Chemotherapy selection for advanced or metastatic pancreatic cancer using machine learning models trained with multi-center datasets.

738 Background: Several factors must be considered when selecting the chemotherapeutic regimen for advanced or metastatic pancreatic cancer. The decision between FOLFIRINOX and Gemcitabine/Nab-paclitaxel (GnP) as the first-line therapy is challenging, as survival is influenced by the efficacy and toxicity profiles of these treatments, along with individual patient characteristics and therapeutic vulnerabilities. This study aims to facilitate the selection of an appropriate first-line regimen for advanced or metastatic pancreatic cancer by employing machine learning (ML) methods trained on multi-center datasets. Methods: Our study is based on a cohort of 191 patients who underwent systemic chemotherapy for advanced or metastatic pancreatic cancer at the Gangneung Asan Hospital and Asan Medical Center in South Korea between 2014 and 2023. The dataset consisted of 17 types of demographic and clinical characteristics, as well as time-course of response and survival outcomes. We divided the cohort into training and test groups (4:1) in a stratified manner. ML models predicting overall survival (OS) were trained on the former group using the CatBoost method. The models utilizing a minimal subset of variables were selected with respect to ROC-AUC during 5-fold cross validation. A patient was classified as having high risk to a therapy when the predicted probability of death following treatment initiation was above the threshold value as determined during training of the respective model. Results: The median age of the entire cohort was 62 years, with 64% being male. The ML models achieved ROC-AUCs of 0.81 and 0.82 when predicting early death within 12 months following the initial administration of FOLFIRINOX or GnP, respectively. The predictive accuracies of the models were 0.77 and 0.80 for the unseen datasets from the two treatment groups, respectively. The median OS of the high- versus low-risk groups of FOLFIRINOX predicted by the ML models were significantly different (9 vs 15 months, P <0.0001), with a hazard ratio (HR) of 2.8. Similarly, the median OS of the high- and low-risk groups of GnP were significantly different (9 vs 18 months, P <0.01, HR of 2.5). We observed that 27% of the patients predicted as high risk to FOLFIRINOX therapy were predicted as low risk to GnP and vice versa for 32% of the patients predicted as high risk to GnP therapy. Conclusions: We developed ML models to compute the probability of early death following FOLFIRINOX or GnP therapy from routinely collected data of the patients with advanced or metastatic pancreatic cancer. After confirming robust predictive performance with multi-center datasets, we believe that these models may assist clinicians in selecting a first-line regimens, potentially enhancing personalized treatment strategies for this challenging disease.