The study protocol was designed in accordance with the Declaration of Helsinki. The data were gathered in a registry which guaranteed the anonymity of the participants and was only known about by the research team.

It was a cross-sectional study carried out at four Avericum centres across the province of Las Palmas (Spain). The records of the haemodialysis sessions of 534 prevalent patients in June 2017 were assessed. The dialysis prescription was 3.5–4h with a blood flow between 300 and 450ml/min, using a standard dialysis concentrate of bicarbonate and high-permeability polyethersulfone haemodialysis filters (Elisio® – Nipro® Model) with a surface area greater than 1.7m2 and an ultrafiltration rate greater than 74ml/h/mmHg). The erythropoiesis-stimulating agent used in all patients was Epoetin Beta (Neorecormon® – Roche Farma AG®).

The exclusion criteria were not receiving treatment with ESAs at the time of the study (n=76) and medical conditions which may have influenced the absolute counts of platelets, lymphocytes and haemoglobin: admission to hospital for an infection, cancer or haematological causes in the previous three months (n=24). Transfusion in the last three months (n=4), active onco-haematological diseases (n=13) and patients undergoing steroid treatment (n=26) were also part of the exclusion criteria. Some of the patients fulfilled more than one exclusion criterion (total number of patients excluded: n=137). Therefore, after applying the criteria, the study sample was 397 patients.

Erythropoietin resistance was defined as those patients who had erythropoietin doses of greater than 20,000IU and haemoglobin values of less than 11g/dl.14

For the PLR calculation, the ratio between the absolute platelet count and the absolute lymphocyte count in the blood count was obtained. In order to obtain the NLR, the ratio between the absolute neutrophil count and the absolute lymphocyte count in the blood count was calculated. And, finally, the ERI was obtained by calculating the weekly dose of erythropoietin (IU) divided by the dry weight (kg) and divided, in turn, by the serum haemoglobin levels (g/dl).11,15,16

Other variables analysed were: albumin, serum iron, transferrin saturation index (TSI) and C-reactive protein (CRP).

The R program version 3.3.2 (2017) was used for the statistical analysis (R Foundation for Statistical Computing, Vienna, Austria. URL http://www.R-projet.org/). The results were considered statistically significant if the p-value was less than 0.05. The continuous variables were summarised as mean±standard deviation (normal variables) or median and quartiles (non-normal variables). The categorical variables were summarised as percentages. The normality of the variables was analysed using the Shapiro–Wilk test. The Pearson coefficient was used for correlations. To evaluate whether there were significant differences according to the ERI quartile, the Kruskal–Wallis test was used for quantitative variables, and the chi-square test for categorical variables.

When significant differences were detected between ERI quartiles, the respective post hoc tests were carried out using the Tukey procedure, when the transformation of the variable to the logarithmic scale enabled normality and homoscedasticity to be achieved, or otherwise using the Dunn's test.

To evaluate the predictive ability of the PLR and the NLR on the ERI, regression models were adjusted in which the log (ERI) was used as a response variable. The models were also adjusted for the age, gender and weight of the individual. In the case of the NLR, to evaluate the particular effect of the neutrophil and lymphocyte counts (which are part of the NLR definition), these variables were introduced into the regression model and a step-by-step method was applied to select the set of variables with the best predictive ability. A similar procedure was carried out with the PLR, introducing in this case the platelets and lymphocytes as additional variables to then apply the step-by-step procedure.

Likewise, in order to carry out an analysis of the ability of the PLR to diagnose erythropoietin resistance, a logistic regression model, adjusted for age and gender, was estimated. A step-by-step method was also applied to simplify the model and to eliminate non-explanatory variables. The calculation of the ROC curve was carried out when only the PLR was used as the diagnostic criterion to predict erythropoietin resistance.

The baseline characteristics of the patients included are described in Table 1. A stratification of the ERI in quartiles (Q1=0.59–4.08, Q2=4.08–7.04, Q3=7.04–12.55, Q4=12.55–80.65) was performed and the values of the different variables among the thus defined groups were compared (Tables 2 and 3). The dry weight, lymphocytes, haemoglobin, concentrations of serum iron, albumin and C-reactive protein showed significant differences among the ERI quartiles. For all these variables (except haemoglobin), the post hoc Tukey tests (with the variables transformed beforehand to the logarithmic scale) showed that the first, second and third quartiles did not display significant differences among them, and that the significant differences occurred with regard to the fourth quartile. In the case of haemoglobin, the Dunn's post hoc test showed that its value is significantly higher in the first quartile of ERI, did not show significant differences between the second and third quartiles and showed a significantly lower value in the fourth quartile.

The correlation between the different study variables and the log value (ERI) was calculated (Table 4). The correlation is significant for all variables apart from age and neutrophils.

Likewise, correlations were carried out between different inflammation and iron kinetics parameters with the log (ERI) (Table 5).

