Predicting Chronic Kidney Disease: A Machine Learning Approach to Early Detection and Risk Assessment

Abstract

Chronic Kidney Disease (CKD) is a major global health concern, often developing silently until reaching advanced stages, which makes early prediction vital for timely medical interven- tion. This research addresses the challenge of predicting CKD onset six months in advance by leveraging both laboratory and clinical data sources. While most existing models either rely on clinical datasets lacking biological markers or laboratory datasets with limited size and availability, our work proposes a hybrid approach to combine the strengths of both. We first trained a Deep Neural Network (DNN) on the UCI laboratory-oriented dataset to detect CKD using biological parameters. This model was then used as a feature extractor in a transfer learning strategy applied to the NHIRD clinical dataset, which contains extensive claims data but lacks laboratory indicators. Our goal was to assess the impact of incorporating learned biological patterns into clinical prediction tasks. The proposed transfer learning-based model demonstrated strong performance, particularly in terms of recall, achieving a true positive rate of 92% for predicting CKD six months before clinical onset. These results confirm the added value of integrating laboratory-derived knowl- edge into large-scale clinical prediction systems, and highlight the feasibility of using such models for real-world healthcare applications, especially in contexts where lab data is scarce.