A novel predictive score for citrate accumulation among patients receiving artificial liver support system therapy with regional citrate anticoagulation

Patients with liver failure may suffer citrate accumulation when using regional citrate anticoagulation for artificial liver support system therapy (RCA-ALSS therapy). This study aimed to develop a predictive scoring system to stratify the risk of citrate accumulation. A total of 338 patients treated with RCA-ALSS therapy were retrospectively enrolled and randomly divided into derivation and validation cohorts. Longer duration of citrate accumulation (LDCA) was defined as the presence of citrate accumulation 2 h after RCA-ALSS therapy. Four baseline variables were found to be independently associated with LDCA: gender, international normalized ratio of prothrombin time, serum creatinine, and serum chloride. A predictive R-CA model and its simplified R-CA score were developed. The R-CA model (AUROC = 0.848) was found to be superior to the MELD score (AUROC = 0.725; p = 0.022) and other univariate predictors (AUROCs < 0.700; all p ≤ 0.001) in predicting LDCA. The R-CA score (AUROC = 0.803) was as capable as the R-CA model (p = 0.369) and the MELD score (p = 0.174), and was superior to other univariate predictors (all p < 0.05) in predicting LDCA. An R-CA score of 0–2 had a negative predictive value of 90.2% for LDCA. Our R-CA score reliably predicts LDCA in patients with RCA-ALSS therapy, and it is easy to use. Patients with R-CA score of 0–2 can safely receive RCA-ALSS therapy, while others should be carefully evaluated before treatment. Trial registration: Chinese Clinical Trial Registry, ChiCTR2000029179. Registered 17 January 2020, https://www.chictr.org.cn/showproj.aspx?proj=48084.


Results
Patient characteristics. A total of 480 patients treated with ALSS therapy were initially screened and enrolled (Fig. 1). Patients treated with non-DPMAS plus PE therapy (N = 7) or non-RCA (N = 9) were excluded from the study. Patients with liver cancer (N = 18) and those without HBV infection (N = 108) were also excluded. A total of 338 patients were enrolled and randomly divided into a derivation cohort (N = 230) and a validation cohort (N = 108) with a ratio of 2:1 using SPSS software. All patients were followed up 2 h after RCA-ALSS therapy.
The patient characteristics are shown in Table 1. There were no significant differences between the two cohorts in gender, age, causes of liver disease, usage of antiviral agents, or laboratory parameters before the initial ALSS therapy. The Model for End-Stage Liver Disease (MELD) score 17 and the proportion who met the HBV-ACLF criteria 18 were similar in the two cohorts. The overall rates of longer duration of citrate accumulation (LDCA) were not significantly different between the two cohorts. There were no significant differences in indicators representing that patients received similar RCA, such as intracorporeal Ca tot before RCA-ALSS therapy, intracorporeal and extracorporeal Ca ion during RCA-ALSS therapy, and intracorporeal Ca tot and Ca ion 2 h after RCA-ALSS therapy (Table 1).

Figure 1.
Flow diagram of patient selection. Of the 480 patients in our database that received ALSS therapy, 142 were excluded from the study. The remaining 338 patients were randomly divided into a derivation cohort (N = 230) and a validation cohort (N = 108). ALSS artificial liver support system, DPMAS double plasma molecular adsorption system, PE plasma exchange, RCA regional citrate anticoagulation, HBV hepatitis B virus. Table 1. Characteristics of derivation and validation cohorts. Measurement data are represented as mean ± SD (normally distributed data) or median (IQR) (non-normally distributed data). Enumeration data are represented as frequencies (proportion). MELD Model for End-Stage Liver Disease, HBV hepatitis B virus, ACLF acute-onchronic liver failure, PT-INR international normalized ratio (INR) of prothrombin time (PT), ULN upper limit of normal, RCA-ALSS therapy artificial liver support system therapy with regional citrate anticoagulation, Ca tot total calcium, Ca ion ionized calcium, Ca tot /Ca ion Ca tot to Ca ion ratio, R-CA model logistic regression model of risk predictors for citrate accumulation, R-CA score risk score for citrate accumulation, LDCA longer duration of citrate accumulation, RCA regional citrate anticoagulation, ALSS therapy artificial liver support system therapy. # Statistical analysis using relative values adjusted by gender (a multiple of upper limit of normal). Testing of R-CA model in validation cohort. Before developing a simplified predictive score, the four independent predictors were tested in the validation cohort. Gender, PT-INR, serum creatinine, and serum chloride were verified as independent predictors of LDCA based on the results of the multivariate analysis of the validation cohort ( Table 3). The R-CA model showed a good predictability with AUROC of 0.848 (95% CI 0.795-0.892, p = 0.000) and 0.856 (95% CI 0.776-0.916, p = 0.000) in the derivation and validation cohorts, respectively. The expected LDCA rates and observed LDCA rates from the derivation cohort (R 2 = 0.909, p = 0.000) matched with the validation cohort (R 2 = 0.778, p = 0.007; Fig. 2A).
