Citations

e15509 Background: Large language models (LLMs) are increasingly explored for oncology decision support, yet comparative evidence evaluating medical-specific versus general-purpose LLMs using expert-reviewed clinical cases is limited. We compared a medical-specific LLM (Google MedGemma 27B) with a general-purpose LLM (ChatGPT-5) using gastrointestinal (GI) and genitourinary (GU) oncology vignettes. Methods: This cross-sectional study evaluated 32 de-identified oncology vignettes (26 GI, 6 GU) created by a U.S. board-certified medical oncologist and informed by real-world cases. Vignettes included text-only information on the clinical context, description of pathology, tumor staging, biomarkers, and relevant comorbidities. Both LLMs were prompted to provide National Comprehensive Cancer Network(NCCN) guideline-based treatment recommendations, monitoring strategies, maintenance therapy considerations, and references, along with a self-reported confidence score (0–100). Board-certified U.S. medical oncologists independently reviewed outputs using a structured instrument; two disease specific expert oncologists evaluated each case. Outcomes included correctness and comprehensiveness (1–10 scales), hallucinations (factually fabricated or unsupported statements), NCCN traceability, potential patient harm, and clinician trust. Associations between accuracy and comprehensiveness were assessed using correlation coefficients. Inter-rater reliability was assessed using a two-way random-effects intraclass correlation coefficient (ICC). Results: ChatGPT-5 demonstrated higher mean correctness (8.65 vs 4.68) and comprehensiveness (8.68 vs 5.22) scores compared with MedGemma 27B. Accuracy and comprehensiveness were strongly correlated overall (Pearson r = 0.93, p < 0.001), with significant model-specific correlations (ChatGPT-5 r = 0.86, p < 0.001; MedGemma r = 0.77, p < 0.001). Hallucinations occurred less frequently with ChatGPT-5 (12.3% vs 65%), and NCCN-traceable recommendations were more common (75.4% vs 36.7%). Responses judged as potentially harmful occurred in 18.5% of ChatGPT-5 outputs versus 70% of MedGemma outputs. Clinician trust favored ChatGPT-5 (84.6% vs 15%). Inter-rater reliability improved with rater averaging (ICC(A,k): correctness 0.72; comprehensiveness 0.64). Conclusions: In expert-reviewed GI and GU oncology cases, ChatGPT-5 demonstrated higher clinical accuracy, comprehensiveness, and guideline alignment than a medical-specific LLM; however, a substantial rate of hallucinations and potentially harmful recommendations persisted across models. These findings underscore that, despite performance differences, current LLMs require rigorous clinician oversight and should not be used as standalone decision-support tools in oncology.

e15501 Background: CP is the backbone of first-line therapy for recurrent or metastatic ASCC. In practice, Q3-week CP has been explored as a potential alternative to the standard (d1,8,15 Q28-day cycle) dosing, though its efficacy is not known. This study evaluated CP scheduling and clinical survival associations in metastatic ASCC. Methods: This multicenter retrospective cohort study included patients with metastatic ASCC treated with first-line CP, identified using the Mayo Data Explorer database. Electronic health records were reviewed to collect baseline clinical characteristics, treatment regimens, responses, progression-free survival (PFS), and overall survival (OS). Descriptive statistics were used to summarize patient demographics. Survival was estimated using Kaplan–Meier methods. Multivariable Cox proportional hazards models were used to compare survival by dosing schedule, adjusting for age, ECOG performance status, HPV (p16) status, liver metastasis, and prior chemoradiation. Results: 54 patients treated between 2020-2025 with first-line CP for metastatic ASCC were included. Baseline patient demographics were balanced between dosing groups (Table 1). Median follow-up time was 26.2 months. For Q3-week CP (n = 21), median PFS and OS were 9.8 (95% CI 8.1, NR) and 27.0 (95% CI 15.7, NR) months, respectively, while median PFS was 6.5 (95% CI 4.5,15.3) months with standard CP (n = 33) and median OS was not reached (95% CI 11.2, NR). Response rate was 76% with Q3-week and 60.6% with standard CP (p = 0.23) dosing. Survival outcomes were similar when comparing Q3-week CP relative to standard CP dosing using multivariable analysis (PFS HR 0.68, 95% CI 0.33–1.41; OS HR 1.67, 95% CI 0.67–4.14). Higher ECOG and liver metastasis were independently associated with inferior outcomes, while HPV was associated with improved survival. Subsequent therapy patterns were similar between CP schedule groups. Ten patients (16.9%) received local therapy, and 24 of 28 patients with progression received further systemic therapy, including 21 (87.5%) treated with immunotherapy. Conclusions: Survival outcomes were similar across dosing schedules, suggesting Q3-week CP may be a reasonable alternative to standard CP, with potential to reduce treatment-related burden in metastatic ASCC. Future analysis will compare toxicity profiles between CP dosing schedules, including in combination with PD1 inhibitors given the results of the POD1UM-303 trial. Patient demographics (n=54). Variable Q3-week CP (n=21) Standard CP (n=33) p-value Sex Female 16 (76.2%) 24 (72.7%) 0.77 Male 5 (23.8%) 9 (27.3%) Age Median (Q1, Q3) 62 (50.5, 69) 64 (55.5, 70) 0.57 BMI Median (Q1, Q3) 27.1 (21.9, 34.4) 27.4 (22.5, 31.7) 0.72 ECOG 0 9 (42.9%) 11 (34.4%) 0.53 ≥ 1 12 (57.1%) 21 (65.6%) Confirmed HPV (p16) Positive 16 (76.2%) 24 (72.7%) 0.77 Prior chemoradiation Yes 17 (81.0%) 21 (63.6%) 0.17

