Citations

Background: Large language models (LLMs) are increasingly used in patient-facing clinical applications, creating a need for scalable methods to evaluate the accuracy, safety, and appropriateness of their responses. Although physician review remains the reference standard, it is resource-intensive and difficult to scale. Objective: To develop and clinically validate an LLM-as-a-judge framework for evaluating responses to patients’ periprocedural questions. Methods: We retrospectively evaluated 1285 patient–AIVA interactions from 112 patients across two Mayo Clinic sites and multiple surgical specialties. Physician reviewers classified interactions using a predefined true-positive (TP), false-negative (FN), true-negative (TN), and false-positive (FP) framework. The finalized LLM judge independently evaluated the same interactions. Agreement was assessed using four-class and category-specific agreement, Cohen’s kappa, and a secondary binary analysis of response correctness. Results: Overall, four-class agreement was 93.1% (95% CI, 91.7–94.5%), with Cohen’s κ = 0.852 (95% CI, 0.823–0.882). Category-specific agreement was 94.1% for TP, 90.7% for FN, 95.5% for TN, and 84.9% for FP classifications. In the binary correctness analysis, accuracy was 93.4% (95% CI, 92.0–94.7%), sensitivity 94.6%, specificity 90.2%, precision 96.2%, and F1-score 0.954. McNemar’s test demonstrated no significant asymmetry between paired classifications (p = 0.11). Conclusions: A clinically grounded LLM judge closely approximated physician evaluation of patient-facing AI responses. These findings support its potential as a scalable assistive monitoring tool while preserving physician oversight for uncertain or safety-sensitive interactions.

Abstract Recent evidence implicates somatic mutations in disease-associated genes across multiple neurodegenerative disorders, including the identification of two somatic TARDBP variants as a putative cause of frontotemporal lobar degeneration with TDP-43 pathology type C (FTLD-TDP type C). However, these findings remain anecdotal. We performed single-nucleus amplicon sequencing of 39,800 neurons from the superior temporal gyrus of sporadic FTLD-TDP type C patients (34,738 neurons, 52 individuals) and non-demented controls (5,062 neurons, 26 individuals) using the Mission Bio Tapestri platform and investigated somatic variants in TARDBP and other FTLD-TDP-associated genes in a case-control setting. We uncovered an extensive landscape of ultra-low-frequency (<1%) somatic mutations across all analysed genes, including TARDBP . Rare somatic occurrences of known ALS/FTD-associated germline TARDBP variants, including p.N267S and p.M337V, were identified exclusively in patients. Among all targeted genes, TARDBP had the highest proportion of neurons harbouring a somatic variant, and this mutational burden was lower in individuals who died at an older age. Strikingly, the C-terminal domain of TDP-43, where most ALS/FTD germline pathogenic variants reside, exhibited a lower neuronal mutational burden than its other domains. This suggests that neurons harbouring damaging mutations may undergo progressive loss prior to autopsy, leaving only rare survivors at frequencies too low to detect by bulk or small-scale single-cell approaches. Together, our findings provide an unprecedented resolution into the somatic mutational landscape of post-mitotic neurons in FTLD-TDP type C and reveal a gene-specific layer of neuronal mosaicism that current single-cell WGS studies fail to detect.

Background: Reliable access to postoperative guidance between clinic visits remains inconsistent, and the usefulness of large language models depends on how clinical knowledge is structured and retrieved. Knowledge graph-only (KG-only), vector retrieval-augmented generation (RAG-only), and hybrid graph-retrieval (GraphRAG) architectures have not been directly compared under controlled conditions for postoperative care. Objective: It was to compare these three pipelines on a curated postoperative question-answering benchmark using automated grounding metrics and blinded clinician review. Methods: A benchmark of 250 expert-authored questions (200 in-scope, 20 emergency-intent, 30 out-of-scope) was evaluated on an identical knowledge corpus using Gemini 2.5 Flash for generation and Gemini Embedding-001 for retrieval. Three blinded clinicians rated all 750 outputs; the primary outcome was majority-consensus accuracy across all 250 questions. GraphRAG is a multi-component pipeline that also includes query expansion, a second retrieval stage, evidence fusion, and a monotonic RAG fallback, and is therefore evaluated as a complete hybrid system rather than as graph retrieval in isolation. Results: Overall accuracy was 96.8% (95% CI: 93.8–98.4) for GraphRAG, 94.8% (95% CI: 91.3–96.9) for RAG-only, and 56.0% (95% CI: 49.8–62.0) for KG-only. Both retrieval-augmented systems outperformed KG-only (adjusted p < 0.001). The GraphRAG versus RAG-only difference was 2.0 percentage points, arose from five discordant question pairs and did not reach the Bonferroni-adjusted threshold (exact McNemar p = 0.062; adjusted p = 0.188). The monotonic RAG fallback determined 44.8% of GraphRAG outputs (112 of 250 queries). GraphRAG showed higher context recall (0.851 versus 0.796; adjusted p < 0.001), whereas RAG-only showed higher context precision. Estimated hallucination risk, derived from an LLM-as-a-judge procedure using Gemini 2.5 Flash rather than independent clinical adjudication, was 0.108, 0.132, and 0.436, respectively; the GraphRAG versus RAG-only difference was not significant (adjusted p = 0.771). Inter-rater agreement was high (Fleiss’ kappa = 0.854). Conclusions: Retrieval augmentation, rather than the choice of retrieval architecture, accounted for most of the difference observed. The incremental advantage of GraphRAG over a well-implemented RAG-only system was small, was not statistically significant after correction and reflects the complete hybrid system rather than graph augmentation alone. Emergency-intent accuracy was 75.0% for both retrieval-augmented systems; the architecture is therefore suited to supervised clinical workflows with clinician review, not autonomous patient-facing triage. Prospective validation on authentic patient messages and formal component ablations are required.

