A recursive embedding and clustering technique for unraveling asymptomatic kidney disease using laboratory data and machine learning.

Abstract

Traditional methods for diagnosing chronic kidney disease (CKD) via laboratory data may not be capable of identifying early kidney disease. Kidney biopsy is unsuitable for regular screening, and imaging tests are costly and time-consuming. Several studies have implemented artificial intelligence (AI) to detect CKD. However, these studies used small datasets, had overfitting problems, lacked generalizability, or used complex algorithms that may require additional computational resources. In this study, we collected and analyzed center-based data and used a recursive embedding and clustering technique to reduce their dimensionality. We identified three clusters from 1600 records. We focused on the second cluster, as most of the characteristics had values in the normal ranges. Normal range values for most indicators generally represent stable kidney function with minor signs of strain, which often remain asymptomatic. Creatinine and eGFR levels within the threshold ranges indicate early kidney stress without filtration issues, which require close monitoring. The gradient-boosting algorithm showed superior performance among all algorithms in detecting these clusters. We evaluated an additional 400 unlabeled records to validate our method. This research can help clinicians automatically detect initial signs in numerous patients via routine tests to prevent the consequences of late-stage CKD detection.