3510 Background: Patients with RAS-mutated mCRC have limited treatment options. Chemotherapy in combination with bevacizumab (Bev) has been the standard first-line treatment for the past two decades, yet the prognosis for these patients remains poor. Onvansertib (Onv), a selective PLK1 inhibitor, has previously demonstrated synergy with chemotherapy in the second line setting. The Phase 2 randomized CRDF-004 trial evaluated onvansertib plus standard-of-care (SoC; FOLFIRI or FOLFOX with Bev) in as first-line RAS-mutated mCRC. Methods: Patients with first-line KRAS or NRAS mutant mCRC were randomized to SoC alone or SoC plus onvansertib (20 mg or 30 mg). The primary endpoint was objective response rate (ORR) as determined by a blinded-independent central review (BICR) using RECIST v1.1. Key secondary endpoints included progression-free survival (PFS), duration of response, and safety. PFS was assessed by BICR and investigator assessments combined. Results are presented in the intent-to-treat population (ITT). Results: At the data cutoff of 22 Jan 2026,110 patients were randomized with a median follow up of approximately 10 months. Baseline characteristics were relatively balanced across arms. Onv 30 mg + FOLFIRI + Bev demonstrated numerically higher ORR (n = 13/18, 72.2%, 95% CI: 46.5, 90.3) compared with the FOLFIRI + Bev control arm (n = 8/19, 42.1%, 95% CI: 20.3, 66.5), the FOLFOX + Bev control arm (n = 8/18, 44.4%, 95% CI: 21.5, 69.2) and the combined control arms (SoC: n = 16/37, 43.2%, 95% CI: 27.1-60.5). Onv 30mg + FOLFIRI + Bev demonstrated a p-value of 0.051 for confirmed ORR vs. SoC; and a p-value of 0.045 for the 6-month ORR compared to FOLFIRI + Bev (n = 10/18, 55.6%, 95% CI: 30.8-78.5 vs. n = 4/19, 21.1%, 95% CI: 6.1-45.6, respectively). The median PFS for the control arms was 10.97 months (95% CI: 7.52, NR) for FOLFIRI + Bev, 9.89 months (95% CI: 9.43, NR) for FOLFOX + Bev and 10.97 months (95% CI: 9.43-15.44) for the combined SoC arms. The median PFS in the Onv 30mg+ FOLFIRI + Bev arm was not reached (95% CI: 9.72-NR). The PFS HR of 30mg + Onv +FOLFIRI + Bev vs FOLFIRI + Bev was 0.38 (95% CI: 0.12-1.17) and vs SoC was 0.37 (95% CI: 0.13-1.02, p = 0.048). The clinical benefit from Onv treatment was further evidenced by the 12-month PFS rates, 61.9% (95% CI: 40.1, 95.8) in the Onv 30 mg + FOLFIRI + Bev arm vs. 28.4% (95% CI: 9.3, 86.9) in the FOLFIRI + Bev arm, and 30.1% (95% CI: 13.8-65.7) in the SoC. Onvansertib was well tolerated, with no unexpected toxicities observed; the most common grade ≥3 adverse event was neutropenia. PK data support moving forward with the onvansertib 30mg dose with FOLFIRI + Bev. Conclusions: Onvansertib plus FOLFIRI and bevacizumab demonstrated improved antitumor activity and manageable safety in first-line RAS-mutated mCRC. These results support further clinical and support planned confirmatory Phase 3 evaluation. Clinical trial information: NCT06106308 .

