A Comparison of Machine Learning Algorithms to Predict Cervical Cancer on Imbalanced Data

Abstract

Cervical cancer is a leading cause of death in women. The present research analyzes, explores, compares and identifies the best method for predicting cervical cancer by applying machine learning techniques. The data is from the University Hospital of Caracas, Venezuela where a selection of variables was made according to the literature in order to predict cervical cancer. Seven algorithms were applied: decision tree (DT), random forest (RF), logistic regression (LR), XGBoost (XG), Naive Bayes (NB), multilayer perceptron (MLP) and K-nearest neighbors (KNN). Furthermore, three imbalanced data techniques were applied: SMOTETomek, SMOTE, and ROS for Hinselmann, Schiller, Cytology and Biopsy as target variables. In addition, accuracy, precision, recall, f-score and AUC were used to evaluate the results. Random forest was the algorithm with the highest results in accuracy, precision and f-score, with 94.57%, 72.46% and 60.70% respectively. Logistic regression and Naive Bayes had the highest values for recall and AUC with 68.37% and 79.11% respectively.