Association between intracatheter Ca²+ trajectory and extracorporeal circuit coagulation in unplanned early withdrawal among patients receiving CRRT: a retrospective cohort study

Abstract

Objective

This study aims to identify the trajectory patterns of intracatheter ionised calcium (Ca²⁺) in Intensive Care Unit (ICU) patients who experienced unplanned early withdrawal of citrate-anticoagulated continuous renal replacement therapy (CRRT) due to extracorporeal circuit coagulation and to analyse the association between these trajectories and coagulation-induced filter lifespan reduction within this selected population.

Design

Retrospective cohort study.

Setting

3 ICU wards of a university-affiliated tertiary teaching hospital in Wuhan, China.

Participants

A total of 732 ICU patients who underwent citrate-anticoagulated CRRT and experienced unplanned early withdrawal between January 2022 and December 2023 were included. Patients who completed CRRT as planned were not included. The study population therefore represents a selected subgroup already enriched for premature circuit interruption rather than the general population of patients receiving CRRT with regional citrate anticoagulation.

Main outcome measures

The primary outcome was the occurrence of extracorporeal circuit coagulation. The secondary outcome was extracorporeal circulation lifespan (filter lifespan).

Results

intracatheter Ca²⁺ trajectories were categorised into three types: low baseline with stable trend, moderate baseline with slow decline and high baseline with rapid decline. Fibrinogen, lactic acid and nurses’ professional titles showed significant differences among the three trajectory groups. Cox regression revealed that the high baseline with rapid decline type was the high-risk group for coagulation-induced filter lifespan reduction (HR=1.0), with the moderate baseline with slow decline type (HR=0.633 (95% CI 0.425 to 0.942), p=0.024) and low baseline with stable trend type (HR=0.534 (95% CI 0.289 to 0.986), p=0.045) having lower hazards. Multivariable analysis identified factors associated with circuit coagulation, including Body Mass Index, Sequential Organ Failure Assessment score, Richmond Agitation-Sedation Scale (RASS) score, use of lipid-lowering drugs and blood-product transfusion. Higher citrate flow rate and lower transmembrane pressure were variables associated with a lower hazard of coagulation-driven filter failure. Several of these variables were measured during ongoing CRRT and may reflect downstream changes related to evolving circuit status rather than baseline causal determinants.

Conclusions

Heterogeneous intracatheter Ca²⁺ trajectories exist in patients receiving CRRT with unplanned early withdrawal due to circuit coagulation, and the high baseline with rapid decline trajectory is an independently associated factor for coagulation-induced filter lifespan reduction. These hypothesis-generating findings provide preliminary observational insights for optimising citrate anticoagulation management in CRRT; the observed intracatheter Ca²⁺ range among low-risk patients requires prospective external validation before any clinical implementation.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• This study employed latent class growth modelling to identify distinct intracatheter ionised calcium (Ca²⁺) trajectory patterns over four time points, moving beyond single-point measurements.

• The retrospective cohort design included a relatively large sample of 732 ICU patients from a tertiary teaching hospital, with strict inclusion and exclusion criteria.

• Cox proportional hazards regression was used to stratify coagulation risk based on Ca²⁺ trajectories while adjusting for potential confounders.

• As a retrospective single-centre study, the findings may be subject to selection bias and residual confounding, and the generalisability to other clinical settings is limited.

• The study did not account for the potential impact of dialyser membrane materials, catheter maintenance protocols or fluctuations in citrate dose during long-term continuous renal replacement therapy on circuit coagulation.

