4211 Background: Celiac plexus radiosurgery is a novel non-invasive palliative treatment for pancreatic cancer pain that was recently added to NCCN guidelines. In the pivotal phase 2 trial (NCT03323489; Lancet Oncol 2024;25:1070-79), pain response was 53% (95% CI 42-64%). Prespecified exploratory analysis identified age and BMI as predictors; however, no clinically actionable patient selection tool was developed. We sought to identify additional predictors and create a practical risk stratification score. Methods: Post-hoc analysis of 90 evaluable patients from the phase 2 trial. Pain response was defined per protocol (≥2-point reduction in average pain from baseline to three weeks, on Brief Pain Inventory–Short Form). Candidate baseline predictors were examined by univariate and multivariate logistic regression. Only variables remaining significant in multivariate analysis were included in the final score to ensure independence and avoid overfitting. Internal validation used bootstrap resampling (500 iterations with replacement) to calculate optimism-corrected AUC. Analysis performed using Stata IC/16.1. Results: Univariate analysis identified 4 significant predictors: neurotoxic chemotherapy exposure (OR 5.33, 95% CI 2.13-13.4, p<0.001), baseline pain intensity (OR 1.73, p=0.003), age (OR 1.06, p=0.014), and therapy line (OR 0.65, p=0.04). On multivariate analysis, only neurotoxic exposure (OR 5.1, p=0.009) and baseline pain (OR 1.8, p=0.003) remained independently significant predictors. Age and therapy line lost significance due to collinearity. Dose-response analysis confirmed pain threshold optimization: >5 (61.5% response), >6 (65.8%), >7 (83.3%), >8 (85.7%), with >6 providing optimal discrimination. We created the SPIN (Severe Pain + Intact Nerves) score (0-2 points): baseline Pain >6 (+1), No prior Neurotoxic chemotherapy (+1). Response rates by score: 0 points: 32% (n=31); 1 point: 53% (n=40); 2 points: 89% (n=19). The 2-variable model achieved an apparent AUC=0.716. Following bootstrapping, the optimism-corrected AUC was=0.714. Conclusions: The simple score allows identification of distinct patient subgroups with markedly different probabilities of pain response following celiac plexus radiosurgery. This suggests the intervention should be considered earlier in the disease course, before exposure to neurotoxic agents. External validation is needed prior to clinical implementation. Financial support: Gateway for Cancer Research, The Israel Cancer Association. Clinical trial information: NCT03323489 .

Patient advocates represent the voice of the patient community and bring a unique perspective to research. We hypothesized that including patient advocates in research training would increase trainees ability to communicate their science, understand the impact of their work and increase their empathy for patients. An IRB-approved survey was administered to assess the impact of patient advocate feedback on trainees who presented their research via posters at our Radiation Oncology Research Symposium. Trainees reported demographics and self-assessed their use of lay language, changes in empathy, understanding the impact of their research, and potential future implications. The binomial proportion test and Fisher's exact test were used to determine significance. The survey was completed by 80% (28/35) of trainees who participated in the poster session and interacted with patient advocates. Trainees were predominantly younger (60.7% under 30yo) and people of color (60.7%). Almost all trainees (96.4%) were comfortable talking to advocates but only 89.29% were comfortable using lay language. Trainees agreed (75%) that interacting with advocates increased their empathy. Most trainees (71.4%) believed patient advocates helped them understand the significance of their research and 64.3% believed advocates helped them develop new research ideas. Most trainees would like advocates at future poster presentations (85.7%), but did not want them to participate in study design or analysis. Gender and training level did not affect trainees' ability to communicate in lay language, their empathy for patients, or understanding of their work's clinical relevance. Including patient advocates in poster sessions may improve trainees ability to present their research in lay language, increase their empathy and understand the clinical impact of their research. Future radiation oncology training should consider including the patient advocate voice to improve the tangible connection between research and real-world impact.

Background/Objectives: This study explored whether a multimodal artificial intelligence (MMAI) model integrating digitized histopathology and clinical features can identify prostate cancer patients who may benefit from neoadjuvant hormonal therapy (NHT) and whole-pelvic radiotherapy (WPRT). Methods: This secondary analysis of NRG/RTOG 9413 included NHT-treated patients with digitized biopsy slides and clinical data who were not part of the MMAI model optimization. A previously validated MMAI model estimated long-term risk, and Fine-Gray models evaluated interactions between MMAI-derived scores and the radiation field (WPRT vs. prostate-only RT [PORT]) for biochemical failure (BF), chosen over progression-free survival because of extended follow-up and distant metastasis (DM), with subgroup analyses by predefined MMAI strata. Results: Among 81 eligible patients, the MMAI-by-treatment interaction for BF did not confirm a differential effect (p = 0.30). Therefore, subgroup findings should be interpreted as descriptive and hypothesis-generating. Nevertheless, the magnitude effect of WPRT was numerically greater in the MMAI high-risk subgroup (5-yr: 41% vs. 79%; 10-yr: 47% vs. 79%; aHR 0.35 [0.14-0.86]) than in the low-intermediate group (5-yr: 18% vs. 33%; 10-yr: 44% vs. 57%; aHR 0.66 [0.29-1.48]). Conclusions: Although no statistically significant treatment-by-MMAI interaction was demonstrated, these findings are hypothesis-generating and support further investigation of MMAI approaches for guiding WPRT in NRG/RTOG 0534, 0924, GETUG-01, and POP-RT trials.

