Improving Lung Cancer Relapse Prediction Using the Developed Optuna_XGB Classification Model

Abstract

Lung cancer is more likely to relapse in the first five years following surgery; even though the operation may have been a complete success, there remains a chance that the lung cancer could return. This return may lead the patient to die after a successful surgery. Because there are no symptoms of lung cancer in its early stage, many researchers use intelligent systems to predict the relapse of lung cancer in its early stages. The outcome of previous works considering this issue still suffers from low prediction accuracy. This study proposed a method to predict lung cancer relapse more accurately. This method has multiple stages: 1st optimization system, feature selection stage, 2nd optimization stage, and extreme gradient boost (XGBoost) classifications stage. It used two datasets (GSE8894 and GSE68465) of a gene expression microarray for NSCLC with its clinical information on relapse state. We obtained three probes (3 genes) with clinical data combinations that can get good prediction results. These genes included 225389_at (BTBD6), 220239_at (KLHL7), and 204832_s_at (BMPR1A). A comparison between the proposed model and the original XGBoost with PSO and Hyperopt as hyperparameter optimization for the XGBoost classification model is performed. Extensive comparisons with four machine learning algorithms, including Deep Forest, K-nearest neighbor (KNN), Support Vector Machine (SVM), and Naive Bayes, are conducted. The proposed model accuracies are 0.93 for the GSE8894 dataset and 0.81 for the GSE68465 dataset.