and future testing of personalized therapies in intact patient-derived tissue.

Background Mupirocin, a widely used topical agent for decolonization of Staphylococcus aureus, is increasingly compromised by resistance. Although plasmid-mediated mupirocin resistance is a recognized cause of decolonization failure, its role in facilitating hospital-wide transmission is unknown. Methods We conducted genomic surveillance of S. aureus at two interconnected urban hospitals where mupirocin decolonization is routine. Genome sequencing of >10,000 isolates was integrated with patient data to identify transmission and resistance determinants. Bacterial phenotypes and fitness were evaluated in vitro and in murine colonization models. Findings Genome sequencing identified 475 hospital transmission events; none were detected by conventional surveillance. The mupA (ileS2) resistance determinant, carried on conjugative plasmids, was enriched eightfold in methicillin-resistant S. aureus (MRSA) relative to methicillin-susceptible strains. mupA was associated with nearly a threefold greater chance of hospital transmission, especially within endemic healthcare-associated MRSA lineages, and was enriched twofold in hospital-onset infections compared with admission colonizing isolates. Multiple independently evolved inactivating mutations in the essential chromosomal gene ileS1 co-occurred with mupA, creating plasmid addiction in which mupA became indispensable for bacterial survival. Addiction arose most frequently within the dominant community-acquired MRSA lineage, where plasmid carriage reduced colonization fitness in mice. Plasmid-containing strains exhibited stringent-response activation, explaining the fitness costs and collateral tolerance to disinfectants, such as ethanol and peroxide. Although addiction reduced S. aureus fitness, it increased plasmid transfer, and addicted variants spread across hosts, demonstrating adaptation that mitigates these costs. Unexpectedly, we identified a mupirocin-dependent vulnerability to isoleucine limitation, revealing a potential strategy to target mupA-mediated resistance. Interpretation Plasmids promote hospital transmission of mupirocin-resistant S. aureus and create an evolutionary trap in which antibiotic use selects for bacterial dependence on otherwise costly resistance elements. This dependence revealed a collateral bacterial vulnerability that could be exploited to target resistant strains and preserve the effectiveness of mupirocin.

Background: Medical-AI publications do more than report technical performance; they also frame AI as beneficial, uncertain, or risky. How this evaluative stance has changed across the medical literature is not well characterized. Objective: To characterize evaluative stance in published medical-AI discourse abstracts from January 2021 through April 2026 and examine variation over time, concern themes, failure mechanisms, specialties, first-author geography, and publication format. Methods: We conducted an LLM-assisted computational content analysis of medical-AI abstracts from first- and second-quartile medical journals. Of 97,492 post-cutoff Q1/Q2 records entering the prefilter, 16,759 were retained as discourse or evaluative. Claude Sonnet 4.6 assigned 16,749 valid stance classifications using Alarm, Caution, Neutral, Cautious Optimism, and Advocacy. Annual analyses used 16,747 records dated 2021-2026. Critical stance was defined as Alarm plus Caution and indexed evaluative scrutiny rather than opposition or author psychology. Each LLM step was validated against blinded human coding by one author: prefilter Cohen kappa = 0.51, stance quadratic-weighted kappa = 0.79 (95% CI 0.72-0.84) for codable, in-scope records, specialty kappa = 0.75, and mechanism-axis kappa = 0.84 for model type and 0.57 for failure mode. Results: Advocacy declined from 2.9% in 2021 to 0.6% in partial 2026, while Cautious Optimism remained the majority stance. Among 16,749 valid classifications, 30.8% were critical. Critical share increased from 25.4% to 32.6%, a 7.25-percentage-point increase based on unrounded estimates. Among critical records, patient safety remained the most prevalent concern. Hallucination/errors increased by 30.9 percentage points. Regulation declined by 22.0 percentage points and ethics/bias by 8.1 percentage points in prevalence share; these declines do not necessarily indicate lower publication counts. Within the hallucination/error theme, factual error was more common than fabrication. Fabrication estimates should be treated as an upper bound because failure-mode agreement was moderate. Specialty patterns were heterogeneous. Critical rate was inversely associated with FDA-cleared device availability (Spearman rho = -0.65, two-sided p = 0.004), which does not measure adoption, deployment, maturity, or clinical use. First-author geography described publication metadata and discourse, not national attitudes or research quality. Reviews were the least critical and most favourable format. In exploratory forward validation, 2 of 78 early Advocacy predictions were fully borne out, although the analysis was single-rater and retrieval-dependent. Conclusions: Published medical-AI abstracts became modestly less promotional and more focused on specific errors and safety concerns. Unqualified promotion declined, but qualified favourable framing remained dominant, and the rise in critical stance was modest. Concern moved toward errors and patient safety, with factual error discussed more often than fabrication. These findings describe published discourse, not AI capability or whether the evaluations were correct.

