Citations

Abstract Large Language Models (LLMs) demonstrate strong capabilities in healthcare applications. However, recruiting human evaluators to assess usability and effectiveness becomes challenging in the large-scale LLM-generated outcomes. Existing metrics are often outdated for evaluating LLM-based healthcare technologies. This paper proposes a Research through Evaluation (RtE) approach that can refine metrics and improve assessments of LLM-generated outputs. We deploy RtE in a retrospective comparative study of prostate cancer patient inquiries, comparing responses from human clinical care teams with those generated by a GPT-4-based in-basket bot. Through three rounds of co-evaluations with clinical professionals and an LLM-as-graders study with four LLMs, the RtE method shows that starting with small, iterative evaluations helps human graders adjust metrics to suit their daily practices better. Additionally, in the LLM-as-graders study, LLMs exhibit similar behavior, preferring the in-basket bot's responses over the clinical care team. Our methodology and findings demonstrate that the RtE method can enhance the transparency and robustness of evaluation metrics for LLM-generated content, with the potential to generalize to evaluations in other clinical practices.

Background: The rapid adoption of digital pathology and artificial intelligence (AI) in oncology creates multiple opportunities for precision diagnostics but also creates an urgent need for evidence-based standards to ensure safe and effective implementation. Goal: This article, developed by the Digital Pathology Association, presents recommendations for digital pathology as well as the validation and clinical utility of AI-enabled digital pathology tools in clinical practice. This guidance addresses analytical and clinical validation, algorithm reliability, and criteria for establishing clinical utility attributed to test use. Key recommendations emphasize separate validation of scanning processes and AI algorithms, concordance studies to support interscanner generalizability, rigorous assessment of accuracy and reliability in real-world settings, and clear description of algorithmic use limitations. These recommendations further provide frameworks for when AI may replace or augment existing diagnostic approaches, such as biomarker scoring, cancer diagnosis, and prognostic risk assessment. Conclusions: By considering payer, regulatory, and clinical perspectives, the recommendations promote transparency, trust, and reproducibility in digital pathology while encouraging value-based care delivery. We support responsible innovation in computational pathology, ensuring that AI applications achieve not only technical performance goals but also deliver measurable clinical benefit to patients.

ABSTRACT Hypomethylating agents (HMA) and allogeneic hematopoietic stem cell transplantation (alloHSCT) have both demonstrated remissions in VEXAS; however, comparative data is lacking. We conducted a multicenter, retrospective analysis of 66 patients diagnosed with VEXAS syndrome treated with HMA ( n = 35) or alloHSCT ( n = 31). Baseline characteristics such as genetics, co‐morbidities, and performance status were balanced between the groups, except older age in the HMA group. Median follow‐up from therapy initiation was 18 months (95% CI: 11–26), and 14 (21%) deaths were reported (alloHSCT n = 3; HMA n = 11). Among all evaluable patients within the alloHSCT cohort, all patients achieved molecular remission, and a substantial proportion of patients discontinued glucocorticoids (58%). In contrast, HMA therapy was associated with lower but meaningful rates of molecular remission (22%) and glucocorticoid discontinuation (6%). In a real‐world setting, HMA therapy was associated with a high discontinuation rate related to toxicity or lack of response. On multivariable analysis adjusted for age and Charlson Comorbidity Index, alloHSCT was associated with improved overall survival (HR = 0.20, 95% CI: 0.05–0.81; p = 0.024). This association remained consistent across multiple ancillary sensitivity analyses, including restriction to transplant‐eligible patients, patients aged ≤ 75 years, 1:1 matching, and propensity score‐based weighted analyses. Although limited by retrospective design, these findings suggest that alloHSCT remains an attractive and potentially curative strategy in selected patients with VEXAS. Prospective validation of these findings is warranted.

Abstract A low-cost, easy-to-adopt noninvasive imaging modality for differentiating bacterial infection from inflammation is desirable to accurately diagnose bacterial infections. Herein, a bacteria-specific PET probe, [68Ga]Ga-BP, was developed and evaluated in rat models of foreign body-associated osteomyelitis induced by methicillin-resistant Staphylococcus aureus, Pseudomonas aeruginosa, and Klebsiella pneumoniae and was directly compared with sterile inflammation and uninstrumented uninfected controls. [68Ga]Ga-BP-based imaging of MRSA-infected tibiae was also compared to [68Ga]Ga-citrate and [18F]FDG. After iron depletion, dynamic PET/CT imaging was performed at 15–120 min postinjection of radiotracers, followed by assessment of ex vivo biodistribution. [68Ga]Ga-BP successfully detected tibial infection and demonstrated >6-fold higher uptake in infected tibiae with all three bacterial species versus contralateral tibiae via PET imaging. [68Ga]Ga-BP outperformed [68Ga]Ga-citrate and [18F]FDG in identifying MRSA infection. [68Ga]Ga-BP represents a promising, noninvasive PET imaging agent for direct bacterial detection and bacterial infection treatment monitoring.

A 46-year-old female was admitted to the hospital for confusion, cognitive decline, and hallucinations with alcohol withdrawal. Her history was notable for Alcohol use disorder (AUD) and family history of Alzheimer's disease (AD). Comprehensive workup was unrevealing for metabolic encephalopathy, infection, or nutritional deficiencies. Phosphatidyl ethanol level was 1181, indicating severe alcohol misuse. Phospho-Tau 217 testing raised suspicion of AD and FDG-PET scan confirmed patterns consistent with early onset AD. AUD can precipitate earlier or more severe onset of AD. Phospho-Tau 217 can be a useful screening marker to identify people at risk of AD to prompt intensive treatment.

