Our study observed a strong correlation between serum [K+] level and ICU mortality with a J-shaped association, and mean [K+] level between 3.0 to 4.0 mmol/L showed the lowest mortality rate. In-hospital and ICU mortality were significantly higher with hyperkalemic (mean serum [K+] ≥ 4.0 mmol/L) patients but increased insignificantly with hypokalemic (mean serum [K+] < 3.0 mmol/L) patients. Furthermore, higher [K+] variability within the first 24 hours after initiation of CRRT indicated increased mortality. In addition, we found that delayed strategy (CRRT initiated > 24 hours after renal failure) would increase mortality risk.
There is a lack of evidence that the mean [K+] level and [K+] variability are associated with mortality in AKI patients requiring CRRT in ICU. Serum [K+] level 3.5 to 5.0 mmol/L is currently accepted as a safe range for critically ill patients.17 In a retrospective study conducted by Hessels et al. in 2015 , a U-shaped association between [K+] and in-hospital mortality for all ICU patients had been reported, and serum [K+] levels between 3.5 and 5.0 mmol/L were associated with the lowest mortality. However, the optimal range of serum [K+] levels for specific patient groups in ICU remains inconclusive. This issue has been studied in several cohorts, such as 3.5 to 4.5 mmol/L  and even 4.5 to 5.5 mmol/L  in acute myocardial infarction and > 3.5 to 4.0 mmol/L in patients with atrial fibrillation . However, in the acute stage of ARDS patients with positive fluid balance, relative hyperkalemia (up to 5.9 mmol/L) is associated with a reduced risk of death . Compared to previous reports, our study observed a tighter [K+] level (3.0−4.0 mmol/L) in AKI patients supported with CRRT in ICU. The possible explanation was that those patients requiring CRRT were often hemodynamically unstable and needed intravenous inotropic agents. These inotropic agents, such as norepinephrine, epinephrine, vasopressin, or dopamine, were commonly known as arrhythmogenic . Both lower and higher serum values have electrophysiological effects of promoting cardiac arrhythmias or myocardial ischemia [22–23]. Those patients with CRRT in ICU are assumed to be more vulnerable to cardiovascular events and may require a stricter control of serum [K+] level. Thus, [K+] level of 3.5 to 5.0 mmol/L should not be regarded as a normal range in this critical population. To go a step further, the [K+] monitoring and correction protocol for CRRT in each institute may need revision to achieve a better clinical outcome.
The variability or fluctuations of serum [K+] level recently emerged as a new focus on investigating its relationship with mortality in the hospital setting. In a monocentric and retrospective observational study by G. Lombardi et al. in Rome, 64,057 hospitalized patients were analyzed. High [K+] variability was reported to be an independent risk factor of in-hospital mortality, even within the normal [K+] range . However, according to this large-scale cohort study, the data was insufficient to address the relationship in specific medical conditions. Some previous studies were designed to investigate the association in critically-ill patients. Using a computerized regulation protocol designed for surgical ICU  to minimize the time in hypo- and hyperkalemia, Hessels et al. reported a low mortality rate with the lower [K+] variability . In the Soroka Acute Myocardial Infarction II (SAMI-II) Project, [K+] variability was associated with increased risk of mortality in patients with acute myocardial infarction (AMI) . Thongprayoon et al. reported hypokalemia (≤ 3.4 mmol/L) and hyperkalemia (≥ 4.5 mmol/L) before CRRT and hyperkalemia (≥ 4.5 mmol/L) during CRRT predicted higher 90-day mortality . Nevertheless, studies evaluating the prognostic value of [K+] fluctuations in patients with CRRT settings are scarce. As far as we know, our study was the first study to describe this issue. Although a direct causal relationship cannot be demonstrated in our study design, there are several possible explanations of the relationship between [K+] variability and mortality in our study. First, fluctuation in cell membrane resting electrical conditions could increase cellular instability and increase the risk of arrhythmogenic deaths. Another potential explanation is that higher [K+] variability may be a surrogate marker of baseline characteristics or disease processes with a poorer prognosis. Interestingly, very low [K+] variability was associated with an increased mortality rate in both hypo- and hyperkalemia (Fig. 4). This phenomenon reminded clinicians to correct dyskinesia more intensively to a normokalemic status.
In line with previous studies, we observed lower mortality in patients with higher serum albumin levels  and pre-existing diabetes mellitus . On the contrary, the delayed strategy of CRRT was associated with increased risk, similar to the result in the ELAIN randomized clinical trial . In addition, vasopressors therapy and the presence of malignancy were also risk factors for mortality, and the association has been reported in previous studies [31–32]. It's noteworthy that our study was observed marginally lower mortality with higher body temperature. Hypothermia was reported to have negative clinical consequences [33–34]. In ICU patients, hypothermia risk is increased by sedation, immobility, paralytic drugs, sepsis, underlying endocrine disorders, and higher CRRT dose. However, the patients’ body temperature could partially be manipulated via extracorporeal blood circulation and warming system in CRRT. Whether the body temperature is a maker or indicator for ICU mortality and the optimal body temperature range during CRRT could not be explained clearly in our study and warranted further investigation.
There are some limitations to be considered in our study. First, this was a retrospective observational study, and therefore, a causal relationship between serum [K+] level or variability and mortality could not be concluded. Second, we pooled all adult patients in medical and surgical ICUs together. However, different [K+] target ranges may depend on individual conditions. Third, some possible residual confounders may not be considered despite the adjusted analysis. Furthermore, we did not record the exact causes of death to better address the mechanistic linkage in our findings. Finally, we conducted this study using a large-scale electronic database from a single center with a primarily Taiwanese patient population. This design might limit the generalizability of our results to other patient populations.