Date of Award
Doctor of Philosophy (PhD)
Health Outcomes and Policy Research
Csaba P. Kovesdy, MD
Oguz Akbilgic, PhD Justin Gatwood, PhD Keiichi Sumida, PhD Fridtjof Thomas, PhD
chronic kidney disease, dialysis, outcomes, potassium, retrospective
Background: The kidneys play a crucial role in maintaining homeostasis of serum potassium levels (K+). Patients with advanced chronic kidney disease (CKD) are at a higher risk of experiencing dyskalemia events (i.e. hyper- and hypokalemia; especially the former) and thus future adverse outcomes. Currently, there is a dearth of literature on prediction models for hyperkalemia and the effects of dyskalemia on outcomes such as incidence of ischemic stroke and short-term hospital/emergency room (ER) utilization in an advanced CKD population transitioning to dialysis. Objectives: Using a nationally representative sample of US veterans with advanced CKD transitioning to dialysis, in the pre-dialysis period we studied the following aims: Aim 1) Develop and validate a prediction model for predicting hyperkalemia in individual patients; Aim 2) Examine the association of dyskalemias with time to first ischemic stroke; Aim 3) Examine the association of dyskalemias with time to short-term hospital/ER utilization. Methods: A retrospective cohort analysis of the Transition of Care in Chronic Kidney Disease cohort (n=102,477), a nationally representative sample of US veterans with advanced CKD transitioning to dialysis between October 1, 2007 through March 31, 2015 identified from the United States Renal Data System was conducted. Across the three study aims, we identified patients with an initial selection criterion (prior to dialysis initiation) of two estimated glomerular filtration rate (eGFR) of /min/1.73m2 90-365 days apart (second eGFR as index); at least one-year each of baseline period (prior to index) and follow-up period (following index but prior to dialysis initiation); and at least one K+ measurement each in the baseline and follow-up period. For each study aim, further inclusion criteria were used to yield a final sample size of 21,654, 21,357, and 21,366 for aim 1, aim 2, and aim 3, respectively. For Aim 1 (Chapter 2), we compared the performance (area under the receiver operating curve [AUROC]) of different machine learning methods including logistic regression (LR), random forest, extreme gradient boosting, and support vector machines using geographical splitting (for creating training and test set) with 10-fold cross validation to predict the outcome of hyperkalemia (K+ >5.5 mEq/L). The method that yielded the best performance was used to build a reduced model with 10 predictors to develop a patient-level hyperkalemia risk score. For Aim 2 (Chapter 3), we assessed the association of baseline time-averaged K+ levels (distant exposure) and time-updated K+ levels (acute exposure) (both categorized as hypokalemia [K+ 5.5 mEq/L] and referent [3.5 mEq/L ≤ K+ ≤ 5.5 mEq/L]) with time to first ischemic stroke using Cox regression models. Finally, for Aim 3 (Chapter 4), we assessed the association of time-updated outpatient K+ levels (categorized as hypokalemia [K+ 5.5 mEq/L] and referent [3.5 mEq/L ≤ K+ ≤ 5.5 mEq/L]) with hospital/ER utilization (as separate events) using generalized estimating equations. Across all the three study aims (Aim 1, 2, and 3) several different sensitivity analyses were conducted to test the robustness of the results. Results: Across the analytic samples (Aim 1, 2, and 3), the mean age was 69 years, ~98% were males; ~28% were African Americans, ~69% had diabetes mellitus, and the one-year baseline averaged K+ was 4.5 mEq/L. In aim 1 (n=21,654), the LR model yielded the best performance with an average AUROC (95% confidence interval [CI]) of 0.765 (0.756-0.774) (training set) and 0.763 (0.753-0.771) (test set) using the geographical splitting with 10-fold cross validation. Using the LR method, the top 10 predictors identified were K+ value prior to index, age, having at least 1 K+ >5.5 mEq/L in the baseline, index eGFR, baseline averaged SBP, baseline averaged HCO3-, number of K+ counts, thiazide use, number of outpatient visits, and NSAIDs use in baseline. The LR parameter estimates for the above listed predictors were used to develop a patient-level risk score for predicting hyperkalemia. In aim 2 (n=21,357), hypokalemia (distant exposure) was associated with higher risk of ischemic stroke (hazard ratio [HR]; 95 % CI: 1.35, 1.01-1.81). Conversely, hyperkalemia (acute exposure) was associated with a lower risk of ischemic stroke (HR; 95% CI: 0.82, 0.68-0.98). Finally, in aim 3 (n=21,366) using outpatient K+ levels, both hyperkalemia (odds ratio [OR]; 95% CI: 2.04; 1.88-2.21) and hypokalemia (OR; 95% CI: 1.66; 1.48-1.86) were associated with higher risk of hospital visit within 2 calendar days of outpatient K+ measurement. Similarly, both hyperkalemia (OR; 95% CI: 1.83; 1.65-2.03) and hypokalemia (OR; 95% CI: 1.24; 1.07-1.44) were associated with higher risk of ER visit within 2 calendar days of outpatient K+ measurement. Across all the three study aims, the results were robust to various sensitivity analysis. Conclusion: In an advanced CKD population transitioning to dialysis, in the pre-dialysis period, we developed an internally valid model for predicting hyperkalemia. We observed that hypokalemia as a chronic exposure is associated with higher risk of ischemic stroke and hyperkalemia as an acute exposure is associated with lower risk of ischemic stroke. Finally, both hyper- and hypokalemia are associated with higher risk of short-term hospital/ER visits. Further studies are needed to externally validate the hyperkalemia risk prediction model; understand the mechanisms underlying the association of dyskalemias with stroke; and expand on the association of dyskalemias with short-term hospital/ER visits by including cost as an outcome.
Dashputre, Ankur (https://orcid.org/0000-0003-0898-959X), "Predictors of Hyperkalemia and Outcomes of Dyskalemia in US Veterans with Advanced Chronic Kidney Disease Transitioning to Dialysis" (2022). Theses and Dissertations (ETD). Paper 589. http://dx.doi.org/10.21007/etd.cghs.2022.0590.