For the present analysis, we linked 2 large, population-based, automated health care databases from our public health system. These databases capture complementary, individual-level, longitudinal health-related information of all residents. For the present analysis, one database was used as the source for sociodemographic characteristics, medical conditions (coded using either the International Classification of Diseases, 9th Revision, Clinical Modification [ICD-9-CM] coding system for in-hospital diagnoses, the International Classification of Diseases, 10th Revision [ICD-10] for diagnoses generated in primary care settings, and the clinical classifications software for both), and medication dispensing (coded using the anatomical therapeutic chemical coding system). The other database was used to identify incident events during follow-up, as well as for health care resource use and associated health care costs calculations. In addition, we linked these databases to a third, laboratory test result database that captures any laboratory test results generated during routine clinical care in both hospitals and primary care settings. This database has been available since January 1st, 2015.

We restricted our analyses to the metropolitan health care area of our hospital, which comprises approximately 1.2 million persons. The study period was defined between January 1st, 2015 and December 31st, 2017. Inclusion criteria were: a) age at study entry ≥ 55 years and b) availability of at least one serum K+ measurement during the evaluation period. Additional criteria were applied for each of the 4 analyses conducted, as detailed below.

We used 4 analytic approaches to evaluate the associations between K+ derangements, and the study outcomes under different sets of assumptions: prevalent case, incident case, prevalent case-prevalent (RAASI) user, and incident case-incident (RAASI) user analyses (table 1).

In the “prevalent case” analysis (figure 1 of the supplementary data), we included individuals with at least 1 prevalent relevant chronic cardiovascular, metabolic or renal condition (CHF, CKD, DM, hypertension, or IHD) as of January 1st, 2016 (regardless of their use of RAASI medications). Evidence of these diagnoses was sought in the database using operational definitions based on ICD-9-CM, ICD-10, and clinical classifications software codes (tables 1-3 of the supplementary data). For each of these individuals, we evaluated any K+ levels recorded in the database between January 1st and March 31st, 2016 (“evaluation period”). Information on other relevant covariates was collected at the end of the evaluation period (ie, March 31st, 2016), except for estimated glomerular filtration rate (eGFR), which was evaluated during the preceding year. Clinical outcomes were assessed between March 31st, 2016 and up to 12 months of follow-up (ie, March 31st, 2017).

In the “incident case” analysis (figure 2 of the supplementary data), included individuals could not have recorded evidence of any of the 5 relevant conditions as of January 1st, 2016, and had a first recording of at least 1 of those conditions between January 2nd, 2016, and December 31st, 2017 (“index date”). K+ levels were evaluated, for each participant, between the index date and the subsequent 90 days (“evaluation period”). Clinical endpoints were assessed for up to 12 months after each patient's evaluation period.

In the “prevalent case-prevalent user” analysis (figure 3 of the supplementary data), participants had to have at least 1 relevant condition as of January 1st, 2016 to be included and to be using at least one RAASI medication (ACEIs, ARBs, MRAs, or RIs) in the preceding 3 months. The use of these medications was identified using operational definitions combining anatomical therapeutic chemical codes (table 4 of the supplementary data). The rest of the analytic design is comparable to that of the prevalent case analysis presented above.

Finally, in the “incident case-incident user” design (figure 4 of the supplementary data), participants could not have recorded evidence of a relevant condition or of RAASI medication use as of January 1st, 2016. The first recording of at least 1 relevant condition had to occur between January 2nd, 2016, and December 31st, 2017 (“index date”), with evidence of subsequent dispensing of at least 1 RAASI medication. The date of the first dispensing of a RAASI drug was the “modified index date”. For each participant, we then evaluated any K+ levels recorded in the database between the modified index date and the subsequent 90 days (“evaluation period”). The rest of the design features are similar to those of the incident case analysis.

