Predicting Chemotherapy Response in Patients With Advanced or Metastatic Pancreatic Cancer Using Machine Learning.
Selecting an optimal first-line chemotherapy regimen for advanced or metastatic pancreatic cancer is challenging because of varying efficacy and toxicity profiles of fluorouracil, leucovorin, irinotecan, and oxaliplatin (FOLFIRINOX) and gemcitabine/nab-paclitaxel (GnP). This study aimed to develop machine learning (ML) models that predict survival outcomes and guide treatment selection using routinely available clinical data.
We retrospectively analyzed 191 patients who received systemic chemotherapy for advanced or metastatic pancreatic cancer at Gangneung Asan Hospital and the Asan Medical Center between 2014 and 2023. Seventeen demographic and clinical variables, along with survival outcomes, were collected. The data set was stratified and split into training and test sets (4:1). CatBoost-based ML models were trained to predict 12-month overall survival (OS) for each regimen. A minimal subset of variables was selected using 5-fold cross-validation to optimize receiver operating characteristic (ROC)-AUC. Patients were classified as high or low risk based on model-derived thresholds.
The median age of the cohort was 62 years, and 64% was male. The ML models achieved ROC-AUCs of 0.81 for FOLFIRINOX and 0.82 for GnP. Predictive accuracies on test data were 0.77 and 0.80, respectively. Median OS differed significantly between predicted high- and low-risk groups: 9 v 15 months for FOLFIRINOX (hazard ratio [HR], 2.8; P < .0001) and 9 v 18 months for GnP (HR, 2.5; P < .01). In addition, 27% of patients predicted to be high risk for FOLFIRINOX were classified as low risk for GnP, and 32% vice versa.
ML models trained on multicenter data can effectively predict early mortality risk and help personalize chemotherapy selection in advanced or metastatic pancreatic cancer, potentially improving clinical outcomes.
We retrospectively analyzed 191 patients who received systemic chemotherapy for advanced or metastatic pancreatic cancer at Gangneung Asan Hospital and the Asan Medical Center between 2014 and 2023. Seventeen demographic and clinical variables, along with survival outcomes, were collected. The data set was stratified and split into training and test sets (4:1). CatBoost-based ML models were trained to predict 12-month overall survival (OS) for each regimen. A minimal subset of variables was selected using 5-fold cross-validation to optimize receiver operating characteristic (ROC)-AUC. Patients were classified as high or low risk based on model-derived thresholds.
The median age of the cohort was 62 years, and 64% was male. The ML models achieved ROC-AUCs of 0.81 for FOLFIRINOX and 0.82 for GnP. Predictive accuracies on test data were 0.77 and 0.80, respectively. Median OS differed significantly between predicted high- and low-risk groups: 9 v 15 months for FOLFIRINOX (hazard ratio [HR], 2.8; P < .0001) and 9 v 18 months for GnP (HR, 2.5; P < .01). In addition, 27% of patients predicted to be high risk for FOLFIRINOX were classified as low risk for GnP, and 32% vice versa.
ML models trained on multicenter data can effectively predict early mortality risk and help personalize chemotherapy selection in advanced or metastatic pancreatic cancer, potentially improving clinical outcomes.
Authors
Koo Koo, Choi Choi, Cheon Cheon, Yoo Yoo, Courcoubetis Courcoubetis, Ryoo Ryoo, Kim Kim, Chang Chang, Oh Oh, Lim Lim, Kim Kim
View on Pubmed