Abstract
BACKGROUND: Data loss often occurs in the collection of clinical data. Directly discarding the incomplete sample may lead to low accuracy of medical diagnosis. A suitable data imputation method can help researchers make better use of valuable medical data. METHODS: In this paper, five popular imputation methods including mean imputation, expectation-maximization (EM) imputation, K-nearest neighbors (KNN) imputation, denoising autoencoders (DAE) and generative adversarial imputation nets (GAIN) are employed on an incomplete clinical data with 28,274 cases for vaginal prolapse prediction. A comprehensive comparison study for the performance of these methods has been conducted through certain classification criteria. It is shown that the prediction accuracy can be greatly improved by using the imputed data, especially by GAIN. To find out the important risk factors to this disease among a large number of candidate features, three variable selection methods: the least absolute shrinkage and selection operator (LASSO), the smoothly clipped absolute deviation (SCAD) and the broken adaptive ridge (BAR) are implemented in logistic regression for feature selection on the imputed datasets. In pursuit of our primary objective, which is accurate diagnosis, we employed diagnostic accuracy (classification accuracy) as a pivotal metric to assess both imputation and feature selection techniques. This assessment encompassed seven classifiers (logistic regression (LR) classifier, random forest (RF) classifier, support machine classifier (SVC), extreme gradient boosting (XGBoost) , LASSO classifier, SCAD classifier and Elastic Net classifier)enhancing the comprehensiveness of our evaluation. RESULTS: The proposed framework imputation-variable selection-prediction is quite suitable to the collected vaginal prolapse datasets. It is observed that the original dataset is well imputed by GAIN first, and then 9 most significant features were selected using BAR from the original 67 features in GAIN imputed dataset, with only negligible loss in model prediction. BAR is superior to the other two variable selection methods in our tests. CONCLUDES: Overall, combining the imputation, classification and variable selection, we achieve good interpretability while maintaining high accuracy in computer-aided medical diagnosis.