Leveraging AI to Classify Kidney Disease Conditions Using Clinical Notes

Mohan Timilsina, Samuele Buosi, Edward Curry and Conor Judge – University of Galway

We applied deep learning models to classify clinical notes for conditions that contribute to chronic kidney disease (CKD), focusing on IgA nephropathy, diabetes
and hypertension. This project used clinical data from Ireland’s Health Service Executive (HSE), specifically from the Kidney Disease Clinical Patient Management System (KDCPMS), which records data on participants with kidney failure. By implementing natural language processing through a BERT model architecture, we aimed to streamline the extraction of diagnosis from unstructured clinical notes. The classifier for IgA nephropathy achieved a remarkable 99% precision and 100% recall, demonstrating high accuracy in identifying positive cases with minimal errors. For diabetes, the classifier had a precision of 75% but encountered a higher false negative rate of 25%, while the hypertension classifier achieved balanced performance with 70% precision and a robust F1 score, indicating effective identification of both true positive and negative cases.
Figure 1 in the study illustrates the end-to-end pipeline, detailing stages from data pre-processing to model output and steps involved in this clinical data classification.

While the models for diabetes and hypertension showed varied levels of success, these findings underscore the potential of AI in healthcare for automating data classification and supporting clinician decision-making. The research process included extensive data anonymisation, physician annotation, and tailored pre-processing to optimise model
performance, balancing computational efficiency with accuracy. As AI continues to advance in healthcare, further refinements to these models could enhance their clinical
applicability, particularly for large-scale patient data analysis.