Predicting potassium trajectories for risk monitoring in outpatients with heart failure, diabetes mellitus or chronic kidney disease.

Aim: To develop a dynamic prediction model for potassium concentration in the outpatient sector for patients with heart failure (HF), chronic kidney disease (CKD) and/or diabetes mellitus (DM).

Methods: We used administrative claims data from Scotland collected at the Tayside Health Informatics Centre and selected patients between 1 January and 30 June 2020 with underlying conditions of HF, CKD and/or DM. The follow-up time of each patient was divided into assessment periods to predict a patient's maximum potassium value within the next 4 weeks (prediction periods). Three linear mixed-effect models were fitted and model performance was assessed using root-mean-squared-error (RMSE), mean absolute error (MAE) and mean squared error (MSE).

Results: Among 5918 patients with a mean age of 76.2 years, a median of 17.0 potassium concentrations were measured per patient corresponding with 1.71 measurements per assessment period. In total, we predicted 5478 maximum potassium values. The final model performed with a RMSE of 0.52, MAE of 0.39, MSE of 0.27 and no apparent trends in the residuals over time. Prediction was more accurate within the potassium reference range and tended to underestimate extremely high and overestimate low observations. Among the strongest predictors were newly acquired acute kidney injury, last measured potassium and use of low ceiling and high ceiling diuretics.

Conclusion: We propose a blueprint of a decision support tool which predicts potassium concentration longitudinally by updating the predictions based on accumulating data. Our findings demonstrate that dynamically reassessing predictors can aid in estimating potassium levels over multiple months with reasonable accuracy in the outpatient setting.