Abstract The paradigm in the field of oncolytic virotherapy proposes that tumor cell killing by an oncolytic virus (OV) culminates in the priming of antitumor CD8+ T cells. However, this ignores the impact a highly immunodominant antiviral response against the OV has on the subdominant antitumor response, particularly with regard to the naturally low affinity of tumor antigens to their cognate T-cell receptors. To investigate this experimentally, we first studied the interaction between the two responses in B16 and MC38 tumor models, both expressing ovalbumin (OVA). Contrary to the accepted paradigm, the inflammatory vesicular stomatitis virus (VSV) failed to expand OVA-specific T cells in vivo unless armed with tumor-associated antigen (TAA). In fact, treatment with the control VSV led to the ablation of antitumor T cells and a shift toward dysfunction as determined by the upregulation of markers, including LAG3 and CD39, correlating with lower rates of survival. Combination with anti–PD-1 immune checkpoint blockade therapy significantly improved survival only when VSV expressed TAA. VSV-expressing altered peptide ligand variants of tumor antigens, which evade mechanisms of central tolerance, expanded near self-antigen–specific T cells with the ability to cross-react with tolerized tumor antigens, thus providing an antitumor effector population capable of proliferating in the presence of a potent antiviral response. Taken together, these findings offer a rational approach to OV design which preserves acute antitumor therapy provided by antiviral innate and T-cell responses while promoting the induction of tumor-specific T cells. Significance: Oncolytic viruses can induce an antiviral T-cell response that occludes an antitumor response, challenging the paradigm in the field and elucidating strategies to optimize the efficacy of oncolytic viroimmunotherapy.

Abstract T-cell immune tolerance is established in part through the activity of the autoimmune regulator (AIRE) transcription factor in medullary thymic epithelial cells (mTEC) of the thymus. AIRE induces expression of peripheral tissue–specific self-antigens for presentation to naïve T cells to promote activation/deletion of autoreactive T cells. This traditional role of AIRE in mTECs is to prevent autoimmunity. Herein, we demonstrate that tumors mimic the role of AIRE in mTECs to evade immune rejection. We found that AIRE induced a profile of “selfness” at the RNA and protein levels which, when presented on major histocompatibility complexes (MHC), shielded the tumor from inherently self-tolerized T cells. Moreover, we describe an in vivo immunotherapy in which engineered changes in AIRE expression in tumor cells altered their profile of selfness, exposing both AIRE-modified and parental unmodified tumor cells to T-cell attack. Therefore, by resetting the immunologic selfness of cancer cells, this AIRE-mediated immunotherapy (i) converted a highly tolerized T-cell compartment into a tumor-reactive T-cell population; (ii) conferred upon nonimmunogenic tumors de novo sensitivity to immune checkpoint blockade; (iii) removed the need to identify potentially immunogenic tumor-associated antigens as targets for generation of T-cell responses; and (iv) lead to potent T cell–mediated rejection of aggressive, immunologically cold, nonimmunogenic tumors. Patient RNA sequencing data showed that the expression of AIRE predicted response to immune therapies with a strong correlation between AIRE expression and markers of T-cell receptor signaling, suggesting that our studies have therapeutic translational value.