Development of R-CA score. Following the validation of our predictive model, the four individual parameters were scored. An ordinal grading system (0-2) with distinct hazard ratios (HR) on logistic regression was performed by comprehensively considering their cut-off values of AUROCs predicting the probability of LDCA, ACLF diagnostic criteria 19,20 , and clinically significant values (Tables 4, 5). The total R-CA score ranges from a minimum of 0 to a maximum of 8. The scoring parameters are easy to collect laboratory measurements or clinical features that showed a distinct hazard ratio on logistic regression in derivation and validation cohorts (all HR > 2 and all p = 0.000; Table 6).  www.nature.com/scientificreports/ The R-CA scores and LDCA rates showed a linear correlation in the derivation cohort (R 2 = 0.912, p = 0.000; Fig. 3). A linear regression equation was developed: LDCA rate = 12.8% × R-CA score − 1.2%. The expected LDCA rates and observed LDCA rates based on the R-CA scores in the derivation cohort (R 2 = 0.845, p = 0.000) matched those of the validation cohort (R 2 = 0.842, p = 0.000; Fig. 2B).

Evaluation of R-CA model and R-CA score as predictors of LDCA in derivation and validation cohorts.
Our R-CA model and its simplified R-CA score were compared with other potential predictors, such as the MELD score, PT-INR and lactate (Fig. 4A,B and Table 7). AUROCs of the R-CA model and the R-CA score in the derivation cohort were 0.848 and 0.803, and those in the validation cohort were 0.856 and 0.816, respectively. Our R-CA model was found to be as capable as the R-CA score, and superior to the MELD score (AUROC = 0.725) and other univariate predictors (AUROCs < 0.700), in predicting LDCA (p = 0.369, p = 0.022 and p ≤ 0.001, respectively). The R-CA score was as capable as the MELD score in predicting LDCA (p = 0.174) and superior to other univariate predictors in predicting LDCA (p < 0.05).
Less than 10% of patients who had an R-CA model ≤ − 1.00 or an R-CA score of 0-2 experienced LDCA. R-CA model ≤ − 1.00 had an AUROC of 0.848, a sensitivity of 74.1%, a specificity of 88.6%, a positive predictive value of 66.7%, and a negative predictive value of 91.8%. R-CA score of 0-2 had an AUROC of 0.803, a sensitivity of 70.4%, a specificity of 84.1%, a positive predictive value of 57.6%, and a negative predictive value of 90.2% (Table 8). Although patients with MELD score ≤ 26.6, PT-INR ≤ 2.43 and lactate ≤ 2.65 had high negative predictive values of LDCA, their AUROCs were all less than 0.750 ( Table 8).