Introduction

Continuous renal replacement therapy (CRRT) is widely used in critically ill patients with acute kidney injury and chronic renal failure.^{1 2} Coagulation in the extracorporeal circuit of CRRT is a major cause of premature withdrawal, with an incidence rate as high as 40%–60%, which can lead to complications such as blood loss and activation of the coagulation response.^{3 4} Therefore, preventing coagulation in the extracorporeal circuit of CRRT is of great significance for maintaining the efficacy of CRRT and preventing coagulation-related complications. Currently, the mainstream anticoagulation methods for CRRT include regional citrate anticoagulation (RCA) and heparin anticoagulation. Compared with heparin anticoagulation, RCA has advantages such as longer circuit life and less systemic bleeding.^{5 6} The Kidney Disease: Improving Global Outcomes (KDIGO) guidelines⁷ and the Chinese CRRT anticoagulation guidelines⁸ both recommend that, if there are no contraindications to citrate, RCA should be the first choice.

In recent years, due to issues such as the supply of calcium-free treatment fluids and complex calcium supplementation protocols,⁹ clinical practice has largely relied on monitoring changes in calcium ions to determine the body’s calcium levels. Citrate anticoagulation works by chelating ionised calcium (Ca²⁺) in the extracorporeal circuit, thereby reducing Ca²⁺ concentration and effectively inhibiting coagulation. Currently, most studies suggest that maintaining a target range of 0.25–0.35 mmol/L for catheter Ca²⁺ concentration can achieve optimal anticoagulant effects.^{7 10–12} This range is also widely used as a reference in clinical settings. Therefore, the trajectory of changes in catheter calcium ion concentration can reflect the anticoagulant effect of citrate. However, this conclusion is based on small sample studies.^{13 14} An increasing number of studies have found that the control range for catheter Ca²⁺ concentration does not need to be too strict; a range of 0.35–0.50 mmol/L still maintains good anticoagulant effects.^14–17 Nevertheless, excessive fluctuation or persistent deviation from the target range in catheter calcium ion concentration may indicate an increased risk of coagulation. For instance, lower target values for catheter Ca²⁺ often require more citrate supplementation in the extracorporeal circuit, which can increase the risk of citrate accumulation and metabolic alkalosis and also place an unnecessary workload on medical staff.^{16 17} By analysing the trajectory of calcium ion changes, it is possible to predict the likelihood of filter coagulation and take pre-emptive actions to reduce the incidence of unplanned withdrawals. Additionally, in clinical practice, the trajectory of calcium ion changes can guide the infusion rate of citrate, ensuring patient safety through timely detection and management of calcium ion abnormalities. Simultaneously, analysing the trajectory of calcium ion changes can help tailor more appropriate anticoagulation plans for each patient, reducing unplanned withdrawals due to coagulation.¹⁶

Therefore, this retrospective study investigated the classification of intracatheter Ca²⁺ variation trajectories in patients receiving CRRT under citrate anticoagulation, analysed their associations with extracorporeal circuit coagulation and related outcomes, including circuit lifespan, and identified the relevant factors associated with them. It is intended to provide scientific evidence for ICU clinicians to optimise the anticoagulation management of patients receiving CRRT and reduce the incidence of unplanned premature withdrawal in RCA-CRRT.

Methods

Study aim

This study aims to explore the associations between intracatheter Ca²⁺ variation trajectories and extracorporeal circuit coagulation in ICU patients with unplanned early withdrawal of citrate-anticoagulated CRRT and to identify the factors associated with such coagulation events. The findings are expected to provide evidence for optimising clinical anticoagulation management, prolonging extracorporeal circuit lifespan and reducing the incidence of unplanned premature CRRT withdrawal.

Study design

A retrospective cohort study was conducted in three ICU wards of a university-affiliated tertiary teaching hospital in Wuhan, China. The CRRT devices used included Prismaflex (Gambro Industries, France), multiFiltrate (Fresenius Kabi, Germany) and Diapact (BRAUN, Germany). Dialysers were either AN69 membrane hollow-fibre (Gambro Industries, France) or polysulfone membrane hollow-fibre (AV1000S/AV600S, Fresenius Kabi, Germany). A dual-lumen 11.5 Fr catheter (ABLE, BAIHE, Guangdong, China) was adopted for vascular access in all patients.

