The association between serum glucose to potassium ratio on admission and short-term mortality in ischemic stroke patients

High serum glucose to potassium ratio (GPR) at admission is implicated for a poor outcome in acute brain injury, acute intracranial hemorrhage, and aneurysmal subarachnoid hemorrhage. However, the relationship between GPR and the outcome of ischemic stroke (IS) remains unknown. In all, 784 IS patients from a large emergency Norwegian cohort were included for secondary analysis. The exposure and outcome were GPR at baseline and all-cause mortality within 30 days after the first admission. Multivariable logistic regression analysis was performed to estimate the risk of 30-day mortality based on GPR levels. In addition, we examined whether there was a nonlinear relationship between admission GPR and 30-day mortality using two-piecewise linear regression with a smoothing function and threshold level analysis. The results of multivariable regression analysis showed that GPR at baseline was positively associated with the 30-day mortality (OR 2.01, 95% CI 1.12, 3.61) after adjusting for potential confounders (age, gender, department, serum sodium, serum albumin, serum-magnesium, hypertension, heart failure, chronic renal failure, and pneumonia). When GPR was translated to a categorical variable, the ORs and 95% CIs in the tertiles 2 to 3 versus the tertile 1 were 1.24 (0.60, 2.56) and 2.15 (1.09, 4.24), respectively (P for trend = 0.0188). Moreover, the results of the two-piecewise linear regression and curve fitting revealed a linear relationship between GPR and 30-day mortality. In IS patients, GPR is positively correlated with 30-day mortality, and the relationship between them is linear. The GPR at admission may be a promising predictor for the short-term outcome in IS patients.

Serum glucose and potassium are two important blood indicators that are commonly used clinically. As the main energy source of cells in the human body, glucose is a critical factor for maintaining cellular metabolism 4 . Potassium ion, the most abundant cation in the cells of a human body, plays a crucial role in physiological processes including neural conduction, cardiac pulsation, muscle contraction, and maintenance of normal renal function 5 . In addition, both serum glucose and potassium disturbances have been revealed to be correlated with the risk of stroke 6,7 . Previous studies have demonstrated that there were complex interactions among potassium and glucose in the human body 8,9 . Given the potential combined effects of glucose and serum potassium, the serum glucose to potassium ratio (GPR) has been used in a few studies and has been shown to be an early prognostic factors for central nerve injures including aneurysmal subarachnoid hemorrhage (aSAH) 10 , acute intracerebral hemorrhage 11 , severe traumatic brain injury 12 , and neuropsychiatric syndrome after carbon monoxide poisoning 13 .
However, the relationship between GPR and the clinical outcome of IS remains unknown. Therefore, we aimed to explore the association between GPR at admission and short-term mortality in IS patients based on a retrospective cohort study.

