Citations

e16018 Background: Intrahepatic cholangiocarcinoma (iCCA) exhibits marked heterogeneity in outcomes that is incompletely explained by stage or treatment timing. Although treatment delay is frequently cited as a modifiable risk factor, its prognostic significance in real-world practice remains unclear. We applied machine learning (ML) to model iCCA as a dynamic disease process and to evaluate whether longitudinal disease trajectories better explain outcomes than treatment timing alone. Methods: We conducted a population-based retrospective analysis of adults diagnosed with iCCA using the Surveillance, Epidemiology, and End Results (SEER) database (2000–2022). Time from diagnosis to first cancer-directed therapy was categorized as ≤30 days, 31–60 days, or > 60 days. Associations between treatment delay, disease characteristics, and overall survival (OS) were assessed using multivariable logistic and Cox regression models. To capture real-world disease evolution, each patient’s clinical course was encoded as an ordered longitudinal sequence of diagnosis, treatment states (surgery, chemotherapy, radiation), and outcome. Unsupervised ML was used to identify latent disease-course trajectories. A supervised ML model using variables available at diagnosis was developed to predict membership in high-risk trajectories. Results: The cohort included 18,312 patients with iCCA. Treatment initiation beyond 30 days occurred in more than half of patients and was more common in localized than advanced disease. In adjusted survival analyses, shorter time to treatment was paradoxically associated with worse OS, consistent with confounding by disease severity rather than a protective effect of delay. Unsupervised ML identified four distinct and reproducible disease trajectories with markedly different outcomes. Median OS ranged from approximately 2 months in early-failure trajectories to approximately 15 months in favorable multimodality trajectories. Trajectory membership was not fully explained by stage or treatment timing. A supervised ML model predicted high-risk trajectory membership with good discrimination (AUC 0.82). Key predictors included tumor grade, age, tumor burden, and incomplete staging, whereas treatment timing had minimal influence. Conclusions: In iCCA, treatment delay reflects underlying disease biology and care prioritization rather than independently determining survival. ML-based trajectory modeling reveals clinically meaningful disease courses that better explain outcome heterogeneity than treatment timing or stage alone and enables early identification of patients at highest risk for unfavorable real-world outcomes.

e16017 Background: Pancreatic ductal adenocarcinoma (PDAC) is the most common form of pancreatic cancer and remains one of the deadliest malignancies worldwide, largely due to late-stage diagnosis. Early-stage PDAC is frequently asymptomatic, limiting opportunities for timely detection and intervention. We aimed to develop prognostic models using routinely acquired pre-diagnostic clinical data to identify individuals at high risk for PDAC prior to clinical manifestation. Methods: We conducted a retrospective study of patients treated at Mayo Clinic sites in Rochester, MN; Phoenix, AZ; and Jacksonville, FL between January 2011 and February 2023. The cohort included 1,151 patients with PDAC and 3,745 noncancer controls. CT scans were standardized to axial abdominal acquisitions with slice thickness between 0.8–3.0 mm. Radiology reports were de-identified, normalized, and parsed into structured sections (Indications, Findings, Impression), with an additional pancreas-specific section aggregating pancreas-related sentences. Demographic and clinical variables, e.g., age, sex, race, etc., were encoded using a templated sentence representation. All textual inputs were embedded using a Sentence-BERT model. A multimodal deep learning survival model was developed to integrate CT imaging, radiology report embeddings, and demographic features. The model was trained using a negative log-likelihood loss to estimate patient-specific PDAC risk scores. Model performance was assessed using the concordance index (C-index). Internal validation included patients from MN and AZ sites, while external validation used the FL cohort. Results: We evaluated unimodal (imaging, text, and demographics) and multimodal inputs. Fusion of imaging, radiology report, and demographic features achieved the highest performance, with internal and external validation C-indices of 0.885 (95% CI, 0.856–0.908) and 0.751 (95% CI, 0.719–0.783), respectively. Subgroup analysis by tumor location demonstrated differential performance. For pancreatic head tumors, internal and external C-indices were 0.916 (95% CI, 0.841–0.960) and 0.778 (95% CI, 0.730–0.823). For pancreatic body tumors, C-indices were 0.954 (95% CI, 0.897–0.997) internally and 0.929 (95% CI, 0.897–0.953) externally. For pancreatic tail tumors, internal and external C-indices were 0.903 (95% CI, 0.830–0.974) and 0.789 (95% CI, 0.689–0.874). Conclusions: A multimodal deep survival model leveraging pre-diagnostic CT imaging, radiology report embeddings, and demographic data can effectively identify individuals at high risk for PDAC prior to clinical diagnosis. Differences in performance were observed across pancreatic head, body, and tail tumors. This approach highlights the potential of opportunistic cancer risk stratification from routine imaging to enable earlier detection and targeted intervention strategies.

