Investigation of the effect of data preprocessing on ensemble learning classification algorithms in disease diagnosis data

In the field of health, while classifying for identification and examination of disease, supervised machine learning methods are used, which are able to compute using artificial intelligence technology, in order to extract meaningful information from complex data. Ensemble learning methods enable establishment of more successful models by training multiple learners at the same time to solve same problem. In this study, it is aimed to compare performance of classification algorithms after data preprocessing to problems such as missing values, class noise and class imbalance that may be encountered in data sets used to diagnose accurate disease in health data. In the study, data collected from KEEL database were used to diagnose diseases such as heart disease, thyroid, hepatitis, lymphedema, breast cancer and diabetes. In order to make classification, while random forest and weighted subspace random forest were used as bagging algorithms; additive logistic regression and gradient boosted machines algorithms were used as boosting algorithms. Accuracy, sensitivity, specificity, precision, Kappa statistic, Youden index, F - measure and ROC measurement metrics were used to compare performance of algorithms. At the same time, run times of algorithms were calculated. All statistical analyzes were performed with RStudio 1.2.1335 - Windows 7+ (64-bit) program. When performances of algorithms were compared for original data and processed data, it was seen that performance success of algorithms increased after data preprocessing. In general, performance of boosting algorithms yielded higher results than bagging algorithms. When algorithms were compared in terms of run time, boosting algorithms were the longest running algorithms. As a result, data preprocessing should not be overlooked if research is aimed at high performance success. In data preprocessing, simulation studies can be performed by adding different topics such as tuning parameters and selecting variables. Keywords: Disease Diagnosis; Data Preprocessing; Missing Values; Class Noise; Class Imbalance; Ensemble Learning