Methods
Data source. Original data were published by Tazmini et al. 14 on the "DRYAD" website (www. datad ryad. org). And Tazmini et al. 15 authorized the ownership of their raw data to the "DRYAD" database. Thus, this secondary research based on the raw data for a different research hypothesis was permitted.
The original research was a single-center retrospective cohort study that included 31,966 unique patients (62,991 registered admission information) who visited the emergency department of the Diakonhjemmet Hospital in Oslo (Norway) from 2010 to 2015. According to the ICD-10 standard of classification, 974 visits (admission information) were diagnosed as IS (ICD-10, I63). The raw data included information on multiple hospitalizations for the same patient, but only the first visit of each patient was considered in this study. Thus, we excluded the second or subsequent admissions (n = 88). Only 886 unique IS patients were then considered during analysis. Subsequently, 102 patients were excluded for missing data concerning serum glucose or potassium levels (n = 6), incorrectly recorded days of death (n = 5), and presence of diabetes mellitus or serum glucose level > 200 mg/dL (n = 91) at admission. Finally, 784 unique participants were included in the study (Fig. 1).
Exposure. All the laboratory indicators were obtained from the first-time laboratory results at admission.
All IS patients were categorized into five subtypes according to the TOAST stroke subtype classification system: large-artery atherosclerosis, cardioembolism, small-vessel occlusion, other determined etiology, and undetermined 16 . The information of patients admitted to the medical or surgical department was identified as a binary variable.
Outcome. The primary outcome was all-cause mortality within 30 days after first admission.
Missing data. All missing data of covariates are stated in Table 1. Considering that missing data may reduce statistical power or even lead to bias, covariates with too much missing data (e.g., concerning serum phosphate and serum magnesium levels) were handled as categorical variables. And dummy variables were used to identify the missing values of the covariate 17 .
In addition, to further assess whether missing data being handled as dummy variables can introduce bias into the results, we used multiple imputations based on five replications and chained equation approach in the R MI procedure to handle all the missing data for sensitivity analysis 18 . The process of multiple imputation and the results of multivariate regression analysis based on five multiple imputation data are shown in Supplementary  Fig. 1  www.nature.com/scientificreports/ relationship between GPR and 30-day mortality, curve fitting and two-piecewise linear regression analysis were performed. We built three models to regulate the potential confounding factors, they are : (1) Crude model, i.e., unadjusted; (2) Model I, adjusted for age and gender; (3) Model II, adjusted for age, gender, department, serum sodium, serum albumin, serum-magnesium tertiles, hypertension, heart failure, chronic renal failure, and pneumonia. Covariates were selected based on their relationship to 30-day mortality or their ability to change the effect value by more than 10% 19 , gender was also included as a basic covariate. Sensitivity analysis. GPR tertiles were also used to test the stability of multiple regression analysis results, and the linear tests were performed by assigning medians to each GPR tertile as a continuous variable in the models 20 .
An E-value was used to explore the potential of unmeasured confounding between GPR and 30-day mortality. The E-value was defined as the required magnitude for an unmeasured confounder to overturn the observed association between GPR and 30-day mortality 21 .
Ethics approval and consent to participate. Ethics of the previous study was approved by the Norway regional committee (Regional Committee for Medical and Health Research Ethics South East) and informed consent was exempt for anonymous data. Thus, our secondary analysis based on this original study did not require separate ethical approval. And our study was carried out following all the relevant guidelines and regulations.

Multivariate logistic regression analysis of GPR and 30-day mortality.
In multivariate regression analysis, we built three models adjusting for different covariates to verify the stability of the results. The results of the crude model without adjusting for confounder factors showed that GPR and 30-day mortality were positively correlated (OR 2.18, 95% CI 1.31-3.63).Model I which was adjusted for age and gender also indicated the same association (OR 2.09, 95% CI 1.24-3.54). Model II which was further adjusted for age, gender, department, serum sodium, serum albumin, serum-magnesium tertiles, hypertension, heart failure, chronic renal failure, and pneumonia also revealed that GPR was independently associated with 30-day mortality (OR 2.01, 95% CI 1.12-3.61) ( Table 3).
Curve fitting and two-piecewise linear regression model of GPR and 30-day mortality. Curve fitting analysis, was adjusted according to three models (crude model, model I, model II) all indicated that a curve that that continues to rise, and the inflection point was approximately 1.6 (Fig. 2). According to the inflection point (GPR = 1.6), two-piecewise linear regression analysis models for different confounding factors (crude model, model I, model II) were used to explore the potential non-liner relationship. The result of model II showed two different effective sizes in two-piecewise linear regression equations, but the P-value for likelihood ratio test was 0.55 which was not statistically significant. The other two models showed similar results and the P-values for the likelihood ratio test were all > 0.05 (Table 4). Thus, the relationship of GPR with 30-day mortality was linear.
Sensitivity analysis. GPR (tertiles) were also plugged into the multiple regression equation for sensitivity analysis. The results of GPR as categorical variable (tertile) were consistent with the results of GPR as a continuous variable, the top tertile had 115% increment of diabetes risk when compared with the bottom tertile in the full model (model II), and found that the trend across the tertiles was significant (P for trend = 0.0188). Other two models (crude model, model I) showed similar results (Table 3). An E-value was calculated to assess the sensitivity to unmeasured confounding. The primary findings were stable unless an unmeasured confounder existed and high positively related to GPR (OR ≥ 3.43) and 30 daymortality (OR ≥ 3.43).