Pulmonary embolism (PE) is a leading cause of cardiovascular morbidity and mortality. Systemic thrombolysis is recommended for high-risk PE, but risk of major bleeding limits its use despite its efficacy. While mechanical thrombectomy (MT) has emerged as an alternative reperfusion strategy, real-world safety data remains limited. The objective of this single-center retrospective observational cohort study was to characterize the safety profile and procedural complications of MT. Methods Adult patients with intermediate-risk and high-risk PE who underwent MT at a quaternary academic medical center between January 1, 2017, and December 31, 2024, were included. Demographic, clinical, laboratory, imaging, procedural, outcome, and mortality data were collected. Major adverse events and bleeding were defined according to established criteria. Results From a total of 103 screened patients, 65 were included. Of these, 22 (34%) were classified as high-risk and 43 (66%) as intermediate-risk. Baseline demographics were similar between groups. High-risk patients demonstrated greater clinical disease severity, with higher mean Pulmonary Embolism Severity Index (PESI) scores (188.5 vs 122.4). Clinical deterioration prior to MT occurred in 35% of intermediate-risk patients. Extracorporeal membrane oxygenation (ECMO) consultation rate (32% vs 7%), peri-procedural ECMO use (18% vs 2.3%), major bleeding (27% vs 4.7%), and in-hospital mortality (36% vs 11.6%) were higher in the high-risk group. Thrombus removal was achieved in 77% and 88% of high- and intermediate-risk patients, respectively. High-risk patients had longer ICU (mean 4 vs 1.3 days) and hospital stays (mean 21.5 vs 6.4 days), higher thirty-day mortality (36% vs 14%), and lower 90-day survival (59% vs 79%). Conclusions MT achieved high rates of thrombus removal in both groups. In this retrospective cohort, high-risk patients had greater baseline clinical severity and was associated with higher bleeding rates, greater resource utilization, including ECMO consultation and peri-procedural use, prolonged ICU and hospital stays, as well as increased mortality.

8519 Background: Non-classical mutations (NCM) of epidermal growth factor receptor (EGFR) represent a broad group of oncogenic mutations in non-small cell lung cancer (NSCLC), often associated with resistance to osimertinib (e.g., P-loop and αC-helix compressing [PACC] mutations). Patients (pts) with NCM have inferior clinical outcomes with a high incidence of CNS metastases (CNSm). Silevertinib (BDTX-1535) is a fourth-generation, covalent EGFR TKI with high CNS penetrance targeting both EGFR classical mutations and NCM. The antitumor activity and safety of silevertinib in pts with advanced NSCLC were evaluated in a Phase 2 trial (NCT05256290). Methods: Pts with advanced EGFR NCM NSCLC who received no prior systemic therapy were enrolled (Cohort 3) based on a local molecular test. Silevertinib 200 mg was administered orally once daily. The primary endpoint of objective response rate (ORR) was assessed using RECIST v1.1. CNS ORR was assessed using Response Assessment in Neuro-Oncology for Brain Metastases (RANO-BM) in pts with no prior CNS therapy. Secondary endpoints include progression-free survival (PFS), duration of response (DOR), disease control rate (DCR), and safety. Results: From February 2024 to July 2025, 43 pts with 35 unique NCMs (including 25 pts with PACC mutations) were enrolled in Cohort 3: median age 70.0 years, 72% female, 74% white, and 42% never smoked cigarettes. As of November 3, 2025, median follow-up was 7.2 months. ORR by RECIST v1.1 is shown in the table. Sixteen pts (37%) had CNSm, including 7 pts (16%) with measurable CNS disease. RANO-BM CNS response was 86% (n=6/7). Complete clearance of EGFR VAF ctDNA was observed in 21/26 (81%) evaluable pts. Serious treatment-related adverse events (TRAEs) were reported in 5 pts (12%), and 4 pts (9%) discontinued treatment due to TRAEs. The most common TRAEs included rash (19%, grade 3), diarrhea (19%, grade 3), stomatitis (9%, grade 3), and paronychia (5%, grade 3). While 77% of pts had a dose reduction, 19/22 (86%) patients with a radiographic response maintained or deepened their response after dose reduction, including all patients with CNS response. Conclusions: Silevertinib demonstrated robust antitumor activity in treatment-naïve pts with a broad spectrum of NCM EGFR driver mutations with high intracranial activity in pts with brain metastases. The safety profile was consistent with the known AEs of the EGFR TKI class. Clinical trial information: NCT05256290 . Data as of November 3, 2025 N ORR n (%) 95% CI DCR n (%) 95% CI ITT Population * 43 26 (60) (44.4, 75.0) 39 (91) (77.9, 97.4) PACC mutations 25 14 (56) (34.9, 75.6) 22 (88) (68.8, 97.5) Compound mutations † 16 11 (69) (41.3, 89.0) 15 (94) (69.8, 99.8) *No pts with classical (E746_A750del or L858R) mutations were enrolled. † Presence of ≥ 2 NCMs. As of December 27, 2025, the 6-month PFS rate was 86% overall and 80% in pts with CNSm; available PFS and DOR data will be presented.