BACKGROUND:Meningiomas exhibit clinical heterogeneity. Radiotherapy (RT) remains the only adjuvant therapy, but tumor-control is variable, and biomarkers are limited. NRG/RTOG-0539 is the first prospective phase 2 trial to stratify meningioma patients for adjuvant RT based on clinical risk. Here, we apply modern molecular tools to this cohort and identify correlates of RT response. METHODS:Tumor tissue from 100 RTOG-0539 patients was profiled using DNA methylation arrays, RNA sequencing, and whole-exome sequencing. Recurrence scores, Molecular Groups, gene expression, and copy number alterations were compared across clinical groups and between RT responders and non-responders; non-response was defined as progression or death within 3 years. RESULTS:Modern grading criteria, including brain invasion, TERT mutation, CDKN2A/B deletion, and 1p/1q status, would reclassify 10% of tumors and alter treatment group assignment in 7%. Non-responders to RT exhibited more frequent 1p and 14q loss, and more copy number alterations. Transcriptomic and epigenetic profiling revealed immune-related signatures in responders and cell cycle-related pathways in non-responders, several of which overlapped with targets of vorinostat, a histone deacetylase inhibitor previously validated in aggressive meningioma models. The Proliferative Molecular Group was an independent predictor of post-RT recurrence in multivariable analysis, outperforming WHO grade. CONCLUSION:Multi-omic analysis of the NRG/RTOG-0539 cohort shows that updated WHO grading criteria, incorporating molecular and cytogenetic features, improve risk stratification. However, molecular classification, particularly the Proliferative group, remains an independent and stronger predictor of RT response. These findings support integrating molecular biomarkers alongside modern grading frameworks to guide treatment and trial design in meningioma. CLINICAL TRIAL INFORMATION:NCT00895622.

Abstract Introduction: Plasma proteomics provides a comprehensive view of the biological processes active in an individual. Studies in healthy populations have shown that the plasma proteome undergoes changes with aging, and that proteomics-based models can predict biological age. The difference between biological and chronological age, known as the “age gap,” reflects an individual’s risk of disease. Extending this approach, organ-specific aging models have demonstrated that accelerated aging of individual organs is associated with organ-related disorders. Here, we examined age-associated effects on the plasma proteome of cancer patients across multiple tumor types and applied organismal and eleven organ-specific proteomic aging models to identify links between aging, tumor characteristics, and clinical features. Methods: Baseline plasma samples were collected from patients with metastatic solid tumors (non-small cell lung cancer [NSCLC], n=818; small cell lung cancer [SCLC], n=99; renal cell carcinoma [RCC], n=297; melanoma, n=163) and healthy subjects (n=278). Deep plasma proteomic profiling was performed using an aptamer-based assay. Bioinformatic analyses identified age-associated proteomic signatures, and organismal and organ-specific predictors were applied to estimate biological ages for each patient and organ. Results: Proteins involved in multiple signaling pathways, including Wnt, PI3K-Akt, IGF, and Ephrin receptor signaling, were upregulated in older patients (≥65 years) compared with younger patients, along with immune-regulatory proteins, reflecting immune remodeling associated with aging. These trends were consistent across tumor types. The organismal biological age gap was significantly higher in all cancer cohorts versus healthy controls, largest in SCLC and smallest in melanoma, with the most substantial effect in younger patients. Organ-level analyses revealed distinct patterns: lung age gap was highest in NSCLC and SCLC, and kidney age gap was most significant in RCC. Elevated organ-specific gaps correlated with corresponding comorbidities (e.g., cardiac age gap with arrhythmia, ischemic heart disease, and vascular disease) but much less with organ-specific metastases. Focusing on the immune age gap, patients treated with immune checkpoint inhibitor-based therapy who exhibited a high immune age gap had significantly shorter overall survival compared with patients with a low immune age gap (median OS, 16.4 vs. 31.8 months; HR = 0.67, p < 0.0001). The effect varied by indication, being strongest in melanoma (HR = 0.27, p = 0.0007) and absent in SCLC (HR = 0.87, p = 0.65). Conclusions: Proteomic aging predictors capture systemic and organ-specific aging processes in cancer. Distinct age-gap patterns across tumor types, along with their association with survival and comorbidities, highlight the biological and clinical relevance of proteomic aging in oncology. Citation Format: Michal Harel, Coren Lahav, Yehonatan Elon, M. Austin Argentieri, Surbhi Singhal, Adam P. Dicker. Proteomic aging biomarkers predict survival in immunotherapy-treated tumors [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 2452.

