USING MACHINE LEARNING FOR PREDICTING SERUM POTASSIUM IN PATIENTS WITH CHRONIC KIDNEY DISEASE: A POPULATION LEVEL STUDY FROM BRITISH COLUMBIA, CANADA

Hyperkalemia is one of the most common, potentially fatal and costly complications in chronic kidney disease (CKD). Despite clear awareness of its clinical implications, predicting hyperkalemia remains challenging, as it often emerges unexpectedly, leading to emergency visits and hospitalizations. Further, managing CKD-related complications consumes substantial resources due to repeated lab testing, clinical follow-up, medication adjustments, and acute care utilization. Such events adversely impact long-term kidney outcomes and add burden to the existing infrastructure. 

Leveraging Machine learning (ML) to analyze complex clinical data helps identify patterns and predict early indicators of potassium imbalance, enabling proactive interventions. In this study, we applied ML models to a cohort of patients with CKD from British Columbia (BC), Canada to predict hyperkalemia. The goal was to enhance proactive monitoring, prevent emergencies, and support standardization and efficient use of resources in care pathways. 

This study utilized a 10-year dataset (2013–2022) from the Patient Records and Outcome Management Information System (PROMIS) database, the province-wide registry for CKD patients in BC, Canada. The data included over 44,000 patients with non-dialysis dependent CKD and 13 million records with demographic, clinical, laboratory, comorbidity, and medication data. A total of 35 key variables were selected based on literature review, feature selection and consultation with CKD clinicians. ML models, including Linear Regression (LR), Neural Networks (NN), Random Trees (RT), Support Vector Machines (SVM), and XGBoost, were trained using IBM SPSS Modeler to predict serum potassium levels. Model performance was evaluated using accuracy, Mean Absolute Error (M Root Mean Square Error (RMSE). Calibration was assessed using quantile-based calibration curves, comparing predicted versus observed potassium levels across deciles to evaluate systematic under- or overestimation. 

Tree-based models performed best in predicting serum potassium levels among patients with CKD. RT achieved the highest accuracy (93.7%, RMSE = 0.31), closely followed by XGBoost (93.4%, RMSE = 0.33). Both models demonstrated strong calibration across the potassium range, indicating reliable performance in detecting both hypo- and hyperkalemia risks (refer Figure 2). LR and SVM underperformed at extreme values, and NN were less consistent due to data exclusions. These findings suggest that tree-based models, particularly RT, are better suited for real-time clinical applications where interpretability and robustness across the physiological range are critical. The model’s ability to generalize well across different potassium levels supports the primary objective of developing an accurate and practical decision-support tool for proactive hyperkalemia risk prediction in CKD.

ML can accurately predict hyperkalemia in CKD using routinely collected data. Integrating these models into clinical systems could improve patient safety, reduce repeated lab testing, emergency visits, and hospitalizations, and optimize healthcare utilization. Next steps include external validation across diverse populations and real-world testing to assess clinical integration, usability, and impact on outcomes.