In each of the 4 analyses, recorded laboratory test results during the evaluation period were reviewed for serum K+ levels. Hyperkalemia was defined as serum K+ levels> 5.0 mEq/L; hypokalemia as serum K+ levels <3.5 mEq/L, and normokalemia as serum K+ levels ≥ 3.5 and ≤ 5 mEq/L. Based on the (1 or more) K+ level measurements available for each participant during the evaluation period, 6 clinical profiles were defined a priori: a) “hypokalemic” patients, those who had evidence of 1 or 2 episodes of hypokalemia (with at least 7 days between the 2 episodes, otherwise they were considered part of the same episode) and no evidence of hyperkalemia; b) “hyperkalemic” patients, those who had 1 or 2 episodes of hyperkalemia during the evaluation period (with at least 7 days between the 2 episodes, otherwise they were considered part of the same episode) and no evidence of hypokalemia; c) “normokalemic” patients, those who had no evidence of hyper or of hypokalemia during the evaluation period; d) “recurrent hyperkalemic” patients, those individuals with 3 or more episodes of hyperkalemia and no hypokalemia; e) “recurrent hypokalemic” patients, those individuals with 3 or more episodes of hypokalemia and no hyperkalemia; and f) “mixed K+ derangements” patients, those with both hyper- and hypokalemic episodes. The very small number of participants in groups d through f (n <5 for all) resulted in their exclusion and analyses were restricted to hyperkalemic, hypokalemic, and normokalemic patients.

Information on other relevant covariates was collected at the end of the evaluation period, including age, sex, comorbidity index (using the adjusted morbidity groups comorbidity and complexity classification system,12table 5 of the supplementary data), individual income, K+ derangement-associated medications (ACEIs, ARBs, MRAs, RIs—when appropriate—, beta-blockers, K+ supplements, loop diuretics, nonsteroidal anti-inflammatory drugs, trimethoprim, and macrolides [table 4 of the supplementary data for lists of anatomical therapeutic chemical codes]), disease duration, number of prior hospitalizations, number of prior emergency department visits, and number of serum K+ level tests.

The baseline characteristics of the study population included in each of the 4 study designs (prevalent case, incident case, prevalent case-prevalent user, and incident case-incident user) were described at the index date, overall and by K+ profile (hyperkalemic, hypokalemic, normokalemic). Categorical variables were compared across K+ groups using chi-squared tests, and continuous variables using ANOVA and nonparametric tests, as appropriate.

The crude incidence rate of the first occurrence of each study outcome was described during follow-up, per 1000 person-years, for each of the K+ profiles. Kaplan-Meier survivor function curves were also used to graphically describe their event-free survival.

We used Cox proportional hazards regression models to calculate the multivariable-adjusted hazard ratios (HR) of each of the clinical endpoints, comparing hyperkalemic and hypokalemic patients, respectively, to normokalaemic patients (reference group), and adjusting for the following potential confounders: age, sex, adjusted morbidity groups, medication use, individual income, and median eGFR levels during the preceding year. In prevalent analyses, we further adjusted for disease duration, number of hospitalizations, number of prior emergency department visits, and number of serum K+ level tests (all in the 6 months before January 1st, 2016). For endpoints other than mortality, Fine and Gray models were used to account for the competing risk with death.

Finally, for the exploratory health care expenditure analysis, we used logistic regression to calculate the multivariable-adjusted odds ratios of an overall yearly expenditure> 85th percentile during 2016, comparing hyperkalemic and hypokalemic, respectively, to normokalemic patients (reference group), with adjustment for the potential confounders listed above.

The Ethics in Research Committee of our University Hospital and the Research Institute provided written approval to the study protocol.

Table 1 displays the number of individuals included in each of the 4 analyses. The largest study population was that of the prevalent case analysis (n=36 269).

