Citations

Abstract Large Language Models (LLM) are being widely adopted into the medical field for their impressive ability to analyze and summarize large amounts of text data. These models enable clinicians and researchers to extract meaningful insights from complex datasets and may assist with decision making. Here we present our workflows and application of LLM for the interpretation and summarization of clinical data related to CAR-T cell therapy. Using an LLM (Gemini 2.5 pro) in the Google Cloud Computing (GCP) environment, two applications were developed to analyze CAR-T cell therapy clinical data: 1) extracting and summarizing CRS and ICANS event-related data to streamline the compliance team workflow, and 2) identifying features available at time of CAR-T infusion able to classify patients into high- or low-monitoring needs 14 days post CAR-T. Patient data (vitals, labs, hematology notes, and EKGs) were extracted from the electronic medical record (EMR) using Google BigQuery into SQL tables in GCP. For each application, relevant data fields were retrieved, formatted into JSON objects, and embedded in the LLM prompt for context-aware processing. Both applications have undergone iterative prompt engineering after analyzing the LLM output against ground truth data in the EMR. For the application extracting and summarizing CRS and ICANS events, the LLM was optimized on Mayo Clinic Rochester data. When compared to the IEC compliance database, the LLM achieved 100% accuracy and F1 score for CRS events, and 96% accuracy and 82% F1 score for ICANS events. The match rate for CRS and ICANS grades were 83% and 89% respectively. We applied the same LLM to Mayo Clinic Arizona (MCA) and Mayo Clinic Florida (MCF) cases, who document clinical notes and toxicity flowsheet differently, and achieved an accuracy (MCA: 93%, MCF:93%) and F1 score (MCA: 96%, MCF: 96%) for CRS and accuracy (MCA: 81%, MCF:82%) and F1 score (MCA: 76%, MCF: 75%) for ICANS. LLM was able to capture events missed by manual reviews. For most of the discrepancies where compliance team final adjudication is needed, LLM will be updated to flag discrepancies for review by the compliance team. For the application related to the monitoring needs 14 days post CAR-T, the LLM identified 5 categories predictive of high or low monitoring needs post CAR-T infusion (disease status, inflammatory and tumor burden markers, hematologic status, renal function and performance status). Our model achieved a sensitivity of 85.7%, specificity of 23.8%, and F1 score of 65.5% for our first cohort, compared with data extracted by the LLM from the EMR. For our second cohort, with demographics statistically similar to cohort 1 and using the same 5 categories, the LLM achieved a sensitivity of 83.3%, a specificity of 28.6%, and F1 score of 65.4%. Our workflow and applications of LLM’s provide examples and guidance to others interested in applying LLM’s for clinical and research applications. Citation Format: Emmanuel Contreras Guzman, Matthew Jankowski, Andre De Menezes Silva Corraes, Malvika Gupta, Monica L. Shaw, Madiha Iqbal, Talal Hilal, Saurabh Chhabra, Ricardo Daniel Parrondo, Jody K. Mclean, Kim R. Riester, Kayla Joseph, Melinda Tan, Holly Ross, Cleyonia Barnett, Sylvia Carter, Semy Girmay, Rachel Wolan, Milana Ramsey, Christian Downhour, Kristy Morgan, Shae Sibley, Erica Rushing, Lucy Holmes, Allison Burgstahler, Stephen M. Ansell, Hassan Alkhateeb, Matthew Hathcock, Ramona Bruno, Allison C. Rosenthal, Hemant Murthy, Patrick B. Johnston, Jonas Paludo, Yi Lin. Applications of large language models to CAR-T cell therapy clinical data using Google cloud computing [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 2740.