Large language models now match clinicians on medical-knowledge benchmarks. Whether they can safely triage a patient message is a separate question, and the highest-consequence failure in triage is the emergency that never gets escalated. We tested two competing explanations for that failure. The first is sycophancy: the model defers to a patient who downplays a danger sign. The second is atypicality: the model reads an early, mild, or atypical presentation of a time-critical disease as benign. To isolate the effect of a model-agnostic cannot-miss discipline, we performed a prompt-level ablation on a fixed frontier model (Claude Opus 4.8), then replicated the same instruction-level intervention across eight models from two families (Claude Opus 4.8, Sonnet 5, Haiku 4.5, and Fable 5; Gemini 3.1 Pro-Preview, 3.6 Flash, 3.5 Flash-Lite, and 3.1 Flash-Lite) under a released harness, and released the whole apparatus as an open, model-agnostic evaluation suite. Minimization turned out to be harmless when danger is overt. Across 40 unambiguous emergencies rendered in five framings on Opus 4.8 (neutral, mild minimization, strong denial, third-party reassurance, and a benign distractor), sensitivity stayed at or above 97.5% in every arm and was flat across framings, and seven of the eight models held 97.5% to 100% on the neutral set (the eighth completed 13 of 40 stems, all escalated). Atypicality is where the hazard lives. Across 76 subtle or atypical emergencies, each adjudicated by a blinded, LLM-simulated three-physician panel and each mapped to a published guideline that names its cannot-miss diagnosis, the base model failed to escalate 13.2% (95% CI 7.3 to 22.6) to the top acuity tier; the governance prompt cut this to 3.9% (1.4 to 11.0; paired exact McNemar 7 to 0, p = 0.016; relative risk 0.30; number needed to treat 11). The effect sits in the harder half of the cohort (36 original stems: 22.2% to 8.3%, 5 to 0; 40 extension stems: 5.0% to 0%, 2 to 0), and its magnitude holds when the three stems an independent panel rated below emergency are excluded (9.6% to 2.7%, 5 to 0, p = 0.063). A one-line distillation reproduced most of the effect, showing that the underlying reasoning principle does not depend on proprietary prompt language. Its residual failures and weaker specificity behavior also show that portability of the principle is not equivalent to deterministic enforcement. The effect replicated in Spanish (27.8% to 11.1%, McNemar 6 to 0, p = 0.031). The effect replicated in a second, released harness on the same model, and then across eight models. In the eight-model replication, baseline under-triage ranged from 2.6% to 34.2% by model and elicitation. The same prompt text, unchanged, was applied above eight models from two vendors, three capability tiers, and two elicitation modes; under-triage fell in every cell where there was material hazard to remove, and no cell of any model gained a false emergency alarm. The identical governance prompt reduced it in seven of twelve model-by-elicitation cells, moved a single stem in three, and was neutral or reversed in the two cells already at the floor. With the stem as the unit of analysis, governance was favored on 19 stems and control on 5 across the eight reasoning-permitted models (sign test p = 0.007); pooled discordant pairs were 27 to 8 (p = 0.002) under forced-choice elicitation in the Claude family and 31 to 10 (p = 0.001; paired risk difference 3.5 points, 95% interval 1.5 to 5.7) under reasoning-permitted elicitation across both families. The misses concentrate: 42 of the 76 stems were never missed by any model in either arm, ten stems account for 71% of all baseline misses, and a handful of conditions (subacute endocarditis, early ectopic pregnancy, early urosepsis, mesenteric ischemia, early appendicitis) are missed by most models and resist most of the prompt's effect. On a 44-stem panel-adjudicated non-emergency cohort, the full prompt added zero escalations to EMERGENCY in any model under any elicitation (0 of 44 per cell; Wilson upper bound 8.0%), and it moved some ROUTINE stems to same-day URGENT (17 added against 2 removed across eight models, p < 0.001). Across more than 5,900 triage classifications, no emergency was ever routed home in any arm; the failure mode is uniformly EMERGENCY to URGENT. The prompt changed nothing on knowledge (MedQA 95.8% in both arms, 4 to 4 discordant), adversarial robustness, or medication-hazard detection, while raising named-guideline citation from 65.6% to 96.9% at roughly 2.3 times fewer output tokens; HealthBench fell 3.6 points on a single seed, concentrated in context-seeking and uncertainty-handling behaviors, and warrants follow-up. Elicitation mattered: forcing an immediate single-shot disposition raised baseline hazard by roughly ten percentage points in two of four Claude models, and the governance increment in the Claude family was largest under that elicitation. Quiet emergencies, defined here as early, mild, or atypical presentations of time-critical conditions, were the principal source of under-triage in this constructed cohort rather than overt emergencies accompanied by patient minimization. A prompt-level cannot-miss discipline reduced that hazard across a heterogeneous set of models without adding false emergency escalations in the primary specificity cohort. However, residual failures remained concentrated, stochastic, dependent on model choice and elicitation, and undefined when output could not be parsed. These findings show both the portability of the reasoning discipline and the limits of implementing it through instructions alone. MedCanon is presented as a hybrid clinical runtime with a deterministic safety spine designed to convert these distributional improvements into versioned, testable invariants. The runtime itself was not evaluated in this study. The Level 2 evaluation specified here is required before claims about its clinical safety effect, determinism, or auditability can be made. Ground truth came from LLM-simulated panels triangulated against an independent cross-family panel, published guideline anchors, and an externally authored vignette set; independent board-certified human adjudication is required before any clinical-deployment claim.