Among the study population included in the prevalent case analysis (table 2), the vast majority were normokalemic (96%) during the 3-month evaluation period, and there were more than twice the number of hyperkalemic (2.7%) than hypokalemic (1.3%) individuals. The prevalence of hyperkalemia was higher in men than in women, and increased with increasing age and with decreasing eGFR (9.6% among those with eGFR <30). Among relevant conditions, the highest prevalence of hyperkalemia was observed in individuals with CKD. Hypokalemia was more frequent among women and among patients with CHF.

Similar trends for hyper- and hypokalemia were observed among prevalent cases using at least 1 RAASI medication as of January 1st, 2016 (table 8 of the supplementary data), although the prevalence of hyperkalemia was slightly higher in this population, particularly in specific subgroups (eg, a prevalence of 11.0% among those with eGFR <30). The use of MRA was associated with a higher prevalence of hyperkalemia (4.8%) compared with the use of ACEIs and ARBs. In incident case and incident case-incident user analyses (tables 7-9 of the supplementary data), the overall frequency of hyperkalemia was higher (4.0% and 4.4%, respectively) than that observed in the prevalent case analyses. The same was true for hypokalemia (2.3% and 2.1%, respectively).

Table 3 presents the crude incidence rates of the study outcomes during follow-up overall and by K+ profile. Emergency department visits were the most frequent event (incidence rates ranging from 441 to 498 per 1000 person-years) while all-cause death and hospital daycare visits were the least frequent. For all study outcomes, higher rates were observed in the incident compared with the prevalent analyses. In all analyses, hyperkalemia was associated with higher crude incidence rates of all-cause death compared with patients with normokalemia (incidence rates in the prevalent case analysis of 39.94 and 78.29 per 1000 person-years, respectively). However, the highest rates of all-cause death occurred among individuals with hypokalemia (incidence rate in the prevalent case analysis of 125.59 per 1000 person-years). Hyperkalemia and hypokalemia were also associated with higher crude incidence rates of hospitalization, emergency department visits, and hospital daycare visits compared with patients with normokalemia. More specifically, for hospitalization and emergency department visits, the crude incident rate for prevalent (case and user) analyses was higher for hyperkalemic patients than for the other groups. In contrast, in terms of incident analyses, these rates were higher for hypokalemic patients than for the other groups.

Consistent results were observed in analyses using Kaplan-Meier survivor function plots (figure 1A-D). There was a trend toward worse event-free survival in participants with hyperkalemia and with hypokalemia compared with normokalemia for the 4 clinical study outcomes. The 95% confidence intervals (95%CI) were narrower in the prevalent analyses and wider in the incident analyses. Finally, differences in event-free survival between individuals with K+ abnormalities and normokalemic participants were particularly salient for all-cause death, the worst prognosis being observed in participants with hypokalemia.

Figure 1.

Kaplan-Meier cumulative survivor function curves for all-cause death (upper left), hospitalization (upper right), ED visits (lower left) and daycare visits (lower right), in the prevalent case analysis (A), incident case analysis (B), prevalent case-prevalent user analysis (C), and incident case-incident user analysis (D). ED, emergency department; Hyper K, hyperkalemia; Hypo K, hypokalemia; Normo K, normokalemia.

(0.84MB).

Table 4 summarizes the results of the multivariable-adjusted regression analyses comparing the risk of each of the 4 study outcomes among hyperkalemic and hypokalemic (compared with normokalemic) individuals, adjusting for potential confounders. Hyperkalemia was associated with an increased risk of all-cause death compared with normokalemia, ranging from a 31% to 68% increased risk depending on the analytic approach used. The 95%CI included the null value only in the incident case-incident user analysis. The risk of death was even higher among individuals with hypokalemia, the HR ranging from 1.92 to 2.70 depending on the analytic approach used, with all 95%CI above 1.00.