Evaluation of R-CA model and R-CA score as predictors of LDCA among patients with or without liver cirrhosis. The R-CA model and R-CA score also showed a distinct hazard ratio on logistic regression among patients with or without liver cirrhosis (all HR > 2 and all p = 0.000; Table 6). The expected LDCA rates and observed LDCA rates based on the R-CA model among patients with liver cirrhosis (R 2 = 0.951, AUROCs of the R-CA model and the R-CA score among patients with liver cirrhosis were 0.851 and 0.804, and those among patients without liver cirrhosis were 0.836 and 0.781, respectively (Fig. 4C,D, and Table 9). Our R-CA model was found to be as capable as the R-CA score, and superior to the MELD score (AUROC = 0.712) and other univariate predictors (AUROCs < 0.700), in predicting LDCA (p = 0.280, p = 0.005 and p = 0.000, respectively). The R-CA score was as capable as the MELD score in predicting LDCA (p = 0.075) and superior to other univariate predictors in predicting LDCA (p < 0.01).
About 12.5% of cirrhotic patients who had an R-CA model ≤ − 1.39 or an R-CA score of 0-2 experienced LDCA. R-CA model ≤ − 1.39 had an AUROC of 0.851, a sensitivity of 76.5%, a specificity of 80.5%, a positive predictive value of 61.2%, and a negative predictive value of 89.5%. R-CA score of 0-2 had an AUROC of 0.804,   Correlation between R-CA model, R-CA score, LDCA and disease severity. R-CA model, R-CA score, LDCA, and Ca tot /Ca ion 2 h after RCA-ALSS therapy were positively correlated with disease severity rated by MELD score in derivation and validation cohorts and among patients with or without liver cirrhosis with all the p < 0.01 (Fig. 6, Table 10).

Discussion
We developed a predictive model of citrate accumulation for patients with liver failure treated with RCA-ALSS therapy. We found that four variables (gender, PT-INR, serum creatinine, and serum chloride) are independent predictors of LDCA. A predictive R-CA model and its simplified R-CA score have both been developed. Prior to our study, the existing predictors in the literature were all univariate, such as PT-INR and lactate 8,16 . There was no score dedicated to LDCA in patients with liver failure treated with RCA-ALSS therapy before our study.
Our R-CA model was constructed based on the four most significant variables found in our multivariate analysis. Female gender, serum creatinine, and PT-INR were positively correlated with LDCA, while serum chloride was negatively correlated. It has been reported that the metabolism of citrate is mainly in the mitochondria of liver, muscle, and kidney cells 21 . This helps to explain the positive correlations between LDCA and gender, serum creatinine, and PT-INR. LDCA may be more likely to occur in female patients because, on average, men have more muscle mass than women, allowing for more citrate metabolism in muscle cells 22 . As serum creatinine is a marker for renal function, it is expected that both citrate and serum creatinine levels will be high if metabolism in the kidneys is slowed. Additionally, as coagulation failure is a symptom of ACLF, and PT-INR values have been correlated with the degree of liver injury, which is related to the function of liver cells' mitochondria, a high PT-INR indicates a lower ability of liver cells to metabolize citrate 1 . The negative correlation between serum chloride and LDCA may be explained by the anion gap. Chloride is one of the most important anions in the body, and its decrease may indicate the increase of other anions, such as lactate, to maintain the anion gap. The correlation coefficient between serum chloride and lactate is − 0.173 (p = 0.001) (Data not shown). Our previous study reveals that baseline Ca tot /Ca ion is an independent predictor for citrate accumulation in patients with ACLF who received DPMAS plus PE therapy with heparin anticoagulation (AUROC = 0.725) 16 . However, the role of the anion gap in predicting LDCA requires further investigation.
We found no correlation between LDCA and standard laboratory liver function parameters, which is in line with the results of previous studies 9, 12 . We predicted that age may be negatively correlated with LDCA, due to decreased muscle mass and weakened mitochondria 23,24 , but that is not supported by our study.