Human tissue transcriptomic studies frequently rely on biopsies obtained during surgery under general anesthesia, yet the impact of perioperative medications on gene expression has not been systematically evaluated. Using paired liver biopsies from patients with obesity at risk for metabolic dysfunction-associated steatohepatitis (MASH), we compared RNA sequencing profiles from biopsies collected under local anesthesia months before surgery with matched intraoperative biopsies obtained under general anesthesia. Intraoperative samples exhibited striking transcriptional changes enriched for canonical glucocorticoid receptor target genes, implicating perioperative dexamethasone administration as the principal driver. In a matched cohort in which dexamethasone was administered immediately after biopsy collection, we directly demonstrate that dexamethasone profoundly alters hepatic gene expression during surgery. These findings were independently validated and shown to substantially distort associations between liver fibrosis stage and inflammatory gene expression, obscuring biologically relevant immune signatures. Similar glucocorticoid-responsive transcriptional changes were observed in adipose tissue, indicating a systemic effect. Our findings identify perioperative dexamethasone exposure as a previously unrecognized but readily modifiable confounder in human tissue transcriptomics. As biopsy-based bulk and single-cell transcriptomic studies continue to expand, careful documentation and consideration of perioperative medication exposure will be essential to improve the reproducibility, interpretability, and biological validity of human gene expression studies.

