PRISM: A REAL-TIME CLINICAL DECISION SUPPORT TOOL FOR PREDICTING HYPERKALEMIA IN CHRONIC KIDNEY DISEASE - TOWARDS SAFER AND EQUITABLE KIDNEY CARE

Hyperkalemia is a frequent and potentially fatal complication of chronic kidney disease (CKD), increasing the risk of arrhythmia, cardiac arrest, and premature death and contributing substantially to emergency department visits and hospital admissions. Fear of hyperkalemia frequently limits the initiation and titration of RAAS-modifying therapies (ACE inhibitors, ARBs, MRAs, SGLT2 inhibitors) that are essential to slow CKD progression and improve cardiovascular outcomes. Detection often depends on intermittent lab testing, which can delay treatment and worsen outcomes. Patients in rural or resource-constrained settings face additional barriers to timely monitoring, amidst deepening inequities in kidney care. There is a pressing need for proactive, real-time risk assessment to guide therapy, reduce avoidable emergency visits, and optimize healthcare resource utilization  

Predictive tools that analyze routine clinical data from electronic health records (EHRs) can help identify high-risk patients before potassium levels become critical. We present a real-time Machine Learning (ML)–based prediction tool to assist clinical decision-making in treating CKD. By enabling clinicians to explore “what-if” scenarios, this tool is designed to enable early/targeted interventions, improve patient safety, reduce hospitalizations, optimize resource use, and lessen the burden on healthcare systems.

PRISM (Predictive Risk Indicator for Serum potassium Measurement) was designed to provide real-time risk prediction of abnormal serum potassium levels using demographic, clinical, and laboratory data. The tool uses a backend tree-based ML predictive engine which was trained on a 10-year cohort of patients with non-dialysis dependent CKD from British Columbia (BC), Canada (2013–2022), encompassing over 44,000 patients. Multiple machine learning algorithms—including Linear Regression, Neural Networks, Support Vector Machines, Random Trees (RT), and XGBoost were trained and evaluated using accuracy, Mean Absolute Error (MAE), and Root Mean Square Error (RMSE). Among these, RT demonstrated the best performance (93.7% accuracy, RMSE = 0.31) and was integrated into the PRISM tool. The intuitive and user-friendly interface developed using Python allows clinicians to input physiological and biochemical characteristics of the patient to predict serum potassium levels in real-time. The tool is integrated with a backend database to pre-populate patient’s demographics, basic vitals, comorbidities, medications, and lab results. 

Predicted potassium levels are displayed in real-time alongside a traffic light–style indicator to visually convey clinical urgency: green for safe, yellow for elevated, and red for critically high potassium. The tool also allows clinicians to adjust variable values to explore “what-if” scenarios, enabling them to simulate potential changes in patient conditions. It helps healthcare providers to better understand the impact of specific risk factors on predicted potassium levels, supporting more informed clinical decision-making and proactive management. Additional features include a reset button to restore default inputs, and a slider for adjusting the serum potassium threshold value, with default thresholds defined in consultation with CKD clinicians.

The developed PRISM tool has the potential to reduce fatal complications, optimize healthcare resource utilization, support proactive clinical decision-making and enhance equity in kidney care, particularly in rural and resource-constrained settings where timely monitoring is most challenging. Customizable thresholds and scenario-based simulations enhance usability, enabling clinicians to adapt to local practices or patient-specific conditions. Future work will involve validating the tool in diverse patient populations and conducting real-world testing to evaluate its clinical integration, usability, and effect on patient outcomes.