As an emerging tire wear-derived environmental contaminant, 6PPD-quinone (6PPDQ) has raised significant concerns regarding its neurotoxic potential, particularly for children exposed to recycled tire crumb rubber in playgrounds. However, the molecular mechanisms by which 6PPDQ influences neurological disorders such as epilepsy remain poorly understood. In this study, we employed an integrative framework combining network toxicology, bulk analysis of human epileptic brain tissues, Mendelian randomization, and molecular dynamics simulations to elucidate these mechanisms. Our findings, validated through CETSA-WB and SPR, identify 6PPDQ as a direct ligand that binds to and stabilizes neuronal TP53. Through a synergistic double-hit mechanism, 6PPDQ directly engages the TP53 pathway while simultaneously triggering microglial interleukin-6 secretion. These converging pathways lead to the suppression of the master antioxidant regulator Nrf2, resulting in glutathione depletion, excessive reactive oxygen species accumulation, and exacerbated neuronal damage under excitotoxic stress. Experimental validation using glutamate-induced HT22 cell models and microglia–neuron crosstalk systems confirmed that targeting the TP53/Nrf2 axis or scavenging ROS significantly attenuates 6PPDQ-induced neurotoxicity. Our findings highlight critical risks to pediatric neurological health posed by tire-derived contaminants and identify the TP53/Nrf2 axis as a promising therapeutic target. Furthermore, this work provides a robust scientific basis for refining risk assessment frameworks and developing regulatory strategies to mitigate environmental exposure to 6PPDQ.

Abstract Rationale Propofol-based sedation in the intensive care unit (ICU) is associated with clinically significant hemodynamic instability, including hypotension. Previous studies have identified hypotension as a key predictor of adverse outcomes in critically ill patients. However, no validated tools currently exist to systematically predict this risk. Therefore, this study aimed to develop and externally validate machine-learning (ML) models for predicting the risk of hypotension in ICU patients receiving propofol sedation. Methods We developed ML models using retrospective data from adult, mechanically ventilated ICU patients who received propofol infusion for at least 24 hours across four Mayo Clinic sites between 2018 and 2024. The primary outcome was hypotension, defined as ≥ 2 consecutive MAP readings ≤60 mmHg within 15 minutes during the first 6 hours of propofol sedation. The MIMIC-IV dataset was used for external validation. Features were selected using a union of Boruta and LASSO-penalised logistic regression. Four algorithms (elastic net-penalised logistic regression, random forest, LightGBM, and multilayer perceptron) were tuned using Bayesian optimization to minimize cross-entropy loss during leave-one-site-out cross-validation (LOSO-CV). The algorithm with the lowest cross-entropy loss during LOSO-CV was chosen as the best-performing algorithm. Classification thresholds were selected to target sensitivity ≥80%. Model performance was evaluated using area under the receiver operating characteristic curve (AUC-ROC), calibration slope and intercept, sensitivity, and specificity. Feature importance was evaluated using SHapley Additive exPlanations (SHAP). Results The development cohort included 16,361 patients from Mayo Clinic, of whom 3,482 (21.3%) developed hypotension. The external validation cohort comprised 4,684 patients from MIMIC-IV, with 115 (2.5%) experiencing hypotension. Patient characteristics included age, admission Sequential Organ Failure Assessment (SOFA) scores, starting propofol dose, ICU admission source, and comorbidities. The best-performing algorithm was LightGBM. The average AUC-ROC during LOSO-CV was 0.70 (95% confidence interval [CI] 0.68-0.71), versus 0.64 during external validation. The average calibration slope and intercept during LOSO-CV was 1.08 (95% CI 0.89-1.27) and 0.08 (95% CI -0.25-0.41), respectively, versus 0.78 and -2.07 during external validation. The average sensitivity and specificity during LOSO-CV was 0.81 (95% CI 0.79-0.83) and 0.47 (95% CI 0.42-0.52), respectively, versus 0.54 and 0.65 during external validation. Conclusions We developed and externally validated an ML-based classifier that demonstrates moderate discrimination for predicting propofol-induced hypotension in mechanically ventilated ICU patients. This tool may facilitate early identification of high-risk patients and enable proactive hemodynamic management strategies. Further prospective validation and assessment of clinical impact are warranted. This abstract is funded by: None

