Salehinejad, Hojjat

Abstract Background: The standard first line therapy for cholangiocarcinoma (CCA), gemcitabine and cisplatin +/- immunotherapy, only induces disease control in approximately 75% of patients (KEYNOTE-966 Trial). While several candidate single genes that predict gemcitabine and cisplatin resistance have been proposed, the overall molecular landscape that leads to resistance is not well understood. Predicting response based on molecular characteristics would benefit the clinical decision-making process when selecting appropriate chemotherapy regimens for CCA patients. Methods: Fifty-five Mayo Clinic CCA patients who underwent whole transcriptome RNA sequencing were included. Forty-eight patients had TempusXR sequencing and 7 patients used institutional labs. Normalized transcript-level data was converted to KEGG pathway-level data using overrepresentation ranking. Patient oncology notes were manually reviewed for response to gemcitabine plus cisplatin therapy, with or without concurrent pembrolizumab or durvalumab. Oncologists used a mixture of formal RECIST criteria and expert opinion to determine whether the response to gemcitabine and cisplatin +/- immunotherapy was progressive disease (PD), stable disease (SD), partial response (PR), or complete response (CR). The binary outcome of PD vs SD/PR/CR based on KEGG pathway overrepresentation was modelled using a Tabular Prior-data Fitted Network (TabPFN). Validation was performed using leave-one-out cross-validation. External model testing was performed on RNA sequencing data from 11 patient-derived xenografts that were treated with gemcitabine and cisplatin and underwent ultrasound measurements of tumor size. Results: There were 34 patients with SD/PR/CR and 21 patients with PD as the best response while receiving gemcitabine and cisplatin +/- immunotherapy. TabPFN cross-validation accuracy was 0.73, recall 0.94, precision 0.71, and F1 score 0.81. External validation accuracy was 0.70, with one xenograft exhibiting mixed response. In contrast, the best trained model using a 14-gene subset previously published for gemcitabine resistance prediction in pancreatic adenocarcinoma had accuracy 0.58. Conclusions: Machine learning models like TabPFN can capture relationships in high-dimensional transcriptomic information in novel ways that enhance our ability to predict chemotherapy response. Integrating machine learning to understand the complex molecular landscape of CCA will be a valuable tool for therapeutic regimen selection. Citation Format: Ellen L. Larson, Erik Jessen, Dong-Gi Mun, Amro Abdelrahman, Jennifer Tomlinson, Danielle Carlson, Hojjat Salehinejad, Caitlin Conboy, Rory Smoot. Machine learning models of RNA expression landscapes help predict overall tumor response to chemotherapy in cholangiocarcinoma [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr A020.

Objective: To develop an artificial intelligence–based pipeline to assess and triage patient-submitted postoperative wound images. Background: The rise of outpatient surgeries, remote monitoring, and patient-submitted wound images via online portals has contributed to a growing administrative workload on clinicians. Early identification of surgical site infection (SSI) is essential for reducing postoperative morbidity. Methods: Patients ≥18 years old who underwent surgery at 9 affiliated Mayo Clinic hospitals (2019-2022) and were captured by the National Surgical Quality Improvement Program (NSQIP) were included. Eligibility required submission of one image via the patient portal within 30 days after surgery. Images were independently screened in duplicate to determine the presence of an incision. SSI outcomes were obtained from NSQIP. The developed model consisted of 2 stages: incision detection and SSI detection in images with incisions. Four pretrained architectures were evaluated using 10-fold cross-validation, with upsampling and data augmentation to mitigate class imbalance. An end-to-end pipeline, image quality assessment and sensitivity analysis stratified by race were also performed. Results: Among 6060 patients, the median age was 54 years (interquartile range: 40–65), 61.4% (n=3805) were female, and 92.5% (n=5731) identified as White. SSIs were confirmed in 6.2% (n=386) images. Vision Transformer outperformed all others, achieving an incision detection accuracy of 0.94 (area under the curve=0.98) and an SSI detection accuracy of 0.73 (area under the curve=0.81). In addition, it demonstrated strong performance in assessing image quality. Sensitivity analysis revealed comparable performance across racial subgroups. Conclusion: This artificial intelligence pipeline demonstrates promising performance in automating wound image assessment and SSI detection, reducing clinical workload and improving postoperative care.

