The Use of Unsupervised Machine Learning in the Prediction of Chronic Kidney Disease

Description

Page: 106-115

Fosu Acheampong¹, Johannie Twum Amponsem,² and Humphrey Danso Bamfo³ (Department of Mathematical Sciences, Ball State University, USA^1,2 and Department of Psychology, University of Cape Coast, Ghana³)

Early identification is essential to slowing the progression of Chronic Kidney Disease (CKD) and improving outcomes, especially in areas with limited access to specialized medical treatment. The potential of unsupervised approaches is still poorly understood, although supervised machine learning has been extensively investigated for the categorization of CKD. To fill this vacuum, this study uses real-world clinical data to assess how well unsupervised models: K-Means, DBSCAN, Isolation Forest, and Principal Component Analysis (PCA) help to distinguish CKD cases from non-CKD cases. The study uses strong feature selection approaches, including ANOVA, Chi-square tests, Pearson correlation, and embedding methods like Random Forest and Lasso regression to maximize model performance. Using extensive internal (Davies-Bouldin Index, Silhouette Score) and external (Precision, Recall, F1 score, Adjusted Rand Index) validation, the results show that K-Means with feature reduction outperforms other techniques with an accuracy of 95%. Eight clinically significant variables make up the chosen feature set, which is in line with the pathophysiology of CKD and improves interpretability for real-world. These results demonstrate the feasibility of unsupervised learning as a label-free, scalable approach to early CKD screening, especially in environments with limited resources. For wider clinical usefulness, next efforts include merging multimodal data and expanding the approach to stage-specific detection.