Abstract Rationale Propofol sedation in mechanically ventilated patients often causes clinically significant hypotension. While we concurrently developed and validated a machine-learning (ML) classifier to predict hypotension risk, accurately estimating the timing of hemodynamic deterioration could enable more precise, proactive interventions. This study aimed to develop and externally validate ML models that predict not only whether hypotension will occur, but also when it is likely to develop after propofol initiation. Methods ML models were developed using retrospective data from adult, mechanically ventilated ICU patients who received propofol infusion for ≥24 hours across four Mayo Clinic sites from 2018 to 2024 and developed hypotension (≥2 consecutive MAP readings ≤60 mmHg within 15 minutes) during the first 6 hours of sedation. The MIMIC-IV dataset was used for external validation. Features were selected using a union of Boruta and LASSO-penalised quantile regression. Two quantile algorithms (LASSO-penalised quantile regression and LightGBM) were trialed to predict the conditional median and interquartile range (IQR) of time to hypotension. In addition, two probabilistic algorithms (NGBoost and Mixture Density Network) were trialed to predict the full conditional distribution. The quantile models were tuned via Bayesian optimisation to minimise pinball loss during leave-one-site-out cross-validation (LOSO-CV), whereas probabilistic models were tuned to minimise the Continuous Ranked Probability Score. The algorithm with the lowest average pinball loss was deemed the best-performing algorithm. Model performance was evaluated using median absolute error (MAE), IQR coverage, and IQR width. Feature importance was evaluated using SHapley Additive exPlanations (SHAP). Results We included 3482 patients from the Mayo Clinic for model development, with 115 patients from the MIMIC-IV dataset used for external validation. Patient characteristics in the modeling dataset included age, sex, admission SOFA scores, weight, ICU admission source, and initial propofol dose. The best-performing algorithm was LightGBM-based quantile regression. The average MAE during LOSO-CV was 4.11 hours (95% confidence interval [CI] 3.98-4.23), versus 4.83 hours during external validation. The average IQR coverage and width during LOSO-CV was 0.49 (95% CI 0.48-0.49) and 7.12 (6.89-7.35), respectively, versus 0.07 and 6.51 during external validation. Conclusions We developed and externally validated a quantile regression model that predicts both the timing and uncertainty of propofol-induced hypotension in mechanically ventilated ICU patients. This model provides clinicians with actionable time estimates that may facilitate preemptive hemodynamic interventions and optimize sedation management. Prospective validation, expanded predictors and clinical impact assessment are warranted. This abstract is funded by: None

Abstract Minimal residual disease (MRD) has been endorsed by the FDA Oncology Drugs Advisory Committee as an endpoint for accelerated approval in multiple myeloma based on individual patient data collected from randomized trials. However, emerging data from recent trials were not included. A novel artificial intelligence (AI)–assisted framework is proposed, which automates information identification and extraction, providing up-to-date analyses that confirm moderate trial-level and strong individual patient–level associations between MRD− rates at suspected complete response (MRD-CR) and survival endpoints in multiple myeloma. Specifically, this study utilized an AI-assisted framework that identifies relevant studies and filters critical information to analyze published data via 2 independent objectives. First, we examined the trial-level association using the coefficients of determination (R2) and its 95% confidence intervals (CI) based on published statistics of treatment effects on MRD and various endpoints. Next, we generated synthetic individual patient data with covariates through AI-curated tools to estimate the individual-level association. The AI tool searched for eligible randomized clinical trials (RCT). A total of 20 two-arm comparisons from 19 RCTs were analyzed. Trial-level analysis showed an R2 of 0.71 (95% CI, 0.52–0.89) pooling disease subpopulations. Furthermore, AI techniques were applied to create synthetic individual data, combining information extracted from Kaplan–Meier curves and subgroup analyses from published literatures. Using generated synthetic data, we estimated the individual-level correlation between MRD-CR rates and progression-free survival outcomes with a bivariate copula model and calculated a global odds ratio of 7.28 (95% CI, 5.60–8.95). Significance: (i) A novel AI-assisted framework is proposed, which automates information identification and extraction, providing rapid up-to-date analyses. (ii) Moderate trial-level and strong individual patient–level associations between MRD and various clinical endpoints in multiple myeloma are confirmed.