714 Background: Celiac plexus radiosurgery demonstrates impressive short-term pain control (3-6 weeks) in pancreatic cancer patients with minimal side effects, leading to NCCN guideline inclusion. In the trial, irradiation of the adjacent tumor was left to physician's discretion, but did not impact short-term pain outcomes (the primary endpoint). In this analysis we 1) report long-term efficacy, and 2) test the hypothesis that concurrent tumor irradiation affects durability. Methods: Single-arm study of 125 patients across five countries with baseline pain ≥5/10 and pancreatic cancer or celiac axis invasion. Subjects lacking pain scores at both 3- and 6-weeks post treatment were excluded. Pain measured via Brief Pain Inventory (weeks 0-6), then Numeric Rating Scale. Linear mixed-effects models with both linear & quadratic time terms analysed pain trajectories. Sensitivity analyses for missing data included multiple imputation, 'last observation carried forward', and worst-case scenarios. Prescription dose was tested as a moderator variable in the mixed-effect model. Results: A total of 107 patients analyzed (median age 64, 54% female, 92% pancreatic cancer, 86% metastatic). Median baseline pain was 6/10 (SD 1.7). Of the 82 cases in which the adjacent tumour was included within target volume, it was prescribed a median dose of 15Gy. Patients alive at 12/26/39 weeks: 79/33/24; pain outcome available for: 52/25/16. Median survival was 18 weeks. Peak effectiveness was at 18.5 weeks (95% CI: 15.7-20.7) with 3.4-point reduction (60% improvement, p<0.001), followed by gradual waning. Multiple imputation suggested sustained efficacy through 39 weeks. Higher prescription doses to adjacent tumor were associated with improved palliative durability (dose×time² interaction: p=0.046). Conclusions: Celiac plexus radiosurgery provides clinically meaningful pain reduction for at least four months. A post-hoc hypothesis generating analysis revealed concurrent tumor irradiation enhanced durability with a dose-response relationship. These exploratory, potentially practise changing findings warrant validation in a randomized trial. Financial support: Gateway for Cancer Research , Israel Cancer Association . Clinical trial information: NCT03323489 .

1636 Background: Conventional analysis of clinical trial data requires statistical expertise, time and effort. Previous AI approaches have used structured, pre-processed datasets on specialized platforms. We hypothesized that widely available commercial large language models (LLMs) would perform rapid analysis of raw data. We evaluated three commercial LLM platforms fed identical data, and subsequently validated the outcomes with a formal statistical analysis. We name this approach "Conversational AI-assisted Exploratory Data Analysis" (CA-EDA). Methods: We used data from a completed phase 2 trial (NCT03323489; Lancet Oncol 2024; 25:1070-9); n=125; 2165 variables) that evaluated the efficacy of celiac plexus radiosurgery in controlling retroperitoneal pain syndrome amongst pancreatic cancer patients. Raw individual patient data was uploaded to three commercial LLMs: Claude Opus 4.1 (Anthropic), ChatGPT 5.2 (OpenAI) in deep research mode, and Gemini 3 Pro (Google). Each LLM received identical inputs: the raw Excel dataset, the codebook, the protocol, and the published manuscript. The prompt requested identification of baseline predictors of pain response and development of a clinical score. Outputs were compared for statistical analyses performed, predictors identified, visualizations generated, and clinical utility. Findings were formally verified using Stata IC/16.1. Results: Analysis completion time ranged from 10 to 30 minutes across all three platforms. Claude performed a comprehensive statistical analysis, generated a 9-panel visualization dashboard, identified prior exposure to neurotoxic chemotherapy as a novel predictor of response (OR 0.20, p<0.001), and created a 4-variable clinical score predictive of response (AUC 0.71). ChatGPT performed a statistical analysis but missed neurotoxic chemotherapy, creating a 3-variable score. Gemini produced an 8-page narrative with biological insights, but did not perform any statistical analysis. Analysis using Stata confirmed the association between prior exposure to neurotoxic chemotherapy and response (35% with prior exposure vs 75% without, p<0.001). Conclusions: Claude Opus 4.1 identified a clinically significant predictor that had not previously been recognized, and developed a helpful clinical score. CA-EDA always requires formal biostatistical verification due to concerns about LLMs’ tendency to hallucinate, may analyse only sampled data and potentially implement incorrect Python code. Despite these concerns, here we found CA-EDA of clinical trial data to be rapid, accurate and cost-effective. Financial support for clinical trial: Gateway for Cancer Research, The Israel Cancer Association.

