Semantic Features Prediction for Pulmonary Nodule Diagnosis Based on Online Streaming Feature Selection

Abstract

Early diagnosis significantly improves the survival rate in lung carcinoma patients. This study attempts to construct a predictive network between the computational features and semantic features of pulmonary nodules using online feature selection and causal structure learning. In this paper, we exploit the causal discovery based on the streaming feature algorithm and causal discovery with symmetrical uncertainty based on the streaming feature algorithm. Different from the traditional learning methods that usually obtain all computational features in advance and then select the optimal subset of features from the computational features, the proposed approach integrates online streaming feature selection with causal structure learning. The critical challenges in this integration include: 1) the dynamic selection of computational features and 2) how to evaluate the feature subsets and implement causal structure learning. In addition, considering that building a causal structure network is a time-consuming process, we improve the process by using support vector machines based on the streaming feature algorithm. The experimental results show that our proposed algorithms improve on other traditional feature selection algorithms and ensemble learning algorithm without feature selection with regard to learning accuracy.