Discussion
To the best of our knowledge, this is the first study to explore the relationship between GPR and the clinical IS outcome. Our study showed a significantly positive correlation between GPR levels and 30-day mortality. Further, the stability of the association was verified by adjusting for potential confounding factors (mode I, OR 2.09, 95% CI 1.24-3.54; model II, OR 2.01, 95% CI 1.24-3.54). In sensitivity analysis, we handled GPR as a categorical variable (tertiles) and the results showed an increasing trend of OR values from tertile 1 to tertile 3 in the three models (P values for trend all < 0.05). Moreover, the curve fittings of GPR levels and 30-day mortality showed a gradual upward curve in smoothing plots for the three different models (Supplementary Fig. 1). According to the inflection point in the curve fitting plot, two-piecewise linear regression analyses with three different adjustment methods were performed and all the results showed a linear relationship between GPR and 30-day mortality.
GPR is a novel parameter that can be measured quickly in clinics. Fujiki et al. first reported the potential association between baseline GPR and H-K grade and Glasgow score at discharge in a retrospective cohort study including 565 aSAH patients 10 . In another study, they investigated cerebral vasospasm after aSAH, and reported that elevated GPR levels were related to cerebral vasospasm grades and ischemic events induced by cerebral vasospasm 22 . In addition, the roles of GPR levels in other acute neurological injury related disease including acute intracerebral hemorrhage. Neuropsychiatric syndrome after carbon monoxide poisoning, and severe traumatic brain injury, have been proven [11][12][13] . In these studies, the baseline GPR levels were all observed in worse clinical outcome group than the normal group. Our results added evidence with regard to association between GPR levels and short-term outcome (30-day mortality) in cerebral ischemic injury. The results of the aforesaid studies suggested that GPR levels were closely related to pathological neurological disorders.
Hyperglycemia is very common in the acute phase of IS, even among non-diabetic IS patients 23 . The phenomenon of post-stroke hyperglycemia was believed to be a type of stress hyperglycemia induced by high cortisol and catecholamine levels after ischemic injury 24 . In addition, stress hyperglycemia had been suggested to be associated with stroke severity. Patients with stress hyperglycemia often had more serious strokes than those with type 2 diabetes mellitus (T2DM) 25 . Guo et.al have reported that IS patients with stress hyperglycemia had a higher risk of 90-day stroke recurrence than those with T2DM 26 . However, inconsistent results were shown in studies to explore the association between hyperglycemia and clinical outcomes of IS patients 25,[27][28][29] . A study by Zonneveld et.al showed that stress hyperglycemia was associated with post-stroke infections and poor functional outcome 27 . Further, among both IS patients treated with intravenous thrombolysis 28 and those treated via mechanical thrombectomy 29 , stress hyperglycemia was proven to be associated with a poor outcome. However, Tziomalos et al. believed that stress hyperglycemia was correlated with stroke severity rather than directly being related to an adverse outcome 25 . Besides, a recent clinical trial showed that glucose-lowering therapy did not help in improving the prognosis 30 . In addition to the heterogeneity of study design, and the potential non-linear relationship between admission serum glucose and outcome reported previously 31 , the complicated and multifaceted path mechanism underlying stress hyperglycemia may lead to the discordance of results. Likewise, as another important clinical blood biomarker, serum potassium levels play a crucial role in maintaining basic www.nature.com/scientificreports/ cellular functions. Normally, potassium ion is mostly stored in the cells and transported outside the membrane by sodium/potassium ATPase when necessary. Some population-based evidences have demonstrated that a potassium-rich diet could lower the risk of stroke 32,33 . Nevertheless, existing studies investigating the association between serum potassium and stroke outcome have shown contrasting results [34][35][36] . In our univariate analysis, admission serum potassium was not associated with 30-day mortality. This may be because of the intermediate factors, such as serum glucose, which impacted the relationship between serum potassium and short-term outcome. Accordingly, current research on the specific relationship between serum potassium level and the outcome of IS patients is still limited. Despite these observations investigating the baseline GPR levels and IS outcome, the mechanisms underlying these findings still remain unknown. In severe stress injury conditions, sympathetic activation would result in an increased secretion of stress hormones including catecholamines, growth hormone, cortisol, and cytokines, and then induce a hyperglycemic response and insulin resistance 24 . In acute IS patients, the regulation of sodium/ potassium ATPase by high catecholamine levels and secretion of insulin all would lead to potassium influx 37 . Thus, post-stroke hyperglycemia and hypokalemia may reflect the stress-related activation and a disorder of the hypothalamic-pituitary-adrenal (HPA) axis. HPA axis dysregulation believed to play a key role in the process of successive energy pump failure and various signaling cascades of IS 38 . In addition, a high cortisol level would activate the renin-angiotensin-aldosterone system (RAAS) to induce low serum potassium 39 . Brown et.al also thought that lower serum potassium level may represent increased activity of RAAS 40 . Current evidences show that RAAS plays a pivotal role in the progression of IS 41 , and angiotensin II receptors blockers could help to stroke prevention 42 . Based on the abovementioned discussion and considering the combined effects of serum potassium and serum glucose levels, the GPR index may be a good indicator for reflecting the status of HPA axis and RAAS dysregulation after IS. Moreover, studies of other stress damage types including acute myocardial infarction 43 , blunt abdominal trauma 44 , pulmonary embolism 45 , and even intermediate syndrome induced by anticholinesterase-containing chemicals poisoning 46 all showed a stable correlation between increased GPR and poor outcomes or more severe symptoms. These studies also suggest that GPR may be a potential marker of stress injury for reflecting the condition of the whole body in severe disease.    www.nature.com/scientificreports/ The pathogenesis of IS is complex, and there was limited knowledge concerning it until now. Therefore, most treatment strategies currently followed that were developed by targeting known key pathogenetic links are often inadequately effective. The role of stress responses 38 and stress-related markers 3 has been proven to play a pivotal role in IS progression and has gained increasing attention. Our results suggested that comprehensive treatments including appropriate potassium supplementation, hypoglycemic treatment, and stress response blocker (β-blocker, RAAS inhibitor) may improve the short-term prognosis with high GPR levels. However, because of the nature of this retrospective study, the causality of high GPR levels and short-term outcome in IS patients could not be established. Future studies were needed to explore the causal relationship and verify the effectiveness of these treatments. In addition, because the raw data of our study were from a large emergency cohort, GPR may have a broad application prospect for IS patients admitted in emergency departments, especially in primary or smaller emergency departments as a brief blood biomarker, information on which could quickly be obtained at admission.