BACKGROUND:Chromophobe renal cell carcinoma (ChRCC) is characterized by the accumulation of abnormal mitochondria, a high rate of mitochondrial DNA (mtDNA) mutations, and altered oxidative metabolism. There are no existing circulating biomarkers to distinguish metastatic ChRCC from clear cell renal cell carcinoma (ccRCC). METHODS:High-throughput plasma proteomic profiling using the SomaScan platform was performed in 18 ChRCC (including 16 metastatic ChRCC) and 197 metastatic ccRCC patients. Data were harmonized to generate a unified 7K-protein matrix. RESULTS:Differential expression analysis was performed using limma (version 3.62.2). Of 7272 quantified human plasma proteins, 209 were differentially expressed between ChRCC and ccRCC. Upregulated proteins in ChRCC included essential β-oxidation enzymes such as ECH1 (enoyl-CoA hydratase 1) and ECI1 (enoyl-CoA delta-isomerase 1), suggesting increased long-chain fatty acid degradation. Creatine and energy-buffering pathways were also represented, with increased CKMT1A (Creatine Kinase, Mitochondrial 1A) in ChRCC. KIM-1 (Kidney Injury Molecule-1) and leptin were lower in ChRCC, consistent with the known upregulation of these proteins in ccRCC. Pathway enrichment analyses revealed an overrepresentation of mitochondrial protein degradation, fatty acid β-oxidation, and respiratory electron transport in ChRCC, suggesting that ChRCC sheds a unique mitochondrial signature into the peripheral circulation. A bootstrap-based LASSO logistic regression restricted to upregulated mitochondrial proteins in ChRCC vs. ccRCC consistently selected ECI1 and CKMT1A. The LASSO model achieved an AUROC of 0.964. CONCLUSIONS:Compared to ccRCC, the plasma proteome of metastatic ChRCC is dominated by mitochondrial metabolic enzymes, revealing a systemic metabolic phenotype strikingly aligned with the known histologic accumulation of abnormal mitochondria in ChRCC cells.

PURPOSE:Cardiac toxicity is a significant concern for cancer patients undergoing radiotherapy. The complexity of evaluating cardiac events, the time required to parse the Electronic Health Record (EHR), and the need for large datasets make cardiotoxicity studies time- and resource-intensive, limiting progress of cardio-oncology research. Large-language models (LLMs) have the potential to automatically identify cardiac events. This work aimed to develop and validate a novel LLM-based automated cardiac event identification framework using a two-institution dataset. METHODS:411 lung and breast cancer patients from two institutions were analyzed. [Institution 1] data (n=266) were divided into a development cohort (lung, n=178) and an internal validation cohort (breast, n=88). External validation used [Institution 2]data (lung, n=145). Cardiac events were physician-adjudicated from the entire EHR patient history. Extracted EHR data comprised structured problem lists and unstructured clinical notes. We introduced the Two-phase Reasoning for Automated Cardiac Event Recognition (TRACER) framework, which combines structured term matching with LLM-based analysis of unstructured notes (including specialty-filtering, temporally aware queries, and few-shot examples). Following initial evaluation of six candidate models, the three top-performing open-source LLMs (DeepSeek-R1, Llama-3.3, Mistral-Large) were validated. Performance was evaluated against physician-adjudicated ground truth using accuracy and processing time. RESULTS:Of 411 patients, 220 had at least one cardiac event. The top three models achieved mean accuracies of 79.4%, 81.0%, and 79.3% for the development, internal validation, and external validation cohorts, respectively. DeepSeek-R1 achieved the highest accuracy on internal cohorts (83.4-85.2%), while Llama-3.3 reached 85.5% accuracy on external validation. TRACER processing time was 20-42 seconds per patient (2.3-4.8 hours total) versus 2 hours per patient for manual review (822 person-hours total). CONCLUSION:TRACER, a locally deployed LLM framework, accurately extracted cardiac events across institutions and disease sites. This approach enables scalable cardio-oncology research by substantially reducing the resources required for cardiac event identification.