Abstract Background Online hemodiafiltration (HDF) enhances the removal of middle molecules, including large-middle molecules, through the combined action of diffusion and convection. Randomized controlled trials (RCTs) and individual participant data (IPD) meta-analyses demonstrate improved cardiovascular outcomes and, importantly, a reduction in all-cause mortality when HDF is delivered with sufficiently high convective volumes in post-dilution mode. These findings underscore that an adequate convective dose is a critical determinant of clinical benefits compared to high-flux hemodialysis (HD). However, translation into consistent clinical practice remains heterogeneous. Therefore, the Italian Society of Nephrology convened an international, multidisciplinary panel to synthesize randomized evidence comparing online post-dilution HDF with high-flux HD in adults receiving maintenance hemodialysis. Methods The critical outcomes included all-cause and cardiovascular mortality, whereas the important outcomes included selected patient-reported health domains, including physical function, patient-perceived cognitive function, pain interference, social participation, and health-related quality of life. The certainty of the evidence was appraised using the GRADE methodology. Structured guidance recommendations and best practice statements were developed through a consensus. Sensitivity analyses were performed to assess the robustness of the observed mortality effect. Results High-volume post-dilution HDF reduces all-cause mortality (hazard ratio 0.84; 95% CI 0 74–0.95; moderate-certainty evidence) and cardiovascular mortality (hazard ratio 0.78, 95% CI 0.64–0.96; moderate-certainty evidence; data from an individual participant data analysis of trials of high volume HDF versus high-flux HD). Emerging evidence suggests slower deterioration in selected patient-reported domains, including physical function, patient-perceived cognitive function, pain interference, and social participation (low-certainty evidence). Safety with on-line post-dilution HDF has been demonstrated to be similar to that of high-flux HD. Conclusions Online post-dilution HDF should be regarded as a dose-dependent therapy. We recommend its use to reduce cardiovascular mortality and suggest high-volume HDF to reduce all-cause mortality and slow deterioration in selected patient-reported health domains. Effective implementation in clinical practice requires a targeted and individualized approach with careful optimization and monitoring of convective volume, along with appropriate treatment time, blood-flow rate, supporting infrastructure, and consideration of patient preferences. This guidance provides evidence-based recommendations and best practice statements for its use.

Abstract Genetic studies have significantly advanced our understanding of tauopathies, yet the genetic aetiology of Pick’s disease, a rare 3-Repeat tauopathy, remains unclear. The MAPT p.A152T variant has been identified as a risk factor for Alzheimer’s disease and progressive supranuclear palsy, but its role in Pick’s disease is unknown. In this study, we examined the prevalence of MAPT p.A152T in the largest series of neuropathologically confirmed Pick’s disease cases to date (n = 401). Through genotyping, we identified a single mutation carrier in the Pick’s disease cohort (minor allele frequency = 0.12%). We previously reported MAPT p.A152T at a 0.20% frequency in healthy controls (n = 2456), suggesting that it does not associate with 3-Repeat tauopathy risk. To further investigate the effect of the variant on MAPT transcript expression, we used bulk RNA sequencing in Alzheimer’s disease and progressive supranuclear palsy A152T mutation carriers. We did not detect significant differences in 4-Repeat tau levels, though preliminary trends may indicate more nuanced effects that need to be examined with long-read sequencing in a larger series. Overall, our study suggests that MAPT p.A152T does not increase Pick’s disease risk and may instead be linked to 4-Repeat or mixed tau pathologies, warranting further functional investigation.

Abstract Background Central venous catheterization (CVC) is a fundamental procedural competency in pulmonary and critical care medicine (PCCM). Although commercially available task trainers are effective for teaching CVC placement, their high cost and logistical requirements can limit access to repeated practice. Objective To evaluate learner and faculty perceptions and usability of a newly developed CVC simulation model compared with a standard commercial task trainer and detail the technical report of the novel model’s construction. Methods We conducted a prospective educational evaluation during the PCCM CVC workshop. Participants practiced CVC placement with both the standard commercial trainer and a locally constructed simulation model developed at our institution. Teaching faculty were asked to evaluate both models from an instructional standpoint. Results Eighteen participants (15 PCCM 1st year fellows and 3 faculty instructors) evaluated the novel CVC simulation model. All fellows (n = 15, 100%) preferred the new model over a standard commercial model, with 14 (93%) reporting improved understanding of CVC insertion steps and 11 (73%) noting ease of ultrasound landmark identification. All fellows (n = 15, 100%) endorsed inclusion of the model in future workshops. Additionally, faculty instructors preferred the novel model for teaching (n = 3, 100%). All faculty supported future use of the model. The initial model development cost was approximately $288.96 with most materials’ reusable. Conclusions The novel task-aligned CVC simulation model offers a realistic, reusable, and cost-effective alternative to commercial simulators, with high learner and faculty preference and strong educational value. Its functional alignment supports integration into fellowship training, with future studies needed to assess objective skill acquisition and retention.