Abstract Background: Digital access, health literacy, and confidence in obtaining cancer information are well-known factors that influence cancer beliefs and behaviors. However, less is understood about whether these information-related factors vary across regions within a cancer center's catchment area and how they collectively relate to cancer beliefs in large community samples. This study investigated regional differences in digital access, health literacy, and confidence in seeking information, and examined the connections between these factors and cancer risk perceptions, fatalistic beliefs, prevention confusion, and willingness to participate in cancer clinical trials. Methods: Data (N=2,334) from the 2023 Mayo Clinic Comprehensive Cancer Center’s Cancer-Focused Needs Assessment, conducted across catchment areas in Arizona, Florida, and the Midwest (Minnesota, Wisconsin, Iowa). Regional differences in digital access (Internet use and device ownership), health literacy, and confidence in obtaining cancer information were analyzed using multivariable ANCOVAs and logistic regressions, adjusting for sociodemographic covariates. Follow-up models examined whether these information precursors are associated with perceived cancer risk, fatalistic beliefs, confusion about recommendations, and willingness to participate in cancer clinical trials, adjusting for the same covariates. Results: Regions showed minimal differences in digital access and health literacy. Health literacy was instead associated with ethnicity, age, gender, and education (all p<.001). Confidence in obtaining cancer information followed similar trends, with significant influences of ethnicity and gender (p<.01). Internet use was strongly associated with sociodemographic factors (all p<.01). Additionally, the information precursors consistently correlated with cancer beliefs across the full sample: Higher health literacy and greater confidence in obtaining cancer information were associated with increased perceived cancer risk, reduced cancer fatalism, and less confusion regarding prevention guidelines (ps < .001). Internet use and device access were significantly related to a greater willingness to participate in clinical trials (ps < .01). These relationships persisted even after controlling for sociodemographic covariates. Conclusions: Regional differences in digital access and health literacy were minimal, indicating a consistent information environment. Within this context, individual sociodemographic and information-related factors played a larger role, strongly shaping cancer beliefs. These findings underscore technology access and information empowerment as key targets for enhancing cancer prevention knowledge and research engagement, and indicate that the effects of information empowerment are primarily driven at the individual rather than the geographic level. Citation Format: Tingyu Zou, Monica L. Albertie, Emelina Asto-Flores, Manisha Salinas, Farhia Omar, Adeline Abbenyi, Noreen Stephenson, Nuwanthi Heendeniya, James R. Cerhan, Kathleen J. Yost, Folakemi T. Odedina. Digital access, health literacy, and cancer beliefs: Findings from a multi-region community survey [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 7553.

Biological heterogeneity among different breast cancer (BC) subtypes results in markedly varying clinical outcomes. Identification and analysis of key gene biomarkers that are differentially regulated during radiation therapy (RT) may pose multiple clinical challenges for BC treatment. The purpose of the study is to identify and analyze the expression of key gene biomarkers and their networks that are differentially regulated after hypofractionated RT. Patients with BC (cT0-T2, N0, M0) were treated with hypofractionated whole breast RT 25 Gy in five fractions, 4 to 8 weeks before breast conservation surgery (BCS). Biopsy (pre-RT; n = 5) and surgical (post-RT; n = 14 or 15) BC tumor samples were used for NanoString targeted sequencing. We identified 165 and 244 differentially expressed genes (DEGs; p < 0.05) in BC tumor samples from BC patients post-RT using the nCounter BC360 and IO360 panels, respectively. Gene networks and pathway analysis revealed that RT increases the gene signature of tumor inflammation (TIS), cytotoxicity, and apoptosis, while downregulating the gene signatures of tumor cell proliferation, differentiation, and cell adhesion, and increases the claudin-low gene score. RT-induced mammary stemness and enhanced infiltration of stroma, mast, and macrophage cells in the BC tumor microenvironment (TME). Further, the nCounter IO360 (immuno-oncology) panel analysis validated the findings of BC360 and demonstrated that RT increased the myeloid inflammation signature and chemokine expression, modulated B, T, NK, and DC cell activities, and enhanced residual cancer burden (RCB) in BC tumors, thus creating an immunosuppressive TME. Collectively, RT sensitized BC tumors by increasing the gene signature of TIS, cytotoxicity, apoptosis, and mammary stemness. RT facilitated an immunosuppressive environment and increased RCB, suggesting that the therapeutic potential of RT is highly individualized for each patient based on their unique tumor biology, genetic makeup, and TME.

Background: Distal femoral osteotomy (DFO) may be performed via either a lateral opening-wedge (LOW) or medial closing-wedge (MCW) osteotomy with meaningful clinical improvements; however, the effects of each DFO technique on leg length have not been well characterized. Purpose: To (1) validate a radiographic measurement technique to predict leg length changes after DFO and (2) compare the change in leg length after DFO for the LOW and MCW techniques. Study Design: Cohort study; Level of evidence, 3. Methods: Preoperative and postoperative full-length standing radiographs obtained from patients who underwent DFO at 1 of 2 academic medical centers were evaluated. The region on the medial (for LOW) or lateral (for MCW) distal femur cortex that would be the “hinge point” during DFO was identified on the radiographs. The distance from the center of the femoral head to the hinge point ( A ), the distance from the hinge point to the center of the talus ( B ), and the resultant angle (α) were measured. The equation presented within the text was used to plot the predicted leg length changes corresponding to the change in alpha angle produced by DFO. Comparisons were made between (1) the predicted and measured leg length changes after DFO and (2) the measured leg length change after LOW versus MCW osteotomies. Results: For both LOW (n = 10) and MCW (n = 10) osteotomies, the predicted leg length change was equivalent to the true change measured on postoperative radiographs (LOW: P = .16; MCW: P = .85). There was a mean lengthening of 0.85 mm (range, 0.5-1.3 mm) for every 1° of mechanical axis correction with LOW DFO, compared to a mean shortening of 0.45 mm (range, 0.1-1.4 mm) per 1° of MCW correction. Conclusion: Preoperative radiographic imaging can be used to predict leg length change after DFO with high reliability. Surgeons can expect approximately 0.85 mm of lengthening per 1° of DFO correction when performing LOW osteotomies, compared to 0.45 mm of shortening per 1° of correction for MCW osteotomies. However, the most accurate prediction results from patient-specific calculations.

