Imputing of Missing Values in Diabetes and Breast Cancer Datasets through a Two-Layer Perceptron Neural Network

Imputation of missing values in a medical data set is one of the important challenges in data mining. Therefore, this study was performed with the aim of imputation the missing values of some features of the diabetes and breast cancer datasets.

In this descriptive study, a breast cancer dataset consisting of 699 specimens including 458 benign and 241 malignant specimens, along with a diabetes dataset consisting of 768 specimens including 500 non-diabetic specimens and 268 other specimens with diabetes, were used. For the purpose of the imputation of missing values in these two datasets, a model based on a two-layer perceptron neural network was developed, and for the purpose of assessment, support vector machine (SVM) and t test were used.

The mean squared errors (MSEs) obtained in the two-layer perceptron neural network model, in the diabetes dataset about 0.03 and in the breast cancer dataset about 0.04, were less than the MSEs obtained in the imputation method with the mean value. The values imputed by the model were closer to the actual value than the values imputed with the mean value. Accuracy and sensitivity of disease classification in the case of missing values imputed by the perceptron neural network increased in comparison with the two conventional methods of mean value and the method of deleting missing values, about 2, 4, 2, and 4 percent in the diabetes dataset, and about 1, 3, 2, 5 percent in the dataset breast cancer, respectively. There was a significant difference between the two methods of imputation of missing values with the mean value and imputation by the model.

The imputation of the missing values in the medical data set by the two-layer perceptron neural network showed better results in the classification of the disease than the two methods of imputation with the mean value and the method of deleting missing values.