Abstract Purpose: VV1 is designed to induce selective oncolysis of tumor cells and amplify cellular antitumor immune responses. This open-label, phase I, multicenter clinical trial assessed the safety and tolerability of VV1 monotherapy administered intratumorally or intravenously or administered intravenously in combination with avelumab. Patients and Methods: Patients with advanced solid tumors received intratumoral (n = 27) or intravenous VV1 monotherapy (n = 33) or intravenous VV1 plus avelumab (n = 16). Infusion durations (15–180 minutes) and viral dose (1.7 × 1010 or 1 × 1011 TCID50) were also evaluated. Study objectives included VV1 safety and tolerability, pharmacokinetics, pharmacodynamics, preliminary efficacy, and immune responses. Results: Intratumoral and intravenous VV1 were well tolerated, and injection reactions were infrequent. The most common adverse events of any grade related to VV1 were cytokine release syndrome (58%), fatigue (33%), and decreased lymphocyte count (32%). Virus infection of tumors was confirmed by NIS imaging and increases in serum levels of virally encoded interferon-β (IFNβ). Histologic tumor analysis at 28 days after intratumoral VV1 administration indicated increases in tumor-infiltrating immune cells. Two durable partial responses were recorded: a patient with pheochromocytoma after one intravenous dose of VV1 and a patient with thymic cancer in the combination arm. Stable disease was observed in 17 of 49 patients (34.7%) who received intravenous VV1. Conclusions: This first-in-human study demonstrates that intratumoral or intravenous VV1 administration had an acceptable safety profile as monotherapy and intravenously in combination with avelumab. There is preliminary antitumor activity. Intratumoral VV1 plus cemiplimab is now showing promise in the neoadjuvant setting. Significance: VV1 is a tumor-selective oncolytic vesicular stomatitis virus engineered to express IFNβ and the thyroidal NIS reporter gene. In this phase I study, we have demonstrated that a single administration of VV1, as a monotherapy or with an immune checkpoint inhibitor is safe, infects tumor lesions resulting in intratumoral infiltration of immune cells, and exhibits early signs of antitumor activity in patients with advanced unresectable and metastatic solid tumors.

Importance TMEM106B is a frontotemporal lobar degeneration (FTLD) genetic susceptibility factor, and TMEM106B protein aggregates are a feature of aging and neurodegeneration. Whether TMEM106B protein levels are associated with clinical features is unknown. Objective To investigate the clinical associations of cerebrospinal fluid (CSF) TMEM106B in FTLD. Design, Setting, and Participants This cross-sectional study was conducted in 2 independent frontotemporal dementia (FTD) cohorts (recruitment from April 2009 through July 2023, with analyses from January 2025 through April 2026), with a 2-year follow up. This multicenter clinical study integrated clinical, genetic, biomarker, and neuroimaging data. Individuals were recruited through the University of California, San Francisco (n = 3733), or ALLFTD (n = 2343). Participants with available CSF were included. A discovery cohort (n = 271) included participants with sporadic neuropathology-confirmed FTLD; presymptomatic or symptomatic carriers of pathogenic variants in C9orf72 , GRN , or MAPT ; or controls. An independent validation cohort (n = 383) included participants with clinically diagnosed sporadic FTD, Alzheimer disease (AD), and controls. Exposures CSF samples for TMEM106B quantification with aptamer proteomics (SomaScan version 3.0 [discovery cohort] and SomaScan version 4.1 [validation cohort]). Main Outcomes and Measures Parametric tests compared the primary outcome, CSF TMEM106B, by disease severity, TMEM106B rs1990622 genotype, sex, clinical syndrome, pathological diagnosis, and pathogenic variant and determined associations with brain volume. Results In the discovery (n = 271; 136 women [51%]; median [IQR] age, 59 [38-80] years) and validation (n = 383; 183 women [48%]; median [IQR] age, 64 [50-78] years) cohorts, lower CSF TMEM106B was associated with more severe disease (β, −0.15; 95% CI, −0.24 to −0.04; P = .003), lower frontotemporal brain volumes (β, 0.42; 95% CI, 0.24-0.61; P &amp;lt; .001), and faster clinical progression (β, −2.21; 95% CI, −3.70 to −0.72; P = .001). Associations of TMEM106B with clinical disease severity were independent of those with neurofilament light chain. TMEM106B levels were influenced by TMEM106B rs1990622 genotype, where individuals with the protective G/G genotype had lower levels than the risk A/A genotype. CSF TMEM106B levels did not differentiate between FTLD subtypes or between FTLD and AD. Conclusions and Relevance Per the results of this cross-sectional study, TMEM106B is detectable in CSF and levels reflect disease severity in sporadic and genetic FTLD and AD, but levels are also influenced by the TMEM106B rs1990622 genotype. CSF TMEM106B could support further studies to understand the mechanisms of disease and develop clinical tools in FTLD and other neurodegenerative diseases.