Dealing with Missing Values for Effective Prediction of NPC Recurrence

This paper aims to investigate missing data techniques for effective prediction of nasopharyngeal carcinoma (NPC) recurrence. The techniques include listwise deletion, imputations by mean, k-nearest neighbor, and expectation maximization. The completed data are used to predict the presence or absence of NPC recurrence in each year by means of logistic regression, multilayer perceptron with backpropagation training, and naive bayes. Five year predictions are carried out. Validity of each predictive model is assured by 10-fold cross validation. Their results are compared in order to determine proper missing data treatment and the most efficient prediction technique. The results showed that EM imputation was superior to the other missing data techniques because it can be efficiently applied to all predictive models. The multilayer perceptron with backpropagation training gave the highest prediction performance and it was the most robust to the data completed by different missing data techniques.