Prediction Model for Pancreatic Cancer-A Population-Based Study from NHIRD

Area of Science:

• Oncology
• Medical Informatics
• Machine Learning

Background:

• Pancreatic cancer is a leading cause of death in Taiwan with a low survival rate.
• Early detection is challenging due to a lack of diagnostic tools, making screening crucial.
• Few predictive models exist for pancreatic cancer, highlighting a need for improved early detection strategies.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting pancreatic cancer.
• To compare the performance of logistic regression, deep neural networks, ensemble learning, and voting ensemble models.
• To create a more accurate predictive tool for early pancreatic cancer screening.

Main Methods:

• Utilized the Taiwan Health Insurance Database, representing over 99% of the population.
• Applied machine learning techniques, including logistic regression, deep neural networks, and ensemble methods.
• Evaluated model accuracy using ROC curves and confusion matrices.

Main Results:

• The logistic regression model showed superior performance in the external testing set across different factor combinations.
• The stacking model achieved the best sensitivity, while the deep neural network model demonstrated the highest specificity.
• The developed models achieved an AUC range of 0.71 to 0.76, surpassing previous studies (0.57-0.71).

Conclusions:

• A novel pancreatic cancer prediction model was successfully developed with enhanced accuracy.
• The model outperforms existing predictive tools, offering a simpler method for early screening.
• Improved early detection via this model can increase patient awareness and facilitate timely treatment during the curable stage.