Using Feature Selection and Ensemble Algorithms for Sophisticated Pancreatic Cancer Diagnosis

Overview & Synopsis

Pancreatic cancer (PC) is the fourth leading cause of cancer death in the United States because it's five-year survival rate has stayed constant at 8.9%. This occurs because of late diagnosis which is affiliated with the hidden location of the pancreas, making current screening methods unavailable. PC is also often misdiagnosed as colon, esophageal, and biliary tract cancer, showing how genetic biomarkers specific to PC are needed. Currently, there have only been four studies aiming to diagnose PC through genetic biomarkers, each achieving accuracy of 84% and AUC of 70%. To address the problems of frequent late diagnosis and misdiagnosis, machine learning models were used to diagnose PC through the expression of significant miRNAs. Feature selection was performed using statistical analysis, recursive feature elimination, feature importance using decision trees, and the CyberT bioinformatics tool to find 155 significant miRNAs out of the 2,500 features. Four machine learning algorithms (a logistic regression, adaboost, neural network, and ensemble model) were then tuned and received an AUC of 92% and total accuracy of 94%. Further, the most important miRNAs out of the 155 were found through LIME and SHAP interpretation and were found to be significant in key PC processes. This research is important as it is the first to assess both misdiagnosis and early diagnosis of PC and beat the state of the art's accuracy by 10% and AUC by 20%. These algorithms need to be improved before clinical trial by adding more diverse data and developing a user interface.