e13648 Background: Artificial intelligence (AI) in cancer care offers opportunities to reduce administrative burden, improve efficiency, and support care delivery, yet it prompts questions around safety, equity, and clinical judgment. To assess real-world perspectives, the Association of Cancer Care Centers (ACCC) conducted a national survey examining how oncology professionals engage with AI, their perceived value, and the barriers to adoption. Methods: ACCC, with guidance from an expert committee, created a national online survey of multidisciplinary staff at US cancer programs from May and August 2025. The 26-item survey assessed experiences and perceptions of AI, organizational adoption and governance, and implementation barriers and facilitators. Completed responses were analyzed using descriptive and stratified statistics in Stata 18, with qualitative data examined using rapid inductive thematic analysis. Results: Respondents (N=168) from 36 states included 61% in care delivery and 39% in administrative/operational roles, primarily from community (46%) and NCI-designated or academic (44%) programs. Care delivery staff were less confident than administrative/operations respondents in describing AI use (21% vs 4% “not at all [confident]”), discussing benefits of AI use (18% vs 7%), and critically evaluating AI systems (26% vs 13%). Administrative/operations respondents reported higher confidence contributing to AI implementation (45% vs 32%) and a higher likelihood to adopt/expand AI for treatment planning (52% vs 34%), clinical trial matching (49% vs 32%), and prior authorization (52% vs 35%). Rural respondents were more likely to express concern about declines in patient-provider communication (63% vs 22% suburban and 30% urban), as were those without AI experience (54% vs 24%). Academic respondents rated improving clinical decision-making/diagnostic accuracy, and establishing performance benchmarks and evaluating AI systems, as higher motivators for integration (0.82 vs 0.41; 0.53 vs 0.26), whereas community respondents rated patient engagement higher (0.36 vs 0.13). Respondents with AI experience reported a higher likelihood to adopt/expand AI for chatbots and virtual assistants (52% vs 16%), real-time alerts (35% vs 14%), personalized patient education (46% vs 27%), and individualized patient navigation (38% vs 11%). Qualitative responses reinforced these findings, highlighting stakeholder engagement, evidence of effectiveness, training, integration, and governance as critical for AI implementation. Conclusions: The results reveal broad AI use despite limited AI tools and governance implementation. Respondents were confident acknowledging AI’s limitations but less confident in evaluating its utility, highlighting opportunities for ACCC to support adoption, governance, and AI integration focused on efficiency and patient outcomes.

5000 Background: Outcomes for some patients with ≥high-risk localized prostate cancer (HR-PCa) who receive radiation therapy (RT) remain poor. Current guidelines recommend intensification with abiraterone/prednisone (AAP) in select patients; however, the current guideline-defined candidate population may be suboptimal. Methods: This is a secondary analysis of available biopsy samples from NRG/RTOG 9202, 9413, 9902, and 0521. The primary objective was to determine if a specific subgroup of ≥HR-PCa patients had sufficiently poor prognoses to derive a clinically meaningful benefit from AAP intensification. The prognostic impact of the Decipher 22-gene genomic classifier (GC; Veracyte, San Diego, CA) was first evaluated for metastasis-free survival (MFS, primary), overall survival (OS), and distant metastases (DM). GC scores were analyzed using both continuous (per 0.1 GC units) and pre-specified categorical cutpoints: ≤ intermediate transcriptomic risk (IR-PCa, GC≤0.6), HR-PCa (GC:0.6-0.85), and very high risk (VHR-PCa, GC > 0.85). A tree-based model was then used to combine clinical and transcriptomic risk category. Results: This study characterized outcomes of 427 patients (64% NCCN HR-PCa; 36% NCCN VHR-PCa) with a median follow-up of 10.4 years. Continuous GC score was observed to discriminate risk when adjusted for age, PSA, grade, T-stage, and treatment received for the endpoints of MFS (HR GC :1.19 [95% CI:1.10-1.29], p < 0.001), DM (HR GC :1.31 [95% CI:1.13-1.51], p < 0.001), and OS (HR GC : 1.18 [95% CI: 1.09-1.28], p < 0.001). MFS and OS were estimated using the Kaplan-Meier method by pre-specified GC subgroup within each NCCN risk group (HR-PCa, VHR-PCa). Based on outcomes by subgroup, a novel clinico-transcriptomic risk stratification system was constructed, where patients receive 1 point for NCCN/GC HR categorization and 2 points for NCCN/GC VHR categorization. The combined risk score ≥3 subgroup had a prognosis similar to the STAMPEDE M0 control arm, whereas the combined score ≤2 subgroup had a substantially better prognosis. Utilizing this system could enhance the specificity of AAP recommendations, increasing the candidate pool eligible for AAP intensification by approximately 20% and identifying approximately 25% of NCCN VHR in whom the therapeutic ratio of AAP may be less favorable. Conclusions: This study defines a novel clinico-transcriptomic risk stratification system to augment current clinical eligibility for AAP intensification. This system may increase the population who benefit from AAP by approximately 20% and simultaneously identify approximately 25% of NCCN VHR patients who may avoid AAP intensification. Acknowledgement: We would like to acknowledge Dr. Felix Feng for his contributions in his capacity as the former GU Committee Chair.

