Predicting and Staging Chronic Kidney Disease using Optimized Random Forest Algorithm

• Published in: 2021 International Conference on Information Systems and Advanced Technologies (ICISAT) pp. 1 - 8
• Main Authors: Samet, Sarra, Laouar, Mohamed Ridda, Bendib, Issam
• Format: Conference Proceeding
• Language: English
• Published: IEEE 27-12-2021
• Subjects: Chronic kidney disease; Data analysis; Data mining; Disease prediction; Feature selection; Kidney; Machine learning; Measurement; Prediction algorithms; Predictive models; Public healthcare; Radio frequency; Random forest; Reliability
• DOI: 10.1109/ICISAT54145.2021.9678441

The silent killer Chronic Kidney Disease (CKD) in wealthy countries and listed with the leading causes of death in impoverished countries. Because of its rising incidence, CKD is included in the most serious public health problems. It is apparent that early detection of CKD may reduce the severity of damage in maturity. The patient must go to a diagnostic facility and consult with a doctor. This significant issue has been solved with the introduction of machine learning. This study's main objective is to build a model that can reliably predict a person's risk of acquiring CKD. Data mining and machine learning techniques have been widely employed for forecasting chronic renal disease, but little research has been done mixing imputation approaches at the pre-processing stage and feature selection strategy so that classification accuracy will be enhanced. The CKD Database, which is used in the experiments and consists of 400 records with 25, is accessible through UCI's machine learning repository. It does, however, have a large number of missing values, which is why we proposed combining several missing data imputation strategies to solve the problem. The chi-square test was used to select features in this work. A supervised machine learning classification model called Random Forest (RF) is utilized and optimized with gridsearch to diagnose CKD at an early stage. Following a cross-validation procedure with 5 folders, several metrics were utilized to evaluate the model. Our RF had a 99.24% accuracy. The model's best result is created by considering the 10 best-selected features. When compared to previous studies, our results are among the best for assessment metrics and the ranking accuracy. However, with only fewer features. In practice, some decision assistance for renal illness' diagnosis, prevention, and prediction are provided by this study.