Predicting chronic kidney disease progression using small pathology datasets and explainable machine learning models

Abstract

AbstractChronic kidney disease (CKD) poses a major global public health burden, with approximately 7 million affected. Early identification of those in whom disease is likely to progress enables timely therapeutic interventions to delay advancement to kidney failure. This study developed explainable machine learning models leveraging pathology data to accurately predict CKD trajectory, targeting improved prognostic capability even in early stages using limited datasets. Key variables used in this study include age, gender, most recent estimated glomerular filtration rate (eGFR), mean eGFR, and eGFR slope over time prior to incidence of kidney failure. Supervised classification modelling techniques included decision tree and random forest algorithms selected for interpretability. Internal validation on an Australian tertiary centre cohort (n=706; 353 with kidney failure and 353 without) achieved exceptional predictive accuracy, with the area under the receiver operating characteristic curve (ROC-AUC) reaching 0.94 and 0.98 on the binary task of predicting kidney failure for decision tree and random forest, respectively. To address the inherent class imbalance, centroid-cluster-based under-sampling was applied to the Australian dataset. To externally validate the performance of the model, we applied the model to a dataset (n=597 adults) sourced from a Japanese CKD registry. To overcome risks of overfitting on small sample sizes, transfer learning was subsequently employed by fine-tuned machine learning models on 15% of the external dataset (n=89) before evaluating the remaining 508 patients. This external validation demonstrated performant results with an ROC-AUC of 0.88 for the decision tree and 0.93 for the random forest model. Decision tree model analysis revealed the most recent eGFR and eGFR slope as the most informative variables for prediction in the Japanese cohort, aligning with the underlying pathophysiology. The research highlights the utility of deploying explainable machine learning techniques to forecast CKD trajectory even in the early stages utilising limited real-world datasets.