Background Immune checkpoint inhibitors (ICIs) have shown substantial benefit for patients with advanced non-small cell lung cancer (NSCLC). However, resistance to ICIs remains a major clinical challenge. Here, we perform a comprehensive bioinformatic analysis of plasma proteomic profiles to explore the underlying biology of treatment resistance in NSCLC.Methods The analysis was performed on 388 “resistance-associated proteins” (RAPs) that were previously described as pretreatment plasma proteomic predictors within the PROphet computational model designed to predict ICI clinical benefit in NSCLC. Putative tissue origins of the RAPs were explored using publicly available datasets. Enrichment analyses were performed to investigate RAP-related biological processes. Plasma proteomic data from 50 healthy subjects and 272 patients with NSCLC were compared, where patients were classified as displaying clinical benefit (CB; n=76) or no CB (NCB; n=196). Therapeutic agents targeting RAPs were identified in drug and clinical trial databases.Results The RAP set was significantly enriched with proteins associated with lung cancer, liver tissue, cell proliferation, extracellular matrix, invasion, and metastasis. Comparison of RAP expression in healthy subjects and patients with NSCLC revealed five distinct RAP subsets that provide mechanistic insights. The RAP subset displaying a pattern of high expression in the healthy population relative to the NSCLC population included multiple proteins associated with antitumor activities, while the subset displaying a pattern of highest expression in the NCB population included proteins associated with various hallmarks of treatment resistance. Analysis of patient-specific RAP profiles revealed inter-patient diversity of potential resistance mechanisms, suggesting that RAPs may aid in developing personalized therapeutic strategies. Furthermore, examination of drug and clinical trial databases revealed that 17.5% of the RAPs are drug targets, highlighting the RAP set as a valuable resource for drug development.Conclusions The study provides insight into the underlying biology of ICI resistance in NSCLC and highlights the potential clinical value of RAP profiles for developing personalized therapies.

BACKGROUND:Immune checkpoint modulators (ICMs) have revolutionized cancer treatment but have unique immune-related adverse events (irAEs). The aim of this study was to develop and validate a brief measure of the most common, distressing, and diagnostically useful symptomatic irAEs. METHODS:Items were generated to assess symptomatic irAEs through a multistep process of (1) literature review and iterative expert input and (2) qualitative interviews of patients, caregivers, and clinicians regarding ICM-related irAEs and quality of life (QOL) impacts. An initial item set was administered across five longitudinal or cross-sectional studies. The final item set was selected using a Delphi method; validity, reliability, minimally important differences (MIDs), and sensitivity to change were evaluated. RESULTS:Qualitative interviews with 14 patients, seven caregivers, and six clinicians informed an initial set of 46 symptomatic irAEs, which was administered to patients (N = 503, 52% female, mean age = 64) treated with ICMs for non-small cell lung cancer (n = 342), head and neck cancer (n = 72), renal cell carcinoma (n = 43), or melanoma (n = 46). A final item set was selected and mapped to FACIT library items. This produced the 17-item FACT-ICM Symptom Index, which showed reliability (α = 0.86), construct validity (comparative fit index = 0.93), convergent validity with validated measures of physical QOL (r = 0.69-0.73), discriminant validity with emotional and social QOL (r = 0.03-0.65), and criterion validity (i.e., better performance status was associated with fewer concerns). Response option anchors adequately captured MIDs and were sensitive to change. CONCLUSIONS:This brief 17-item FACT-ICM Symptom Index demonstrates initial construct, convergent, and divergent validity, reliability, MID, and sensitivity to change and is ready for use in research and clinical care.

Background/Objectives: Weight management is directly linked to cancer recurrence and survival, but unfortunately, nutritional oncology counseling is not typically covered by insurance, creating a disparity for patients without nutritional education and food access. Novel ways of imparting personalized nutrition advice are needed to address this issue. Large language models (LLMs) offer a promising path toward tailoring dietary advice to individual patients. This study aimed to assess the capacity of LLMs to offer personalized dietary advice to patients with breast cancer. Methods: Thirty-one prompt templates were designed to evaluate dietary recommendations generated by ChatGPT and Gemini with variations within eight categorical variables: cancer stage, comorbidity, location, culture, age, dietary guideline, budget, and store. Seven prompts were selected for four board-certified oncology dietitians to also respond to. Responses were evaluated based on nutritional content and qualitative observations. A quantitative comparison of the calories and macronutrients of the LLM- and dietitian-generated meal plans via the Acceptable Macronutrient Distribution Ranges and United States Department of Agriculture’s estimated calorie needs was performed. Conclusions: The LLMs generated personalized grocery lists and meal plans adapting to location, culture, and budget but not age, disease stage, comorbidities, or dietary guidelines. Gemini provided more comprehensive responses, including visuals and specific prices. While the dietitian-generated diets offered more adherent total daily calorie contents to the United States Department of Agriculture’s estimated calorie needs, ChatGPT and Gemini offered more adherent macronutrient ratios to the Acceptable Macronutrient Distribution Range. Overall, the meal plans were not significantly different between the LLMs and dietitians. LLMs can provide personalized dietary advice to cancer patients who may lack access to this care. Grocery lists and meal plans generated by LLMs are applicable to patients with variable food access, socioeconomic means, and cultural preferences and can be a tool to increase health equity.