Background: Cholangiocarcinoma is a molecularly heterogenous cancer arising from the biliary epithelium. Individual omic techniques have identified clinically relevant, targetable subtypes but most tumors have no targetable alterations based on exome or transcriptomic evaluation. Integrated multiomics approaches incorporating transcriptomic, proteomic, and phosphoproteomic characterization provide deeper understanding and identify non-mutated, activated pathways that can be targeted. To optimize the potential clinical utility of these approaches, minimal classifier features must be defined. Methods: Approximately 60,000 RNA, protein, and phosphoprotein features were extracted from 35 cholangiocarcinomas treated at the primary US study site. 206 patients treated and profiled at an institution in China were also included. Using the Multiomics Factor Analysis clustering method, patients were sorted into 3 distinct molecular profiles. Important molecular features were selected using competing feature selection (FS) methods: Boruta, RreliefF, variable selection using random forests (RF), recursive factor elimination (RFE), median decrease accuracy in RF, and Concrete Autoencoders. Machine learning classification prediction models including Extreme Gradient Boosting, gradient-boosted trees, multinomial logit, and support vector machine (SVM) were tested on each FS. Each combination of FS and predictive model was tested with 10 replicates of 10-fold cross-validation of a randomly-selected balanced test set and an oversampled, balanced training subset. Accuracy, F1 score, AUC, precision, and recall were assessed for each FS and predictive model combination. A custom function of physical assay cost and classification accuracy was optimized to find the best assay. KEGG pathway overrepresentation analysis was used to analyze feature subsets. Results: A consistent set of 50 important RNA and protein features was identified, including previously reported prognostic factors, like SpryD4 protein expression, as well as previously unreported pathway surrogates, such as Slc2A2 or Pipox protein. Pathway overrepresentation analysis found insulin signaling and resistance and amino acid biosynthesis/metabolism pathways were overrepresented in the discriminative feature set. The gradient-boosted trees algorithm combined with the RFE FS method had the best performance, with a multiclass macro average (±sd) cross-validation accuracy 89% (±6%), F1 score 89% (±8%), AUC 0.98 (±0.02), precision 90% (±10%), recall 89% (±6%). Conclusions: Feature selection methods and predictive machine learning models work synchronously to develop accurate molecular subtype assays based on just 50 protein and RNA levels. This strategy has the potential to make molecular subtyping for cholangiocarcinoma and other molecularly complex cancers fast and widely available. Citation Format: Ellen Larson, Erik Jessen, Dong-Gi Mun, Jennifer Tomlinson, Amro Abdelrahman, Danielle Carlson, Hojjat Salehinejad, Rory Smoot. Feature selection and machine learning strategies optimize an affordable molecular assay for cholangiocarcinoma subtype [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 1107.

Abstract Objectives In the general hospital wards, machine learning (ML)-based early warning systems (EWSs) can identify patients at risk of deterioration to facilitate rescue interventions. We assess subpopulation performance of a ML-based EWS on medical and surgical adult patients admitted to general hospital wards. Materials and Methods We assessed the scores of an EWS integrated into the electronic health record and calculated every 15 minutes to predict a composite adverse event (AE): all-cause mortality, transfer to intensive care, cardiac arrest, or rapid response team evaluation. The distributions of the First Score 3 hours after admission, the Highest Score at any time during the hospitalization, and the Last Score just before an AE or dismissal without an AE were calculated. The Last Score was used to calculate the area under the receiver operating characteristic curve (ROC-AUC) and the precision-recall curve (PRC-AUC). Results From August 23, 2021 to March 31, 2022, 35 937 medical admissions had 2173 (6.05%) AE compared to 25 214 surgical admissions with 4984 (19.77%) AE. Medical and surgical admissions had significant different (P <.001) distributions of the First Score, Highest Score, and Last Score among those with an AE and without an AE. The model performed better in the medical group when compared to the surgical group, ROC-AUC 0.869 versus 0.677, and RPC-AUC 0.988 versus 0.878, respectively. Discussion Heterogeneity of medical and surgical patients can significantly impact the performance of a ML-based EWS, changing the model validity and clinical discernment. Conclusions Characterization of the target patient subpopulations has clinical implications and should be considered when developing models to be used in general hospital wards.