e15722 Background: Despite overall improvements in colorectal cancer (CRC) outcomes, concerning trends have emerged in younger populations. We examined two decades of US mortality data to characterize age-specific trajectory differences and identify inflection points in the epidemiologic landscape of CRC. Methods: Using CDC WONDER data (1999–2020), we analyzed age-adjusted CRC mortality rates (ICD-10: C18–C20) stratified by age: early-onset (≤44 years) versus late-onset (≥45 years). Joinpoint regression analysis (Version 5.4.0.0) with permutation testing identified significant trend changes. Annual Percent Change (APC) estimates with 95% confidence intervals assessed temporal patterns, with statistical significance set at α = 0.05. Results: Late-onset CRC (≥45 years): Mortality rates declined substantially from 69.26 to 40.49 per 100,000, representing a 41.6% reduction. A critical inflection point occurred in 2014 (95% CI: 2011–2015), marking a deceleration in the rate of improvement. The pre-2014 period showed robust decline (APC: −3.06%; 95% CI: −3.18 to −2.93; p < 0.000001), which slowed significantly post-2014 (APC: −1.61%; 95% CI: −2.12 to −1.09; p = 0.000006)—a 47% reduction in the pace of mortality improvement.Early-onset CRC (≤44 years): In stark contrast, mortality rates increased from 0.92 to 1.13 per 100,000, representing a 22.8% rise. No inflection points were identified; instead, a persistent upward trajectory characterized the entire 22-year period (APC: +0.78%; 95% CI: +0.54 to +1.02; p = 0.000001). This yielded a continuous, unabated increase in deaths among young adults. Conclusions: This analysis reveals a troubling paradox in US colorectal cancer mortality: while older adults experienced substantial but decelerating improvements, young adults faced steadily worsening outcomes throughout the study period. The 2014 inflection point in older adults, coinciding with slowed screening expansion and potential changes in risk factor prevalence, coupled with the relentless rise in early-onset mortality, signals an urgent need to reassess prevention strategies, lower screening age thresholds, and investigate biological and environmental drivers of early-onset disease. The nearly 23% increase in mortality among those ≤44 years represents a growing public health crisis demanding immediate action.