Abstract Background Minimally invasive colorectal surgery is characterized by significant procedural variability, a difficult learning curve, and complications that affect patient outcomes. Video-based assessment offers an opportunity to generate data-driven insights to reduce variability, optimize training, and improve surgical performance. This mixed-methods study aimed to develop and validate a video-based assessment tool for workflow analysis across minimally invasive colorectal procedures. Methods An international steering committee panel of seven members coordinated a three-round modified Delphi process. Experts in colorectal surgery and video-based assessment were invited to reach consensus (≥ 70% agreement) on the phases and steps that describe workflows across colorectal surgery procedural videos. The resulting framework informed the development of ColoWorkflow. Four independent raters then applied ColoWorkflow to a multicentre video data set of laparoscopic and robotic colorectal procedures. Applicability and inter-rater reliability were evaluated. Results Of 66 invited experts, 41 (62%) from 11 countries completed round 1, 40 (98%) completed round 2, 20 attended the final online consensus meeting, and all (40 of 40) approved the final workflow descriptors. Consensus was achieved on nine procedure-agnostic phases (port placement and abdomen exploration; vascular dissection and ligation, mesocolon/mesorectum dissection, additional lymphadenectomy; colon and/or rectum mobilization; colorectal transection; anastomosis; completion of operation; preplanned additional procedures; unplanned procedures; extracorporeal procedures) and 34 procedure-specific steps that describe colorectal surgery workflows. ColoWorkflow was developed and applied to 54 colorectal surgery videos (left and right hemicolectomies, sigmoid and rectosigmoid resections, and total proctocolectomies) from five centres in four countries. The tool demonstrated broad applicability, as all phase and step labels, except one, were represented in at least one video. Inter-rater reliability was moderate, with a mean Cohen’s κ value of 0.72 for phases and 0.67 for steps. Most discrepancies arose at phase transitions and step boundary definitions. Conclusion Based on this preliminary work, ColoWorkflow is a promising tool for workflow analysis across minimally invasive colorectal procedures. Its adoption may help standardize training, accelerate competency acquisition, and advance data-informed surgical quality improvement.

Abstract Within the field of regenerative medicine for musculoskeletal regenerative medicine, orthobiologics and cellular therapies are evolving at a rapid pace. Exuberant ideas for new treatments often outpace their scientific evidence, which may create confusion for both patients and providers on what treatments are evidence based. Orthobiologics and cellular therapies are a subset of regenerative therapies focused on treatment of musculoskeletal disease. For the field of orthobiologics to progress, it is essential that musculoskeletal providers understand the necessary components for translating novel therapies from discovery into practice. In this article, we provide an overview of both the lexicon and regulatory processes related to translation of orthobiologics and cellular therapies along with a comparison of how to translate orthobiologics and related novel cellular therapies in hopes to dispel misinformation while promoting responsible advancements and ethical practices in medicine.

Variants in TMEM106B and GRN , which encode lysosomal proteins, interact through unknown mechanisms to increase the risk of age-related cognitive decline and neurodegeneration. Here, we show that these variants converge on a single molecular intermediate: the cleaved intra-lysosomal fibril core of TMEM106B, a precursor to amyloid fibrils that accumulate in the aging brain. A protein-coding TMEM106B risk variant (p.T185) drives fibril core accumulation by impairing its degradation and GRN risk variants amplify this effect. Mice over-expressing the fibril core develop hallmarks of neurodegeneration, and cryo-electron tomography reveals intra-lysosomal fibrils in cultured neurons, mice, and diseased human brain. In GRN -mutation carriers, in whom fibril burden is greatest, fibrils extrude through ruptured lysosomal membranes. These findings identify intra-lysosomal TMEM106B fibrillization as a convergent neurodegeneration mechanism and potential therapeutic target.