Background: Prior research on symptom clusters (SCs) in colorectal cancer (CRC) has relied primarily on patient-reported outcome surveys, which capture symptom experience at discrete assessment points rather than the continuous documentation generated during routine care, leaving open whether SCs derived from electronic health record (EHR) text carry the same clinical meaning and predictive value. To construct and validate patient-level symptom co-occurrence networks from large language model (LLM), extracted symptom data in CRC patients, and to test whether resulting SCs predict clinical outcomes. Methods: Using a zero-shot LLM extraction pipeline previously benchmarked against a manually annotated ground truth (Macro F1=0.70 for the best-performing model), we extracted 46 symptoms from 2,728 discharge notes of 1,507 CRC patients in MIMIC-IV. Patient-level symptom co-occurrence networks were constructed independently from Gemini 3.5 Flash and Claude Haiku extractions using phi correlation and Louvain community detection, with sensitivity analyses across correlation thresholds, random seeds, note-aggregation strategy, and bootstrap resampling. Per-cluster symptom burden scores were tested as predictors of in-hospital mortality, 30-day readmission, and 1-year mortality using logistic regression adjusted for age, sex, and (in sensitivity models) metastatic disease. Results: Both LLMs' networks converged on three clinically coherent SCs (Systemic, Gastrointestinal, and CRC Disease-Specific) across sensitivity analyses (cross-model Adjusted Rand Index=0.727; 100-seed Louvain ARI=0.985; bootstrap ARI=0.733). The Systemic Symptom Cluster was the most consistent predictor of in-hospital mortality (OR=1.33) and 1-year mortality (OR=1.41), while the CRC Disease-Specific Cluster specifically and independently predicted 30-day readmission (OR=1.20); both associations were robust to adjustment for metastatic disease. Conclusion: LLM-extracted symptom data recover clinically coherent, reproducible SCs from unstructured discharge notes that carry independent prognostic value for mortality and readmission, supporting the clinical validity of automated, EHR-derived symptom profiling in CRC.

Background: Recent work has brought large language models (LLMs) into clinical variant interpretation. In most existing approaches, the model is confined to text-level operations, retrieving, extracting, compacting, and summarizing heterogeneous evidence, while the genuine genetic evaluation is delegated to a separate deterministic engine. Other tools employ LLMs to mine functional evidence for a single ACMG criterion, to rerank a classifier's output, or to answer isolated variant questions. Integrative genetic reasoning, by contrast, demands multi-source evidence weighing for each variant, reconciliation of inheritance with observed zygosity and parental genotypes, joint interpretation of variants in the same gene, and the derivation of a case-level conclusion. Moreover, to protect patient privacy, this must be done on hospital premises. Methods: o build an end-to-end variant interpreter that runs on local infrastructure, we benchmarked 13 small language models (SLMs) on 150 multiple-choice genetics questions and selected Qwen3.5-9B. In IVA (Intelligent Variant Analyser), a four-stage pipeline, the model assigns ACMG/AMP criteria, applies de novo, dominant, recessive, and X-linked inheritance logic to trios, and writes a structured report, while deterministic code handles retrieval, phasing, point summation, citation checking, and formatting. We evaluated IVA on synthetic scenarios, 16 solved trios, and 327 solved singleton cases of visual impairment, on a single GPU. Results: On the benchmark, Qwen3.5-9B reached a mean accuracy of 0.81, compared with 0.89 for a 1.6-trillion-parameter cloud model. In the 16 trios, IVA reported every previously identified pathogenic, likely pathogenic, or uncertain-significance variant of interest (sensitivity 100%; precision 39.5%). In the 327 singleton cases, it recovered 93% of the expert-reported causative variants and classified 82% of them as pathogenic or likely pathogenic (precision 43%). Failures fell into four recurrent classes: omission, unverified mechanism, mismatch between reasoning and label, and arithmetic errors. Conclusions: A small language model running on local hardware can serve as the reasoning layer for case-level and trio-based variant interpretation when it is surrounded by deterministic safeguards. Because its outputs are not deterministic, analyst review and prospective multicenter validation remain necessary.