e15713 Background: Fruquintinib demonstrated overall survival benefit in heavily pretreated metastatic colorectal cancer (mCRC) in FRESCO and FRESCO-2. The FDA approved fruquintinib on Nov 8, 2023 for adult patients with mCRC after standard chemotherapy plus anti-VEGF therapy, with anti-EGFR therapy when appropriate (e.g., RAS wild-type). Observed geographic variation in post-approval uptake suggests potential disparities in the delivery of evidence-based medicine across U.S. regions. Methods: We conducted a retrospective cohort study using Epic Cosmos, a database of over 300 million patient records from over 1,500 hospitals nationwide. We identified adults with metastatic colorectal cancer (mCRC) with prior standard systemic therapy exposure and recorded fruquintinib exposure following FDA approval beginning Nov 8, 2023. Fruquintinib uptake was analyzed by U.S. Census region and state, expressed as the proportion of eligible mCRC patients receiving fruquintinib within each geography. Results: Among 116,116 pre-treated mCRC patients, 6,722 (5.79%) had recorded fruquintinib exposure during the post-approval period. Uptake varied nearly five-fold across states (2.27%–11.33%) (Table). High-uptake states were disproportionately concentrated in the central U.S./Midwest, whereas multiple low-uptake states were observed in the West and Northeast, suggesting broader regional variation beyond state-to-state differences. Conclusions: In a large national EHR-derived cohort of pre-treated mCRC patients after U.S. approval, fruquintinib uptake demonstrated substantial geographic variation. These findings raise concern for inequities in delivery of evidence-based later-line therapy and highlight the need to identify modifiable system- and site-level drivers (e.g., access, referral patterns, treatment infrastructure) to promote equitable adoption. Fruquintinib uptake among eligible metastatic colorectal cancer patients by U.S. state. Rank Geography Eligible mCRC (N) Fruquintinib uptake (%) Overall Overall 116,116 5.79 Top 1 Oklahoma 1,236 11.33 2 Kansas 1,312 10.75 3 Indiana 2,448 10.70 4 Missouri 1,521 9.47 5 Connecticut 2,208 9.24 Bottom 1 Washington 1,232 2.27 2 Nevada 441 2.72 3 Rhode Island 476 2.73 4 Tennessee 1490 2.82 5 Maine 614 2.93 Fruquintinib uptake is reported as the proportion of eligible pre-treated metastatic colorectal cancer (mCRC) patients with recorded fruquintinib exposure following U.S. FDA approval. States are ranked by uptake percentage.

e15708 Background: AFTT is characterized by physical, nutritional, and psychosocial decline. In GIC, AFTT arises from malnutrition, disease progression, or treatment-related toxicity resulting in poor quality of life. We evaluated the association between AFTT and IM, length of stay (LOS), and hospital charges (HC) in hospitalized GIC patients. Methods: The National Inpatient Sample (2016-21) was used to identify adult hospitalizations with GIC using ICD-10 codes, which were stratified as primary or secondary diagnoses by AFTT. Clinical characteristics and primary outcomes including IM, LOS, and HC were extracted. Multivariable regression models adjusting for demographics, insurance, comorbidity burden, census tract income, and hospital teaching status were evaluated. Results: Among 2,732,505 GIC hospitalizations, colorectal (46.9%) and pancreatic (20.4%) cancer predominated; 121,035 (4.4%) had AFTT. Patients with AFTT vs non-AFTT, were older (mean = 69.3 vs 66.5 yrs), more often male (58.1% vs 57%), and from the lowest income quartile (27.7% vs 26%). AFTT cohort had higher comorbidity burden (mean CCI = 