400 Background: The ArteraAI Prostate Test (v1.2), a digital pathology-based multimodal artificial intelligence (MMAI) biomarker, was developed and validated using clinical data (age, T stage, PSA) and whole slide images (WSI) from prostate biopsies to prognosticate risk of 10-year distant metastasis for men with localized prostate cancer (PCa). This study aimed to apply the MMAI biomarker to prostatectomy (RP) tissue microarray (TMA) samples and, for the first time, explore the impact of using RP TMA in place of RP WSI on MMAI scores. Methods: The analysis cohort included men with localized PCa who had undergone RP. Black men were matched in a 4:1 ratio to White men with similar baseline characteristics. MMAI scores were generated using digitized TMA and WSI from each patient’s RP specimen. The normality of score distribution was examined using the Shapiro-Wilk (SW) test. Wilcoxon signed rank (WSR) test and Spearman rank coefficients were used to compare TMA-derived and WSI-derived scores. Analyses were performed in the entire cohort and Black or White subgroups. Results: Paired MMAI scores were generated for 98 men with a median age of 58, PSA 5.4 ng/mL, the majority were Gleason grade groups 1-2 (84%), 82% were Black. The distribution of TMA-derived scores (SW test, P=0.04) was more skewed than that of WSI-derived scores (SW test, P=0.39). The median MMAI score was significantly higher for WSI images than for TMA images in all men and in Black men (WSR test, P<0.01). There was no notable variation in either set of scores by race (Table). The correlation coefficient was 0.496 (0.519 for Black men). Using pre-specified cutoffs in MMAI scores (high, intermediate, low), 30/98 (31%) men were classified into lower risk groups by TMA than WSI scores (23/80 [29%] Black men). Conclusions: In this application of the MMAI biomarker to RP samples within this cohort of primarily Black men, we found WSI-derived and TMA-derived MMAI scores to be significantly different but moderately correlated. Given that TMA represents only a portion of WSI and sampling location may impact the results, TMA may be insufficiently reliable for MMAI score generation. MMAI biomarkers that are robust across different specimen preparation methods have potential for clinical and research utility; these hypothesis-generating results support further research along these lines. Continuous MMAI scores and categorical MMAI risk group distribution using WSI- and TMA-derived images by subgroup. WSI-derived score TMA-derived score Patients Continuous 1 Low 2 Int 2 High 2 Continuous 1 Low 2 Int 2 High 2 All 0.47(0.44-0.53) 0(0%) 69(70%) 29(30%) 0.37(0.33-0.42) 6(6%) 87(89%) 5(5%) Black 0.47(0.45-0.51) 0(0%) 58(73%) 22(27%) 0.37(0.33-0.43) 4(5%) 73(91%) 3(4%) White 0.48(0.43-0.54) 0(0%) 11(61%) 7(39%) 0.37(0.32-0.41) 2(11%) 14(78%) 2(11%) 1 Presented as median (IQR). 2 Presented as n (%).

Glioblastoma is the most aggressive adult brain tumor, with a median overall survival of approximately 15 months. It is important to build accurate prognostic models for glioblastoma patients to inform clinical management and trials. This study proposes a self-supervised learning-based approach with multimodal data integration for survival prediction and prognostic stratification of glioblastoma patients on the ReSPOND consortium. We curated a multi-parametric MRI dataset (T1, T1CE, T2, FLAIR) of 3,119 glioblastoma patients from 22 institutions across 3 continents. Masked autoencoder (MAE) was adapted to pretrain a Vision Transformer (ViT) encoder by reconstructing the masked image patches. The encoder was utilized for extracting patch embeddings for survival tasks, with cross-attention mechanism to incorporate the molecular and clinical information (age, sex, extent of resection, MGMT) to guide imaging feature aggregation. Imaging and clinical embeddings were fused through a multi-layer perceptron (MLP) for log-risk hazard estimation, optimized using Cox partial likelihood. Model performance and generalizability were assessed via k-fold cross-validation on the ReSPOND consortium and the leave-one-site-out validation was performed on 11 institutions comparing with CoxPH, DeepSurv and DeepHit. Prognostic risk stratification via Kaplan-Meier analysis divided the patients into low-, medium- and high-risk subgroups per site. Multimodal data integration using the proposed framework achieved the highest C-index (0.674 ± 0.017) on the ReSPOND consortium. Integration of clinical information and MGMT consistently boosted the performance of the proposed model across sites (0.615 ± 0.046 vs. 0.662 ± 0.044). The imaging-based approaches, i.e., radiomics and convolutional neural network (CNN) features performed less robustly. The Kaplan-Meier curves and log-rank tests suggested the proposed framework achieved more separable prognostic subgroups. The proposed self-supervised multimodal learning framework shows promise for survival prediction and prognostic risk stratification in glioblastoma. It highlights the challenge for clinical model deployment due to the data heterogeneity in multi-institutional cohort.