General-purpose language models generate fluent health reports that can fabricate derived clinical metrics. In an illustrative comparison on identical two-week CGM and meal data, leading foundation models produced reports with invented MAGE values, inflated meal counts, and unreferenced complication-risk projections: failures invisible to non-expert readers and plausible enough to mislead clinicians. We describe the HPP Personal Health Agent (PHA), a metabolic health agent that grounds generation in four layers: the Human Phenotype Project (HPP), a deep-phenotyped cohort of 13,000+ participants supplying population references and trained predictive models; 21 domain-expert tools and trained-model wrappers that compute clinical metrics and risk predictions; declarative behavioural skills that constrain what the model may claim; and 21 automated evals across 8 categories developed via a test-driven cycle in which each eval encodes a failure mode discovered during iterative development. In a 210-report matrix (14 participants x 3 prompts x 5 system conditions), the gains are largest on the system's primary use case (meal-grounded metabolic reports, the report it was designed for), where the full system raises a deterministic form/provenance score from 0.37 (the same foundation model with no tools or skills) to 0.91; this score measures structural completeness, numerical accuracy, tool grounding, and clinical-language compliance: a necessary condition for trustworthy health reporting, with clinical quality as a complementary axis examined qualitatively. A skills-vs-tools decomposition shows the two layers act on different axes: tools drive numerical accuracy (from about 14% to 90% of reported metrics correct), while the declarative skills add most of the remaining gain in citations, completeness, and structure (tools alone recover only part of the gap, 0.49 from the same 0.37 baseline). The lift generalises beyond the primary use case: to a second metabolic prompt (0.72) and a cardiovascular extension (0.70), each from a 0.37-0.39 baseline. The architecture extends across clinical domains: adding a SCORE2 cardiovascular risk tool and a corresponding skill (with no changes to orchestration, eval harness, or existing tools) produced a cardiovascular risk report from the same system. Trustworthy domain-specialised health AI is a systems design problem: deep-phenotyped cohort data, domain-expert tools and models, and eval-driven development together form a replicable pattern.

Clinical AI agents may be assigned to individual patients, but hospital resources are shared across many patients. We tested what agents do when helping their assigned patient would violate the hospital's rule for a scarce resource. We analyzed 22,916 simulated cases comprising 274,992 logged agent actions across 20 AI models. In each scenario, the agent could claim a scarce resource for its patient even though the hospital rule gave another patient priority. We varied only the agent's assigned role, from responsibility for the whole ward to strong advocacy for one patient. Violations of the hospital rule rose from 32.5% under whole-ward responsibility to 69.4% under strong patient advocacy, a 36.9-point increase (95% CI, 25.7-48.0). Agents correctly identified which patient should receive the resource in 95.7% of tests, yet still took it for their own patient in 65.9% of those episodes. Asking the agent to apply its own allocation judgment immediately before acting reduced violations to 0-2% in a three-model follow-up experiment. Assigned roles can shape how clinical AI agents use shared hospital resources, even when they identify the correct priority patient. Patient-focused agents should not independently control shared resources without an allocation check.

Artificial intelligence (AI) has the potential to transform healthcare, with advanced multimodal approaches showing great promise in leveraging diverse health-related data. Here, we applied multimodal AI to entire electronic health record (EHR) and complete pathogen genome data to predict patient outcomes from life-threatening infection. An automated, scalable pipeline was developed for EHR data preprocessing, quality control, and standardisation. A deep learning fusion model was trained to predict in-hospital mortality, need for ICU admission, prolonged length of stay and 30-day unplanned readmission. We then developed a novel genomic large language model (gLLM) architecture to incorporate bacterial genomic features into the multimodal fusion model. The cohort comprised 2,656 bloodstream infection hospitalisations involving 2,535 patients. Deep learning fusion models using entire structured and unstructured EHR data outperformed traditional APACHE II score mortality prediction (AUROC [95% confidence intervals] 0.93 [0.92-0.94] versus 0.77 [0.77-0.78]). The model also showed strong performance for predicting the need for ICU admission (AUROC 0.978 [0.966 - 0.986]), prolonged hospital length of stay (AUROC 0.803 [0.790 - 0.812]) and unplanned readmission (AUROC 0.696 [0.690 - 0.701]). As proof of principle, incorporating entire microbial genomic features from the causative pathogen further enhanced prediction and enabled identification of key bacterial virulence pathways relevant for human disease. Multimodal AI integrating harmonised EHR and genomic data can accurately identify hospitalised patients at risk of poor outcomes. These approaches are scalable to other subspecialities of medicine.