7.24 vs 6.26), more Black patients (19.1% vs 12.8%), higher Medicare use (62.7% vs 56.0%), increased palliative care consultation (37.4% vs 13.6%), risk of malnutrition, AKI, and dementia (all p< 0.001). AFTT was associated with higher IM, longer LOS, and lower HC. Higher age, males, and self-pay were associated with increased IM and HC, but shorter LOS. Lower CCI was associated with reduced IM, LOS, HC, whereas urban hospitals showed lower IM but higher LOS and HC. Black patients had higher IM and LOS but lower HC, contrary to Hispanic patients. Conclusions: The grave impact of AFTT and variability in survival and practice patterns is driven multifactorially, raising concerns about inequitable care access and differences in healthcare utilization. Standardized guidelines and multidisciplinary interventions are urgently needed to combat AFTT in GIC. Predictor IM aOR (95% CI) LOS Adj. Ratio (95% CI) HC Adj. Ratio (95% CI) AFTT Status (vs. Non-AFTT) 1.96 (1.93-2.00) 1.13 (1.12-1.14) 0.86 (0.85-0.86) Age (/year increase) 1.02 (1.02-1.02) 1.00 (1.00-1.00) 1.00 (1.00-1.00) Sex (vs. Female) Male 1.18 (1.17-1.19) 0.99 (0.98-0.99) 1.05 (1.05-1.06) CCI Score (vs. High, ≥5) Low (≤2) Mid (3-4) 0.33 (0.32-0.33) 0.48 (0.47 - 0.48) 0.87 (0.87 - 0.88) 0.93 (0.93 - 0.94) 0.93 (0.93 - 0.94) 1.02 (1.02 - 1.02) Insurance (vs. Medicare) Private Medicaid Self-Pay 1.03 (1.01 - 1.05) Ref in model 1.29 (1.24 - 1.33) 0.92 (0.91 - 0.92) 0.96 (0.95 - 0.97) 0.98 (0.97 - 1.00) 1.03 (1.02 - 1.03) 1.06 (1.05 - 1.06) 0.96 (0.95 - 0.97) Race (vs. White) Black 1.03 (1.00 - 1.06) 1.07 (1.06 - 1.08) 0.95 (0.94 - 0.96) Facility Status (vs. Rural) Urban Non-Teaching Urban Teaching 0.87 (0.85 - 0.89) 0.78 (0.77 - 0.80) 1.10 (1.09 - 1.11) 1.18 (1.17 - 1.19) 1.64 (1.63 - 1.65) 1.86 (1.84 - 1.87)

e15687 Background: Colorectal cancer (CRC) screening has been shown to reduce CRC-related mortality, yet substantial racial and ethnic disparities persist. Whether CRC screening confers equivalent mortality benefit across racial and ethnic groups, and the mechanisms underlying potential heterogeneity if any, remains unelucidated. Methods: We conducted a secondary analysis of the Prostate, Lung, Colorectal and Ovarian (PLCO) Cancer Screening Trial, including 154,852 participants aged 55–74 years randomized to CRC screening or usual care. After excluding those with incomplete diagnostic data, 2,123 were eligible for analyses (929 in the screening arm, 1,195 in the control arm). Primary outcomes included cumulative incidence and CRC-specific mortality. Age-adjusted CRC mortality rates (AR, per 100,000) and crude rates (CR) were estimated by race/ethnicity and trial arm. Effect heterogeneity was assessed descriptively with trial-center analyses given correlations between center and racial/ethnic composition. Results: Overall, CRC screening was associated with lower cumulative CRC incidence and CRC-specific mortality compared with usual care. However, the magnitude of mortality reduction varied substantially by race/ethnicity (Table). Trial-center analyses demonstrated wide heterogeneity and screening–control separation across. For instance, data from Pennsylvania and Hawaii centers which has a high proportion of either non-Hispanic Black participants or Native Hawaiian/Pacific Islanders, had no mortality separation after 20 years of follow-up. Conclusions: Within a large randomized trial data, CRC screening effectiveness on mortality was not uniform across racial and ethnic groups. To enhance CRC screening benefits and to address the racial differences, downstream care processes have to be optimized