Computational competitions are the standard for benchmarking medical image analysis algorithms, but they typically use small curated test datasets acquired at a few centers, leaving a gap to the reality of diverse multicentric patient data. To this end, the Federated Tumor Segmentation (FeTS) Challenge represents the paradigm for real-world algorithmic performance evaluation. The FeTS challenge is a competition to benchmark (i) federated learning aggregation algorithms and (ii) state-of-the-art segmentation algorithms, across multiple international sites. Weight aggregation and client selection techniques were compared using a multicentric brain tumor dataset in realistic federated learning simulations, yielding benefits for adaptive weight aggregation, and efficiency gains through client sampling. Quantitative performance evaluation of state-of-the-art segmentation algorithms on data distributed internationally across 32 institutions yielded good generalization on average, albeit the worst-case performance revealed data-specific modes of failure. Similar multi-site setups can help validate the real-world utility of healthcare AI algorithms in the future.

582 Background: Kidney injury molecule-1 (KIM-1) is overexpressed in clear cell and papillary renal cell carcinoma (RCC) and in proximal tubular kidney injury. While circulating KIM-1 is a minimally invasive biomarker for RCC, it is unknown whether kidney disease, a common comorbidity among RCC patients, impacts the association of KIM-1 with RCC outcomes. We evaluated the association between KIM-1 and outcomes in metastatic RCC after adjustment for multiple kidney injury biomarkers using plasma proteomics. Methods: Plasma samples from patients with metastatic clear cell and papillary RCC were obtained prior to 1 st line systemic therapy. Samples were analyzed using a high-throughput aptamer-based proteomics assay (SomaLogic), and results were log-transformed for analysis. Clinical and laboratory data, as well as cancer outcomes, were retrospectively curated. Spearman’s ρ was used to evaluate correlations between circulating KIM-1 and kidney injury biomarkers (cystatin C, TNFR1, TNFR2, eGFR). Cox regression analyses were used to evaluate the association between KIM-1 as a continuous variable and overall survival (OS) and progression-free survival (PFS), after adjusting for kidney injury biomarkers. Performance of KIM-1 tertiles versus IMDC risk groups for prognosticating OS was evaluated using the C-index. Results: Among 210 patients, higher baseline KIM-1 was associated with worse PFS (p = 0.004) and OS (p < 0.001) in univariate Cox regression analysis (Table). The prognostic value of KIM-1 was consistent across clear cell and papillary RCC (p-value for interaction = 0.98). KIM-1 remained prognostic for PFS (p = 0.01) and OS (p < 0.001) after multivariable adjustment for kidney injury biomarkers and eGFR (Table). Median follow-up was 22.1 months. Kidney injury biomarkers (cystatin C, TNFR1, TNFR2, eGFR) were correlated with each other but not with plasma KIM-1. KIM-1 tertiles (high/medium/low) were more prognostic for OS than the IMDC risk groups (C-index, KIM-1 0.63 vs. IMDC groups 0.58) and the addition of KIM-1 to the IMDC model improved its performance (C-index, KIM-1 + IMDC groups 0.64). Conclusions: Plasma KIM-1 was associated with PFS and OS in metastatic clear cell and papillary RCC. Plasma KIM-1 was not correlated with kidney injury biomarkers, suggesting that at least in metastatic RCC, circulating KIM-1 derives predominantly from tumor rather than benign kidney. The addition of KIM-1 improves IMDC model performance and may be useful for risk prognostication in RCC. Association of KIM-1 with PFS and OS in metastatic RCC. Multivariable models are adjusted for kidney injury markers (cystatin C, TNFR1, TNFR2) and eGFR. log KIM-1 HR (95% CI) p-value OS (univariate) 1.4 (1.2 – 1.7) <0.001*** PFS (univariate) 1.2 (1.1 – 1.4) 0.004** OS (multivariate) 1.4 (1.2 – 1.6) <0.001*** PFS (multivariate) 1.2 (1.1 – 1.3) 0.01**