Abstract Metagenomic sequencing can detect a broad range of pathogens, but interpreting which detections are clinically relevant requires expert adjudication that is difficult to scale and standardize. Here we present diagnostic classifiers that formalize expert adjudication by combining structured decision trees with large language model reasoning to assign diagnoses and select pathogen candidates. We first developed a short-read metagenomic assay for sterile-site specimens (cerebrospinal and ocular fluid) in the META-GP study (Victoria, Australia, 2024-2025) and evaluated classifiers on a validation dataset (n = 96; clinical samples, spike-ins and controls). Locally deployed, open-weight reasoning models (Qwen3) achieved diagnostic performance comparable to expert consensus, improving with clinical context (n = 79, above experimental limit-of-detection; without clinical notes, 94.4% sensitivity, 95.4% specificity; with clinical notes, 97.2% sensitivity, 100% specificity). Automated adjudication enabled systematic benchmarking of computational parameters and regression testing for pathogen detection tasks. In a heterogeneous development cohort (n = 78), reviewers and classifiers identified clinically significant pathogens missed during routine testing. By reproducing consensus detections without requiring a full review panel, diagnostic classifiers enable scalable, standardized metagenomic interpretation that complements expert adjudications.

BACKGROUND & AIMS: Inflammatory bowel disease (IBD) is often diagnosed months to years after symptoms begin, and diagnostic delay is associated with complications. We tested whether longitudinal machine learning using electronic health record (EHR) data can identify incident IBD before diagnosis, and quantified lead time and deployment burden. METHODS: We studied adults with an IBD-compatible symptom, sign, or finding at UCSF (2012-2024). Incident cases were confirmed by structured filtering and validated large language model note review. We trained a state space model (Mamba) and comparators on pre-diagnostic EHR data, excluding diagnostic-workup features, and assessed discrimination at 6- and 12-month landmarks. We externally tested Mamba without recalibration at ZSFG, a safety-net hospital serving a demographically and socioeconomically distinct population, and simulated diagnosis-code-only deployment at UCSF. RESULTS: Six months after presentation, among patients still undiagnosed (67% of UCSF test cases), Mamba achieved an AUROC of 0.969 (95% CI, 0.928-0.999) at UCSF and 0.965 (0.952-0.976) at ZSFG. Median lead time was 20 months at UCSF and 15 months at ZSFG. Among future cases with an initially normal colonoscopy (N=32) or a histological finding of acute inflammation without chronicity (N=63), the model flagged 84% and 89% of these patients, respectively, before those procedures, indicating the potential for pre-diagnostic detection. Among 358,886 symptomatic patients not used for training (60 incident cases), the diagnosis-code-only model achieved an AUROC of 0.926; reviewing 20 alerts/month identified 34 cases (PPV, 1.09%; 64-fold above baseline). CONCLUSIONS: Longitudinal machine learning can identify increased IBD risk months to years before diagnosis and may define a population for prospective studies of early IBD biology. Prospective evaluation should determine whether earlier identification reduces diagnostic delay.

Clinical AI deployment rarely reaches the bedside because both workstation integration and governed backend provisioning are challenging. We introduce a workflow that allows rapid and governed access to AI functionalities through a browser interface or API, enabling privacy-compliant data analysis. The complex setup for compliant healthcare applications such as a secure backend, private AI model endpoint, and managed identity configuration is encoded by itself, a concept known as Infrastructure as Code (IaC). IaC is a concept in which networks, access rights, and service endpoints are written down as machine-readable files that can be reviewed once and executed identically in any institution. Here, we apply this concept to a clinical use case. The process of deploying a governed, private clinical AI solution using this approach in a tertiary care center took, on average, 310.3 +- 196.6 seconds (n=25, mean +- SD). To further demonstrate the clinical value of this approach, we developed and deployed an LLM-based inconsistency checker, which identifies textual inconsistency in medical documents. When evaluating the tool on a set of 1,347 physician-approved clinical documents, the tool flagged a clinically severe inconsistency in 105 clinical documents (7.8%, approximately 1 in 13). In a severity-stratified physician audit, 149 of 150 sampled findings were confirmed (population-weighted PPV 98.2%; CI 94.8 - 100.0). In summary, we present the application of IaC frameworks to a healthcare use case and demonstrate that clinically useful AI applications can be deployed to physician workstations in hospital-governed cloud environments within a few minutes.