rather than focusing on screening exposure alone. Future prospective data is warranted to address inequities in the full screening-to-diagnosis-to-treatment continuum and find out in which stage of care a care delivery disruption occurs. Colorectal cancer mortality by race/ethnicity and screening arm in the PLCO trial. Race/Ethnicity Age-standardized Mortality Rate per 100,000 (95% CI) Risk Reduction (95% CI) Screening Control Absolute Relative (%) White 362.1(313.8–410.3) 538.5(479.4–597.6) 176.4(100.1–252.7) 32.8(20.1–43.4) Black 680.8(405.8–955.7) 808.6(499.1–1118.2) 127.9(-286.2–541.9) 15.8(-46.9–51.7) Hispanic 455.8(86.7–824.8) 654.3(198.2–1110.4) 198.5(-388.2–785.3) 30.3(-102.8–76.1) Asian 333.0(84.7–581.2) 588.2(252.3–924.0) 255.2(-162.4–672.9) 43.4(-44.8–77.9)

e15647 Background: Colorectal cancer (CRC) screening offers an opportunity to prevent invasive disease by detecting and removing precancerous lesions. Colonoscopy-based detection of adenomas and serrated polyps is strongly associated with reduced risk of interval CRC. While the performance of noninvasive CRC screening tests is well described, comparative evidence on adenoma detection across screening colonoscopy and stool-based strategies with follow-up colonoscopy remains limited. We conducted a systematic review to compare adenoma detection across CRC screening modalities. Methods: MEDLINE (PubMed) and Embase were searched from January 1, 2000, through July 10, 2025, for peer-reviewed, U.S.-based studies reporting colonoscopy among average-risk adults aged ≥45 years, who underwent screening colonoscopy or follow-up colonoscopy after a positive stool-based screening test [fecal immunochemical test (FIT) or multi-target stool DNA (mt-sDNA) test]. The primary outcome was the detection of one or more adenomas at colonoscopy. High-risk or surveillance cohorts, AI-assisted colonoscopy, diagnostic accuracy trials, and non-U.S. studies were excluded. Results: Among 7,710 de-duplicated records, 147 studies underwent full-text review, and 47 studies were included in quantitative synthesis, with reported adenoma detection for screening colonoscopy alone (n = 31), follow-up colonoscopy after a positive FIT alone (n = 4), follow-up colonoscopy after a positive mt-sDNA test alone (n = 4), and for more than one of the screening strategies of interest (n = 8). Across included studies, overall adenoma detection differed by screening pathway. Follow-up colonoscopy after a positive mt-sDNA test demonstrated the highest adenoma detection (66.3%; 95% CI, 61.2-71.1), with lower adenoma detection observed for colonoscopy after a positive FIT (50.6%; 95% CI, 43.4-57.8) and screening colonoscopy alone (33.8%; 95% CI, 30.9-36.8) (Table 1). Conclusions: Detection of adenomas was highest following colonoscopy performed after a positive mt-sDNA test compared with FIT-positive follow-up or screening colonoscopy alone. These findings highlight differences in downstream detection across CRC screening tests and suggest that, in settings where colonoscopy resources are constrained, initial mt-sDNA screening with timely follow-up colonoscopy may help prioritize patients most likely to benefit. CRC screening modality No. of Studies Adenoma detected Study population Proportion, % (95% CI) Heterogeneity, I² (%) Between-study variance (τ 2 ) P-value* mt-sDNA 8 17,436 29,098 66.3 (61.2-71.1) 94.6 0.0924 <0.0001 FIT 11 17,588 32,630 50.6 (43.4-57.8) 94.0 0.2187 <0.0001 Screening colonoscopy 38 2,351,801 6,378,891 33.8 (30.9-36.8) 99.8 0.1710 NA *P-value represents adenoma detection vs. screening colonoscopy adenoma detection.

