Research on early risk predictive model and discriminative feature selection of cancer based on real-world routine physical examination data

摘要

Most cancers at early stages show no obvious symptoms and curative treatment is not an option any more when cancer is diagnosed. Therefore, making accurate predictions for the risk of early cancer has become urgently necessary in the field of medicine. In this paper, our purpose is to fully utilize real-world routine physical examination data to analyze the most discriminative features of cancer based on ReliefF algorithm and generate early risk predictive model of cancer taking advantage of three machine learning (ML) algorithms. We use physical examination data with a return visit followed 1 month later derived from CiMing Health Checkup Center. The ReliefF algorithm selects the top 30 features written as Sub(30) based on weight value from our data collections consisting of 34 features and 2300 candidates. The 4-layer (2 hidden layers) deep neutral network (DNN) based on B-P algorithm, the support machine vector with the linear kernel and decision tree CART are proposed for predicting the risk of cancer by 5-fold cross validation. We implement these criteria such as predictive accuracy, AUC-ROC, sensitivity and specificity to identify the discriminative ability of three proposed method for cancer. The results show that compared with the other two methods, SVM obtains higher AUC and specificity of 0.926 and 95.27%, respectively. The superior predictive accuracy (86%) is achieved by DNN. Moreover, the fuzzy interval of threshold in DNN is proposed and the sensitivity, specificity and accuracy of DNN is 90.20%, 94.22% and 93.22%, respectively, using the revised threshold interval. The research indicates that the application of ML methods together with risk feature selection based on real-world routine physical examination data is meaningful and promising in the area of cancer prediction.