[Validation of a hyperkalemia prediction model in chronic kidney disease]

Objective: To validate the accuracy and consistency of a previously established prediction model for the occurrence of hyperkalemia in non-dialytic chronic kidney disease (CKD) patients. Methods: All patients diagnosed with CKD from Outpatient Department of Shanghai Changzheng Hospital during the 4th quarter of 2020 were recruited. Demographic data, clinical characteristics and prediction model-related parameters of the patients were collected and analyzed. Receiver operating characteristic (ROC) curve was drawn to evaluate the effectiveness of the model, and the specificity and sensitivity were calculated based on the cut-off value of 4 obtained from the previous model. The improved Hanley method was used to compare the area under the curve (AUC) between the previously established model and current validation dataset. The calibration curve was drawn to verify the model calibration degree. Results: A total of 434 patients diagnosed with non-dialytic CKD were enrolled, among whom 233 were males and 201 were females, with an average age of (55±16) years. According to the measured serum potassium values, the prevalence of hyperkalemia was 7.6%. And 33 patients were allocated to the hyperkalemia group and 401 patients were to the normal potassium group. There was no significant difference in age and sex between the two groups (both P>0.05). A combination of hyperkalemia and heart failure (27.3% vs 3.7%, P<0.001), diabetes (42.4% vs 19.7%, P=0.002), and acidosis (51.5% vs 7.0%, P<0.001) were more frequently in the hyperkalemia group, compared with the normal serum potassium group. Patients in the hyperkalemia group were more likely to have a past history of serum potassium ≥5.0 mmol/L (48.5% vs 2.5%, P<0.001). For the drugs that could increase serum potassium levels, there was a significant correlation between Chinese herbal medicine and the occurrence of hyperkalemia, while renin-angiotensin-aldosterone system inhibitor (RAASi) and potassium supplementation showed no significant difference between the two groups. The results of ROC curve analysis showed that the AUC was 0.914, with the sensitivity of 84.8% and the specificity of 79.8% with the cut-off value of 4. The difference of AUC between the previously established risk assessment model of hyperkalemia in patients with non-dialytic CKD and current validation dataset was not statistically significant (Z=1.924, P=0.054), indicating the good accuracy and consistency of the prediction model. In the calibration curve, when the predicted risk of patients was below 0.4 or above 0.6, the prediction effect of the model was better. Conclusion: The previously-constructed hyperkalemia prediction model in non-dialytic CKD patients had good accuracy and consistency, and could be used to evaluate the risk of hyperkalemia in all stages of non-dialytic CKD patients.