Implementation of decision tree and Mamdani fuzzy inference system for Erythropoietin resistance prediction

• Published in: Biomedical signal processing and control 2025-06, Vol.104, p.107496, Article 107496
• Main Authors: Kusumadewi, Sri, Rosita, Linda, Wahyuni, Elyza Gustri, Mulyati, Sri, Arifin, Aridhanyati
• Format: Article
• Language: English
• Subjects: Centroid; Defuzzification; Fuzzy inference system; Imputation; Resistance
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2025.107496

•An early detection of ESA resistance is important.•The J48 algorithm is able to classify data into Responsive and Resistance classes.•Hb, weight, MCHC, and hemodialysis period are factors that influence ESA resistance.•Fuzzy Inference System can predict hemoglobin after hemodialysis.•The Centroid method has been proved to be the best defuzzification method. Erythropoietin Stimulating Agents (ESA) therapy is one way to treat anemia in chronic kidney disease (CKD) patients. Unfortunately, not all patients are responsive to ESA. Some patients show resistance to ESA. Resistance is characterized by low hemoglobin levels after ESA therapy (HbPost). This study aims to develop a knowledge base model for classifying ESA resistance in CKD patients on hemodialysis, as well as a model for predicting patient hemoglobin levels after ESA therapy. We trained 122 data samples containing 17 variables using the J48 classification method. Four features remain after feature selection: HbPre, Weight, MCHCPre, and Period. The classification process generates a decision tree. Furthermore, we build the Fuzzy Inference System (FIS) with the Mamdani method to predict HbPost levels. The centroid, bisector, SOM, MOM, and LOM methods are implemented as defuzzification methods. We tested the performance of FIS using MAPE and RMSE. The results showed that centroid was the most accurate method, with the lowest MAPE of 16.28% and the lowest RMSE of 1.64. On the other hand, LOM was the method with the worst performance because it had the highest MAPE and RMSE of 21.99% and 2.45, respectively. Some of the data used in this study are still incomplete or missing. While we can perform imputation on missing data, our imputation method assumes similar characteristics to the data. Selecting an imputation method that can anticipate uncertainty requires further research.