e15598 Background: Immune checkpoint inhibitors (ICIs) have transformed the treatment landscape of gastrointestinal (GI) malignancies. Emerging evidence suggests that gut microbiome composition plays a critical role in modulating the efficacy of immunotherapy. Antibiotics, by disrupting intestinal microbial diversity, may attenuate antitumor immune responses and compromise ICI outcomes. We evaluated time-dependent clinical outcomes associated with antibiotic exposure using real-world data. Methods: We conducted a retrospective cohort study using a large federated real-world database (2015-2025). Adults (18–89 years) with gastrointestinal or hepatobiliary malignancies (including colorectal, gastric, esophageal, and liver cancers) treated with ICI, receiving systemic antibiotics, were compared with matched controls without antibiotic exposure. Cohorts were balanced using 1:1 Propensity score matching for demographics, comorbidities, and treatment variables. The primary outcome was all-cause mortality at 30 days of ICI initiation; the secondary outcomes were hospitalization secondary to infectious and organ-related complications, overall survival at 1 and 3 years. Pre-specified subgroup analyses were performed in older adults ( > / = 65 years). Effect estimates were reported as risk ratios( RR), hazard ratios (HR), and absolute risk differences. Survival was analyzed using Kaplan–Meier curves and Cox proportional hazards models. Results: Within 30 days, antibiotic exposure was associated with significantly increased risks of Clostridioides difficile infection (RR 2.39), pneumonia (RR 2.93), acute kidney injury (RR 1.40), acute respiratory failure (RR 2.11), and neutropenia (RR 1.18) (all p < 0.05). Absolute risk differences translated to a number needed to harm (NNH) of approximately 38 for pneumonia, 143 for acute kidney injury, and ~140 for C. difficile infection. One-month mortality was modestly higher in the antibiotic group (5.4% vs 4.5%; HR 1.22; p = 0.014). However, no significant differences in mortality were observed at 6 months, 1 year, or 3 years, and survival curves converged over time. In the subgroup analysis of adults ≥65 years, early adverse event patterns were preserved, while long-term mortality remained unchanged. One-year hospital admission rates did not differ significantly between groups. Conclusions: Antibiotic exposure is associated with substantial early infectious and organ-related harm and a small increase in short-term mortality, without improvement in long-term survival, including among older adults. These findings highlight a critical opportunity for risk-adapted antibiotic stewardship and careful reassessment of empiric antibiotic use. Key words: ICI- Immune checkpoint inhibitors, RR- Risk ratio, HR- Hazard Ratio, NNH- Number Needed to Harm.

e15557 Background: In metastatic colorectal cancer (mCRC), systemic chemotherapy can enable conversion surgery in a subset of patients with initially unresectable disease. Although R0 resection considered a key determinant of long-term survival, its clinical impact across heterogeneous trial populations and treatment strategies remains insufficiently defined. We aimed to clarify the survival benefit of R0 resection after first-line chemotherapy using large-scale individual patient data. Methods: Using the ARCAD database, which comprises 67 randomized clinical trials and approximately 47,000 patients, we conducted a pooled analysis of individual patient data from six first-line randomized clinical trials (COIN, COIN-B, 03-TTD-01, ATOM, SOFT and PARADIGM) for which data on R0 resection were available. Patients with initially unresectable mCRC and available surgical data were categorized into three groups: R0 resection, non-R0 resection, and no surgery. All R0 cases were included, comprising liver-only, lung-only, combined liver and lung metastases, and peritoneal dissemination or other metastatic sites. Overall survival (OS) and progression-free survival (PFS) were evaluated using multivariable Cox proportional hazards models adjusted for age, sex, and performance status. Pre-specified subgroup analyses examined the consistency of outcomes according to tumor response, metastatic site, primary tumor sidedness, RAS status, and targeted therapy class. Results: Among 4,445 eligible patients, 263 underwent R0 , 233 underwent non-R0 resection, and 3,949 received no surgical intervention. The R0 resection group comprised patients with liver-only metastases (n = 176), lung-only metastases (n = 23), and combined liver and lung metastases (n = 27). R0 resection was associated with significantly improved OS and PFS compared with both non-R0 resection and no surgery. These survival benefits were consistently observed across key clinical subgroups, including patients achieving complete or partial response to chemotherapy, those with liver or lung metastases, left- and right-sided primary tumors, and both RAS wild-type and mutant disease. Importantly, the benefit of R0 resection was maintained irrespective of anti-EGFR– or anti-VEGF–based first-line therapy. Conclusions: R0 resection following systemic chemotherapy is strongly associated with durable survival benefits in patients with initially unresectable mCRC across diverse clinical settings. These findings support the pivotal role of conversion strategies and highlight the importance of multidisciplinary evaluation to identify appropriate candidates for curative-intent resection.

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.