With regards to the other 3 clinical outcomes, weaker trends toward increased risks of hospitalization, emergency department visits, and hospital daycare visits were observed in individuals with hyperkalemia compared with those with normokalemia, although all 95%CIs included the null value. For hypokalemia, in incident case analyses there were strong, statistically significant associations with hospitalization (HR, 1.41; 95%CI, 1.09-1.82), emergency department visits (HR, 1.30; 95%CI, 1.06-1.60) and hospital daycare visits (HR, 1.95; 95%CI, 1.23-3.08). The incident case-incident user analysis yielded similar findings in patients with hypokalemia, although 95%CIs were wider.

Subgroup analyses by prevalent/incident condition were limited by the small absolute number of events occurring in specific strata (tables 10-12 of the supplementary data), and could not be conducted in the incident case-incident user population. In the prevalent case analysis, the strongest associations between K+ derangements and all-cause death were observed in CKD patients.

In the exploratory analyses assessing yearly health care expenditure among prevalent cases (table 5), patients with hyperkalemia had higher multivariable-adjusted odds of having a yearly health care expenditure> 85th percentile compared with normokalemic patients, the increased odds ranging from 21% to 29%. This association was even stronger for hypokalemic than for normokalemic patients (increased odds ranging from 81% to 85%).

Our study has important strengths compared with prior research in this field. We used a very large population database and included a large number of patients with each of the relevant conditions. This increased overall statistical power and precision compared with previous studies conducted in smaller clinical cohorts. Modeling a time-varying, potentially recurrent exposure such as serum K+ levels is challenging, and a number of modeling approaches have been described in the literature. The design becomes even more complex when considering exposure to specific drugs, and pharmacoepidemiological methods are necessary. Therefore, for the present analysis, we used a variety of study designs and analytic approaches, aimed at maximizing the robustness of the results and at minimizing their sensitivity to specific study design assumptions and potential biases. The consistency of the observed results across analytic approaches further reinforces the validity of our findings.

In addition, to minimize the possibility of residual confounding, adjustments for a number of potential confounders were conducted. Specifically, detailed, updated data on baseline eGFR was used to ensure adequate adjustment for this key covariate. Because eGFR and K+ levels are strongly correlated, this information was obtained before (rather than during) the evaluation period for each participant, which allowed adjustment for baseline eGFR while minimizing collinearity and over-adjustment.

The present study also has some limitations. First, as for any analysis using information generated from routine care and recorded in large administrative health care databases, under-recording of health conditions is possible. This may have led to some underestimation of the relevant cardio-metabolic conditions assessed, but this is unlikely to have affected the reported associations among individuals with recorded conditions. In addition, some residual confounding on adjustment for comorbidities in the multivariable analyses cannot be ruled out. Nevertheless, the multivariable analytic strategy used was very comprehensive, including adjustment for adjusted morbidity groups, a summary measure of complexity and comorbidity, and for other clinically relevant features such as disease duration and prior hospitalizations, which are expected to be able to capture baseline differences in the likelihood of incident events across groups.

Second, although the overall sample size of the study was very large, the number of patients in specific subgroups was relatively small. This limited our ability to evaluate the associations in some specific subgroups such as those patients with recurrent hyper- or hypokalemia and those with mixed K+ derangements. Of note, inclusion in the study population of individuals with no K+ measurements—who were excluded from the analysis—would not have ameliorated this, as the small counts issue affected hyper- and hypokalemic strata rather than normokalemic groups. Since small numbers of events are a frequent issue in pharmacoepidemiological research studies, larger studies combining multiple population-based databases may be necessary to characterize the associations between K+ derangements and incident events in specific subgroups of patients.

Finally, our health care expenditure analyses used a published calculation for overall health care expenditure. Nonetheless, this calculation is intended to be used for a given whole calendar year (eg, 2016) rather than for time periods of varying lengths, such as those allowed for in our incident analyses. Consequently, such analyses could not be performed in the incident designs. More generally, due to their limited granularity, these analyses should be considered hypothesis-generating, and more detailed health economics evaluations should be considered in smaller clinical cohorts.