Our R-CA model is based on some of the same parameters used to determine the MELD score (serum creatinine and PT-INR), which has been used for organ allocation in patients with end-stage liver disease 17 . We found an AUROC of 0.725 when using the MELD score to predict LDCA, which is inferior to our R-CA model. Our results did not support lactate and PT-INR as predictive factors, though they have been reported in previous studies 8, 16 . Although our R-CA model is able to successfully predict LDCA, complex calculations are required, much like when calculating a MELD score 17 . To make the model more useful, we developed the R-CA score, a simplified version that does not require complex variables and can be used easily at the patient's bedside. The R-CA score is constructed based on the four most significant and readily available variables found in our multivariate analysis, and it is as capable as our R-CA model and the MELD score in predicting LDCA. Thus, www.nature.com/scientificreports/ it has a distinct advantage over our R-CA model and the MELD score. The simple scoring system assigns 0-2 points for each of the four parameters, much like the Child-Turcotte-Pugh (CTP) score for patients with liver cirrhosis and the CLIF-C ACLF score for patients with liver failure 25,26 . The cutoff values were determined using AUROC, ACLF diagnostic criteria 19,20 , and clinically significant values. The R-CA score was verified using a validation cohort, and found to be successful in predicting LDCA (AUROC = 0.803 and 0.816 in the derivation and validation cohorts, respectively). The CLIF-C ACLF score, one of the most important but easy-to-use prognostic scores for disease severity in patients with ACLF 26 , probably also has the capability in predicting LDCA because of some of the same parameters (serum creatinine and PT-INR). We had tried to compare the CLIF-C ACLF score to our score, but the respiratory indicator (arterial blood gas analysis result) was missing in most cases in this retrospectively study. The role of CLIF-C ACLF score in predicting LDCA requires further investigation.
In critically ill patients with acute kidney injury without liver disease treated with RCA-CVVHD, ROC data supports lactate (cutoff value 2.39 mmol/L) as a strong negative predictive factor of citrate accumulation, with a predictive value of 99.28% 27 . In our study, lactate has a negative predictive value of 82.1% at a cutoff of 2.65 mmol/L, but its AUROC is only 0.627, and it was not found to be an independent risk factor. These results www.nature.com/scientificreports/ agree with those of our previous study 16 . Similarly, the MELD score and PT-INR also showed high negative predictive values but had low AUROCs (AUROCs < 0.750). In our study, an R-CA score 0-2 predicted the absence of LDCA with a negative predictive value of 90.2%. Approximately 70% of the patients in this study had an R-CA score 0-2. In addition to predicting LDCA, the R-CA score may have the potential to guide initial anticoagulant prescriptions in the future. Patients with an R-CA score 0-2 can safely receive RCA, but the initial dosage should be reduced, and intervention may be needed, in patients with an R-CA score of 3-8. Automated anticoagulant technology may provide a user-friendly and safe system for patients with liver failure treated with RCA-ALSS therapy in the future 28 .
In this study, we found that R-CA model, R-CA score, and LDCA were positively correlated with disease severity rated by MELD score. As the R-CA model and R-CA score are based on some of the same parameters as the MELD score 17 , the CLIF-C ACLF score 26 , and the ACLF criteria 19,20 , and an elevated MELD score, CLIF-C ACLF score, and Ca tot /Ca ion being associated with increased mortality in patients with liver failure 17,26,29 , the R-CA model, R-CA score, and LDCA could also have the ability to predict patients' prognosis. The direct relationship between these values and patients' prognosis requires further investigation.
Our study has several limitations. First, as a retrospective study with a small number of patients from a single geographic area, the patient characteristics may not represent the general population, which may have affected the results of our study. Second, the patients in our study all had chronic HBV infections presenting as acute hepatic decompensation. Data derived from this subset may not be applicable to all patients with liver disease, and requires further investigation. Lastly, we used the Ca tot /Ca ion ratio instead of a direct measurement of plasma citrate concentration to reflect citrate accumulation. The upper normal and toxic levels of citrate in the blood are not well established, and as citrate is a physiological metabolite, it may not be toxic, but instead   www.nature.com/scientificreports/  www.nature.com/scientificreports/ lead to metabolic lag when it accumulates 8,15 . Further investigation on normal citrate levels, as well as the effect of citrate accumulation in the plasma, is needed.