Abstract Glioblastoma (GBM) remains the most common and lethal adult malignant primary brain cancer with few treatment options. A significant issue hindering GBM therapeutic development is intratumor heterogeneity and plasticity. GBM tumors contain neoplastic cells within a fluid spectrum of diverse transcriptional states. Identifying effective therapeutics requires a platform that predicts the differential sensitivity and resistance of these states to various treatments. Here, we develop scFOCAL (Single-Cell Framework for -Omics Connectivity and Analysis via L1000), to quantify the cellular drug sensitivity and resistance landscape. Using single-cell RNA sequencing of newly diagnosed and recurrent GBM tumors, we identify compounds from the LINCS L1000 database with transcriptional response signatures selectively discordant with distinct GBM cell states, and leverage this capability to predict combination synergy. We validate the significance of these findings in vitro, ex vivo, and in vivo, and use the Olig2 inhibitor CT-179 as a reference drug to identify additional small molecules that would maximize the cell-drug discordance across the GBM transcriptional landscape. Our analysis leads to the combination of Olig2 inhibition with treatment with Depatux-M, of with which we demonstrate synergy in vivo. Our studies suggest that scFOCAL identifies cell states that are sensitive and resistant to targeted therapies in GBM using a measure of cell and drug connectivity, which can be applied to identify new synergistic combinations. Citation Format: Robert K. Suter, Anna M. Jermakowicz, Rithvik Veeramachaneni, Matthew D'Antuono, Longwei Zhang, Rishika Chowdary, Simon Kaeppeli, Madison Sharp, Pravallika Palwai, Vasileios Stathias, Grace Baker, Luz Ruiz, Winston Walters, Maria Cepero, Danielle Burgenske, Edward B. Reilly, Anatol Oleksijew, Mark G. Anderson, Sion Ll. Williams, Michael E. Ivan, Ricardo J. Komotar, Macarena I. De La Fuente, Gregory Stein, Alexandre Wojcinski, Santosh Kesari, Jann N. Sarkaria, Stephan C. Schürer, Nagi G. Ayad. Drug and single-cell gene expression integration identifies heterogeneity-aware synergistic combinations for glioblastoma [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Brain Cancer; 2026 Mar 23-25; Philadelphia, PA. Philadelphia (PA): AACR; Cancer Res 2026;86(6_Suppl):Abstract nr B032.

PURPOSE To compare mortality risk among long-term survivors of high-grade serous ovarian cancer (HGSOC) with that of the general US population and to examine how mortality and recurrence trends have evolved under changing treatment approaches. MATERIALS AND METHODS This is a retrospective cohort study including patients with histologically confirmed HGSOC treated at a single referral cancer center from 1991 to 2022. Mortality risk, expressed as standardized mortality ratios (SMRs, ratio of observed deaths in the study population to expected deaths in an age-matched US population), was assessed at diagnosis and yearly on patients without events (disease persistence, recurrence, or death from any cause). RESULTS A total of 2,074 consecutive patients were included. The median follow-up was 12.5 years (IQR, 11.7-13.8). Most patients were stage III (1,396, 69.2%), 1,299 (62.8%) underwent primary cytoreductive surgery, and 1,688 (87.4%) received platinum-taxane as first-line therapy. At diagnosis, the mortality rate was 7.40 times higher than that in the general population (95% CI, 7.04 to 7.77). By 7 event-free years, the 95% CI for the SMR included 1 for the first time, and at 10 years, the SMR was 1.05 (95% CI, 0.72 to 1.49), indicating no statistically significant excess mortality compared with the general population. In those event-free at 10 years, the 5-year cumulative incidence of ovarian cancer–related deaths (8.2% [95% CI, 4.2 to 16.0]) was comparable with that of non–ovarian cancer–related deaths (6.3% [95% CI, 2.9 to 13.8]). CONCLUSION In this cohort, mortality risk steadily declined for patients who remained event-free, approaching that of the general population by 7 years after diagnosis. These results emphasize the importance of individualized, long-term follow-up to address both recurrence risk and survivorship needs.

Abstract INTRODUCTION Cognitive impairment in hospitalized older adults without delirium is often undiagnosed. We evaluated the diagnostic accuracy and prognostic value of two brief bedside screening tools—the 10‐point Cognitive Screener (10‐CS) and Short Portable Mental Status Questionnaire (SPMSQ). METHODS Multicenter prospective cohort of 2003 patients ≥65 years of age without delirium, admitted to 43 hospitals in five countries of the Creating a Hospital Assessment Network in Geriatrics (CHANGE) Study. Screening was completed within 48 h of admission. Dementia was defined as informant‐based Clinical Dementia Rating ≥1. RESULTS Dementia prevalence was 22%, with 35% undiagnosed. Both tools showed good accuracy for detecting dementia, with the 10‐CS slightly more accurate than SPMSQ (area under the curve [AUC] = 0.87 vs 0.85; p = 0.02). After adjustment, cognitive impairment detected by either tool was associated with increased risk of incident delirium, hospital‐associated disability in self‐care activities, and 90‐day mortality, but not prolonged stay. DISCUSSION Brief bedside screening identified dementia and predicted adverse outcomes, supporting its integration into routine hospital care.