Background Improved understanding of how variants cause breakthrough infection, despite pre-existing immunity, is needed to advance development of next-generation SARS-CoV-2 vaccines, including those that may provide cross-variant protection or block transmission. SARS-CoV-2 controlled human infection models (CHIMs) may therefore identify correlates of protection and accelerate the development of new interventions. Methods Healthy vaccinated adults aged 18-30 years were inoculated intranasally in a stepwise dose-escalation CHIM with doses from 1x102 TCID50 to 1x106 TCID50 of SARS-CoV-2 Delta variant. Within the 1x106 TCID50 group, participants were selected for serum neutralising antibody titres (NT50) less than or equal to 1:80. Post-inoculation, participants were quarantined for up to 14 days. Outpatient follow-up continued for 12 months. The primary aim was to elicit safe, well-tolerated Delta SARS-CoV-2 breakthrough infection at a rate of over 50%. Findings Forty-six participants were inoculated; 22 during dose-escalation with no resultant infections, and 24 at the highest dose, of whom 18 were screened for low serum neutralising antibodies. Sustained infection with mild-to-moderate symptoms occurred in 33% (6/18) of the sero-selected group, with highly variable viral loads, viral emissions and symptoms. Serum neutralising antibody, anti-N IgG and to a lesser extent, mucosal anti-S IgA and N-specific T cells most strongly predicted protection from virologically-defined infection. Higher neutralisation, serum and nasal anti-N IgG level, and N-specific T cell responses correlated with lower viral load, while baseline nasal anti-S IgA was associated with lower symptom scores. Transient infection was additionally observed in 6 participants and was associated with higher baseline N-specific T cell responses than those that developed sustained infection. Interpretation Susceptibility to SARS-CoV-2 breakthrough infection in those with hybrid immunity is strongly associated with low levels of pre-existing antibody, but the diversity of immune markers associated with protection implies that additional benefits may be conferred by multi-pronged immunity.

Biologically rich clinical trials generate heterogeneous datasets whose relationships and clinical context must be preserved. Here we present the POETIC-AI environment, a trial-centric digital environment for integrating, analysing, and exploring multimodal data from POETIC, a randomised phase III trial of peri-operative aromatase inhibition in hormone receptor-positive breast cancer (NCT02338310). We implemented a graph-based ontology linking patients, treatments, sampling timepoints, specimens, assays, images, spatial regions, and derived measurements; and designed code-based and graphical analytical workflows. Using proliferation as a test case, we reproduced POETIC's established reduction in Ki67% and examined variation underlying sample-level measurements. In response-enriched cohorts, mean Ki67% decreased while relative variation among pathologist-assessed fields increased (paired Wilcoxon, P < 0.0001). Pathologist-assessed Ki67% correlated strongly with an 18-gene transcriptomic proliferation signature (r = 0.834), and both decreased significantly during treatment (P < 0.0001). An annotation-free pixel-wise Ki67 H-score showed weaker correlations and a similar but non-significant reduction (r = 0.327-0.472; P = 0.15), highlighting differences between tumour-focused and whole-section image quantification. Digital pathology and spatial molecular measurements remained linked to their source images and annotated tissue regions, enabling exploration within their morphological and spatial context using bespoke interactive applications. POETIC-AI demonstrates how the structure, provenance, and tissue context of evolving multimodal clinical-trial data can be preserved within a connected research environment, supporting reproducible analyses and the laying the groundwork for future development of biologically interpretable computational biomarkers.