conclusions
Our R-CA model, based on gender, PT-INR, serum creatinine and serum chloride, reliably predicts LDCA in patients with HBV-ACLF who are undergoing RCA-ALSS therapy. The simplified R-CA score is also reliable, and easy to use. Patients with an R-CA score 0-2 can safely receive RCA-ALSS therapy, while others should be MELD Model for End-Stage Liver Disease, R-CA model logistic regression model of risk predictors for citrate accumulation, R-CA score risk score for citrate accumulation, LDCA longer duration of citrate accumulation, Ca tot total calcium, Ca ion ionized calcium, Ca tot /Ca ion Ca tot to Ca ion ratio, RCA regional citrate anticoagulation, ALSS therapy artificial liver support system therapy. RCA-ALSS therapy. All patients received standard medicinal treatment and DPMAS plus PE therapy. The composition of DPMAS plus PE therapy was the same as that previously described (Fig. 7) 15,16 . All patients received DPMAS therapy for 2 h, followed immediately by PE therapy with half the total plasma volume (approximately 1,500 mL) for approximately 1 h. As shown in Fig. 7, the parameters of the CRRT machine were set to a blood flow of 130 mL/min, a plasma separation flow of 1,500 mL/h, and a plasma return flow of 1,500 mL/h. Patients received 4% citrate serum sodium anticoagulation with an initial speed of 120 mL/h in the arterial segment during DPMAS plus PE therapy and received 10% calcium gluconate supplement in the venous segment with a speed of 13 mL/h during DPMAS therapy and 74 mL/h during PE therapy.
Definition. Under normal conditions, citrate's half-life is approximately 5 min and citrate is metabolized completely within 30 min of discontinuing a citrate infusion 31,32 . Here, we defined LDCA as the presence of citrate accumulation 2 h after RCA-ALSS we defined longer duration of citrate accumulation (LDCA) as the presence of citrate accumulation 2 h after RCA-ALSS therapy concluded. Citrate accumulation was defined as the value of the ratio of Ca tot to Ca ion greater than or equal to 2.5 (Ca tot /Ca ion ≥ 2.5) 8,33 .

Figure 7.
Schematic diagram of the artificial liver support system therapy with regional citrate anticoagulation. Circuit 2 therapy is initiated immediately when circuit 1 therapy is completed.
Scientific RepoRtS | (2020) 10:12861 | https://doi.org/10.1038/s41598-020-69902-2 www.nature.com/scientificreports/ Statistical analysis. Patients were randomly divided into a derivation cohort and a validation cohort using SPSS software (IBM SPSS) with a ratio of 2:1. The t test and the U test were performed for quantitative data of normal distribution and that of non-normal distribution, respectively. The chi-squared test or Fisher's exact test was performed to calculate differences between qualitative data. The predictors for LDCA in the derivation cohort were analyzed by logistic regression in univariate analysis. For any variables with p ≤ 0.1 in the univariate analysis, the backward stepwise (likelihood ratio) method was performed in a multivariate analysis. The predictors obtained from the derivation cohort were then tested in the validation cohort. The predictive model was also tested in the validation cohort. Predictive factors with an AUROC > 0.750 in the derivation cohort that was equivalent or greater in the validation cohort was used to derive our predictive R-CA model. With the purpose of deriving a simple, specific predictive score for patients treated with RCA-ALSS therapy, we included clinically relevant characteristics and laboratory parameters observed at baseline. An ordinal grading (0-2) was performed for individual parameters by comprehensively considering their cut-off values of AUROCs predicting the probability of LDCA, ACLF diagnostic criteria 19,20 , and clinically significant values. A score was obtained by combining the individual grade of all the significant parameters. Multiple comparisons of the score with other predictors were performed by AUROC. Statistical significance was set at p < 0.05. The statistical tests were performed using SPSS v.24 (IBM SPSS), except multiple comparisons of AUROCs, which were performed using MedCacl v.19 (MedCalc Software). The figures of AUROCs and linear regression lines were drawn using Med-Cacl v.19 (MedCalc Software) and GraphPad Prism 6 (GraphPad Software Inc.), respectively.

Data availability
The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.