Conclusion
Among IS patients, GPR is positively correlated with 30-day mortality, and the relationship between them is linear. Thus, GPR at admission may be a promising predictor of the short-term outcome of IS patients.

Strengths and limitations.
There are several advantages to our study. First, the results of univariate analysis, regression coefficient change, and previous literature were used to select covariates. Second, curve fitting and two-piecewise linear regression analysis were performed to explore the potential non-liner relationship, which had been shown in a previous study. Third, one crude model and three models which had been adjusted for potential confounding variables were used to test the stability of the results. Fourth, to avoid the contingency of analysis, GPR was considered as a continuous variable and categorical variable in the multiple regression equation, and sensitivity analysis and trend test were performed.
However, this study also has some limitations. First, the presence of unmeasured confounders could not be excluded. Since the secondary analysis originated from a retrospective cohort, variables that were not collected could not be adjusted. E-value was used to explore the potential for unmeasured confounding between GPR and 30-day mortality and the result showed that an unmeasured confounder was unlikely to explain the entirety of the mortality effect. Second, ICD-10 codes for renal failure may not be clear enough to identify renal function status, and hospitalization information after emergency admission including intensive care unit duration and length of hospital stay were not included in the analysis; future studies collecting renal function indicators (serum creatinine and baseline eGFR), intensive care unit duration, and length of hospital stay are required to more accurately explore the association and mechanisms. Second, there was no record concerning the levels of serum hormones such as catecholamines, glucagon, and corticosteroids in the original data; thus, we could not clarify the reason for high GPR in patients with severe IS. Third, lacking treatment information before the first blood tests (dextrose, potassium, or insulin) and after admission may lead to bias. However, given that the treatment would tend to a bias toward the null, we believed that the unmeasured confounding of medication treatment may underestimate the observed effect. Fourth, though first-time laboratory results at admission, which are more likely to reflect the initial state of the patient at the onset, were used, it would be better to examine the dynamic changes in GPR in future studies to understand the potential mechanism of the associations. Because of the retrospective study design, we could not confirm the time of blood collection, which will influence the GPR level. Thus, further prospective studies with predesigned identical examination time are required. Finally, the participants of this study are Norwegian populations, and the findings do not necessarily apply to other populations.

Data availability
The data are available from the 'DataDryad' database (www. datad ryad. org).