An Ensemble Learning Approach for Pancreatic Tumor Detection

Pancreatic cancer remains one of the most lethal cancers worldwide, primarily due to its asymptomatic nature in early stages and limited availability of reliable diagnostic tools. This research addresses this critical gap by proposing a machine learning-based diagnostic system that leverages non-invasive urinary biomarker data to predict the risk of pancreatic tumors. The study utilizes biomarkers such as LYVE1, REG1B, TFF1, and CA19-9, which have been clinically correlated with tumor development. To ensure robust and accurate predictions, we implemented an ensemble learning approach combining three powerful models: Random Forest, XGBOOST, and CATBOOST. The models were trained and evaluated using an 80-20 split on a cleaned version of the Debernardi et al., 2020 dataset. The CATBOOST classifier achieved the highest individual accuracy (~94.07%), while the ensemble model yielded a balanced F1-score of ~91.5%, demonstrating its effectiveness in reducing false negatives—a critical requirement for medical screening systems. The complete solution was deployed as a web-based application using Flask, with user input forms connected to the trained model. The deployment was made publicly accessible via the Render cloud platform, making the tool easily usable by researchers or clinicians. This report presents the design, implementation, results, and deployment strategy of the system, and discusses its potential use in early pancreatic cancer detection scenarios. Future work includes extending the dataset, integrating real-time clinical inputs, and enhancing model explainability for medical adoption.