INTRODUCTION: Nosocomial infections (NIs) in cirrhosis are associated with high mortality but could be preventable. Logistic regression (LR) models have failed to identify high-risk patients. We aimed to develop machine learning (ML) models to predict NI. METHODS: The CLEARED consortium consists of prospectively enrolled cirrhosis inpatients from >120 centers. Using day-of-admission clinical data, 3 ML approaches (random forest [RF], extreme gradient boosting, and neural networks [NNs]) were used to predict NI. Data were split 80:20 for training and testing stratified by the outcome. Models were compared using area under the receiver operating characteristic curve (AUC). RESULTS: In total, 8,263 patients (55.90 ± 13.34 years; 64.1% men) from 127 centers in 37 countries were included. NI developed in 869 (10.5%), a median of 6 (4–11) days of postadmission. Major NIs were respiratory (29.6%) and urinary tract infection (15.7%), and spontaneous bacterial peritonitis (13.5%). NIs occurred more frequently in patients from low/low-middle income countries and those with severe liver disease, alcohol etiology, and admission infections. NIs were associated with inpatient mortality (31.9% vs 8.1%, P < 0.001) and liver transplantation (4.8% vs 1.9%, P < 0.001). Although the RF model (AUC 0.69) showed good calibration (Brier score 0.09), outperforming extreme gradient boosting, neural network, and LR models (AUC 0.66 for all; LR comparison P = 0.043), no model achieved AUC ≥0.80 for clinical utility. At 10% predicted probability threshold, the RF model demonstrated only 75.4% sensitivity, 52.9% specificity, and 15.9% positive predictive value (PPV). DISCUSSION: NIs cannot be accurately predicted from day-of-admission data using ML models, even in a large, prospective, global cirrhosis cohort. Every hospitalized patient with cirrhosis should receive protocolized infection control measures.

PURPOSE 177Lu-PSMA-617 (LuPSMA) is an approved prostate-specific membrane antigen (PSMA)–targeted radiopharmaceutical therapy for men with metastatic castration-resistant prostate cancer (mCRPC). Older patients, particularly octogenarians, represent a substantial proportion of men with mCRPC but have been traditionally underrepresented in key registrational trials of LuPSMA. METHODS This retrospective multi-institutional study evaluated patients age 80 years and older with mCRPC after chemotherapy treated with ≥1 cycle of LuPSMA from August 2022 to December 2024. Clinical and demographic data were abstracted from electronic medical records. Outcomes included prostate specific antigen (PSA) response (PSA50 and PSA90), progression-free survival (PFS), overall survival (OS), and toxicities. Kaplan-Meier methods estimated PFS/OS; descriptive statistics summarized baseline and safety data. RESULTS Ninety-five patients (median age, 83 years) were included, of whom 36 (38%) and 21 (22%) had a history of cardiac disease or chronic kidney disease, respectively. Median follow-up was 24.1 months (IQR, 6.7-28.1). Patients received a median of five cycles of LuPSMA. Fifty-two (57%) and 21 (23%) patients achieved a PSA50 and PSA90, respectively. Median PFS and OS were 7.3 months (95% CI, 6.4 to 8.7) and 13.5 months (95% CI, 9.7 to 18.8), respectively. Grade ≥3 hematologic toxicities included anemia (n = 19; 20%) and thrombocytopenia (n = 4, 4%); grade ≥3 acute kidney injury occurred in one patient. Fourteen (15%) patients had dose delays, seven (7%) required reductions, and 10 (11%) discontinued therapy due to toxicity. Thirty-seven patients (39%) were hospitalized during therapy, with intensive care unit-level care required in two patients (2%), and there was one treatment-related death. CONCLUSION LuPSMA had comparable outcomes in octogenarians with mCRPC to patients on registrational trials, although 2/5 of patients required hospitalization during therapy. These findings support the feasibility and efficacy of LuPSMA in well-selected older men with mCRPC and suggest a role for closer monitoring of older patients.

Abstract Pancreatic cancer disproportionately affects Black individuals in the United States, but they have limited representation in genetic studies of pancreatic ductal adenocarcinoma (PDAC). To address this gap, we performed admixture mapping and genome-wide association analysis (GWAS) in genetically inferred African ancestry individuals (1,030 cases and 889 controls). Admixture mapping identified three regions with a significantly higher proportion of African ancestry in cases compared to controls (5q33.3, 10p1, 22q12.3). GWAS identified a genome-wide significant association at 5p15.33 ( CLPTM1L , rs383009:T>C, T Allele Frequency=0.51, OR:1.45, P value=1.24×10 -8 ), a locus previously associated with PDAC. Known loci at 5p15.33, 7q32.3, 8q24.21 and 7q25.1 also replicated (P value <0.01). Multi-ancestral fine-mapping identified two potential causal SNPs (rs3830069 and rs2735940) at 5p15.33. Collectively these findings identified novel PDAC risk loci and expanded our understanding of this deadly cancer in underrepresented populations, emphasizing the multifactorial nature of PDAC risk including inherited genetic and non-genetic factors. Statement of Significance To understand how genetic variation contributes to PDAC risk in Black people in North American, we studied individuals of genetically-inferred African ancestry. We identified novel risk loci and differences in the contribution of known loci. This demonstrates that ancestry-informed genetic analyses improve our understanding of PDAC risk and enhances discovery.