Given the asymmetry of the NLR variable for each ERI quartile, the comparison of the NLR values between the four ERI quartiles was performed initially using the non-parametric Kruskal–Wallis test: this test showed significant differences in the distribution of the ERI for the different ERI quartiles (p=0.00058). To determine the quartiles among which these differences occur, ANOVA and a post hoc test were used. For this, the asymmetry was corrected, transforming the NLR values to a logarithmic scale. The analysis of variance of the log (NLR) confirmed that there are significant differences in the log (NLR) values according to the ERI quartiles (p=0.0004296). The Shapiro–Wilk test did not find evidence of lack of normality in the residuals (p=0.3808) and Levene's test did not detect heteroscedasticity among groups (p=0.4492), which validates the performance of ANOVA (Fig. 1). The Tukey multiple comparison test was used to determine among which quartiles the NLR differences were found: the log values (NLR) for the fourth ERI quartile differ significantly from the values in the first and second quartile. The log value (NLR) in the third ERI quartile does not differ significantly from the others.

Fig. 1.

log(NLR) vs. ERI quartiles.

(0.05MB).

To evaluate the predictive ability of the NLR on the log (ERI), a regression model was adjusted; although the effect of the NLR on the log (ERI) turned out to be significant (standardised beta=0.173, p<0.0001), the R2 value turned out to be very low (R2=0.029), which indicates that, although the ERI value tends to grow with the NLR, the predictive ability of the latter variable is limited. Fig. 2 shows the overall regression line adjusted for all the data, in which the growing trend of the ERI with the NLR is observed and in which the presence of considerable variability can be seen. If we analyse this graph carefully, we can observe that, in particular, the majority of the subjects (90%) have NLR values lower than 6; within this group, only 3.9% are resistant to erythropoietin; subjects with NLR >6 only make up 10% of the total; but within this group there are proportionally more subjects with erythropoietin resistance (18.9%). In other words, the risk of erythropoietin resistance is higher among subjects with high NLR values.

Fig. 2.

Regression analysis between log(ERI) and NLR.

(0.13MB).

This motivates adjusting a logistic regression model to predict the erythropoietin resistance depending on the NLR. With this model, we obtain that the OR of suffering from erythropoietin resistance is 1.32 (p<0.0001; 95% CI: 1.118–1.544) for each unit of increase in the NLR.

As with the NLR, the asymmetry of distributions of the PLR for each ERI quartile led to the comparisons of the PLR according to ERI quartiles being carried out using the Kruskal–Wallis test, with the differences turning out to be statistically significant (p<0.00001) (Fig. 3).

Fig. 3.

log(PLR) vs. ERI quartiles.

(0.08MB).

Similarly, to decide in which quartiles these differences occur, an ANOVA of the log (PLR) was carried out, with the differences being significant (p<0.0001). The normality of the log (PLR) was analysed using the Shapiro–Wilk test (p=0.65) and homoscedasticity with Levene's test (p=0.737).

The Tukey test indicated that the differences were in the fourth ERI quartile compared to the rest. No significant differences in the distribution of the log (PLR) were detected in the other three quartiles.

In the regression adjustment to predict the log (ERI) based on the PLR, a significant positive slope (standardised beta=0.32, p<0.0001) was obtained, although, once again, the R2 value turned out to be low (R2=0.103) although higher than that obtained with the NLR, which means that the predictive ability of the PLR is better than that of the NLR (Fig. 4).

Fig. 4.

Regression analysis between log(ERI) and PLR.

(0.14MB).

As occurred with the NLR, it is observed that the log (ERI) increases with the PLR value; as in that case 90% of the PLR values are below 233; in this group, some 3.9% of patients are resistant to erythropoietin; however, when the PLR is above 233, 17.5% of the subjects are resistant to erythropoietin; therefore, the risk of being resistant to erythropoietin increases with the PLR value.

To assess this risk, the logistic regression model indicates that an increase of one unit in the PLR translates into an OR of 1.011 (p<0.0001; 95% CI: 1.005–1.015) of presenting erythropoietin resistance. An additional calculation makes it possible to estimate that the OR of suffering from erythropoietin resistance doubles for each increase of 67 units in the PLR value.

Furthermore, a multivariate regression analysis was carried out, in which variables such as the TSI, serum iron, C-reactive protein and the PLR were included. The correlation of the PLR with logERI continued to be significant (p<0.0001), with the total predictive capacity of the model, measured in terms of R2, of 17.26%.

To analyse the capacity of the PLR to predict erythropoietin resistance, the area under the ROC curve was measured (AUC=0.681) (95% CI: 0.541–0.821). A PLR cut-off point of 125.5 would result in 80.95% sensitivity and 42.82% specificity (Fig. 5).

Fig. 5.

ROC curve which analyses the capacity of the PLR to predict erythropoietin resistance.

(0.07MB).