Toward Early Detection of Pancreatic Cancer: A Machine Learning Study on Urinary and Plasma Biomarker Data

Abstract—Pancreatic ductal adenocarcinoma (PDAC) is usually diagnosed late, and the only widely used blood marker, CA19-9, is neither sensitive nor specific enough for screening. This paper asks whether a machine learning model that combines four urinary proteins (LYVE1, REG1A, REG1B, TFF1), urinary creatinine, plasma CA19-9, age and sex can separate PDAC from non-PDAC samples (healthy controls and benign hepatobiliary disease) in the public Debernardi et al. dataset of 590 participants (199 PDAC, 391 non-PDAC). The pipeline combines median imputation with missingness indicators, three engineered features, an 80/20 stratified split, SMOTE oversampling of the training data, and a comparison of six classifiers by five-fold cross-validated ROC-AUC. Random Forest performed best (CV ROC-AUC 0.967 ± 0.014) and was tuned by grid search. On a held-out test set of 118 participants the tuned model reached accuracy 0.924, precision 0.878, recall 0.900, F1-score 0.889, ROC-AUC 0.967 and average precision 0.948, detecting 36 of 40 PDAC cases with 5 false positives. Plasma CA19-9, urinary LYVE1 and a CA19-9/creatinine ratio had the highest feature importance. These results come from a single split of a single-source dataset, were not stratified by tumor stage, and were not compared with a CA199-only model; they are therefore a proof of concept, not evidence of early-stage screening performance. The limitations and the validation still required are discussed. Keywords—pancreatic cancer, early detection, machine learning, urinary biomarkers, CA19-9, random forest, class imbalance, ROCAUC