CRRT protocol and citrate anticoagulation regimen

CRRT management was performed by a multidisciplinary nursing team, with treatment decisions (initiation, filter selection and discontinuation/restart) made by a renal critical care group in collaboration with critical care specialists. Dialysis nurses were responsible for circuit setup and regular evaluation, and ICU nurses for bedside CRRT operation.

The standard CRRT mode was continuous venovenous haemodiafiltration on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) platform, with oXiris or M100 filters. The blood flow rate was 80 mL/min for extraction and 120–150 mL/min for treatment, with a prescribed therapeutic dose of 30–35 mL/kg/hour. The circuit was pre-flushed with heparin and normal saline for two cycles each. RCA was initiated with 4% citrate solution (Chengdu Qingshan Likang Pharmaceutical) at 3.0–4.0 mmol/L. Replacement fluid and dialysate were administered at a 1:1 ratio via post-dilution, combined with 5% bicarbonate solution at 80 mL/hour. The ultrafiltration rate was adjusted for fluid balance, and the circuit was intended for 72 hours of operation unless filter clogging occurred or CRRT was no longer indicated. The anticoagulation targets were peripheral ionised Ca²⁺ 1.0–1.2 mmol/L, serum total Ca²⁺ 2.2–2.7 mmol/L and intracatheter ionised Ca²⁺ 0.25–0.35 mmol/L. Figure 1 shows the operational diagram of CRRT under citrate anticoagulation in this study.

Figure 1. Diagram showing continuous renal replacement therapy operation under citrate anticoagulation. Low starting point with stable development type (average estimated baseline value range: 0.296–0.311 mmol/L); middle starting point with slow decline type (average estimated baseline value range: 0.353–0.379 mmol/L) and high starting point with rapid decline type (average estimated baseline value range: 0.382–0.489 mmol/L).

Outcomes and definitions

The primary outcome was the occurrence of extracorporeal circulation coagulation events. According to the Chinese Blood Purification Operation Standards, extracorporeal circulation coagulation is defined as follows¹⁸: (1) dialyser coagulation: no coagulation or several fibres coagulated (Grade 0); <10% fibres coagulated (Grade I); 10%–50% fibres coagulated (Grade II) and >50% fibres coagulated (Grade III). Grade II and above are defined as circuit coagulation; and (2) venous pot coagulation: coagulation occurring in over one-third of the venous pot is defined as extracorporeal circulation coagulation. The coagulation grade is classified based on the extent of coagulation in the venous pot: 1/3 coagulated is Grade 1; 1/3–2/3 coagulated is Grade 2 and >2/3 coagulated is Grade 3. The presence of either type of coagulation meeting the respective criteria constitutes a judgement of extracorporeal circulation coagulation.

Secondary outcome: extracorporeal circulation lifespan: the time from the start of CRRT treatment to the end of CRRT treatment.

Unplanned early withdrawal was defined as CRRT termination occurring before the intended maximum treatment duration of 72 hours. This 72-hour threshold was selected based on local institutional clinical practice, in which RCA-CRRT circuits are routinely planned for up to 72 hours of use unless filter malfunction or loss of treatment indication occurs. The 72-hour cut-off also aligns with prior published CRRT observational studies focusing on premature circuit loss. Nevertheless, this definition excludes coagulation-related filter failures arising after 72 hours of treatment and may therefore under-represent events occurring in patients receiving prolonged CRRT support.

Participants and setting

The study included patients who underwent citrate-anticoagulated CRRT in the hospital’s ICU from January 2022 to December 2023, with data finalised in March 2025. A retrospective, consecutive sampling approach was employed. All patients who received CRRT in the three ICU wards during the study period were identified through the hospital’s electronic medical record system and the CRRT treatment logbook. The detailed sampling and screening process, including the stepwise application of inclusion and exclusion criteria, is presented in figure 2.

Figure 2. Flowchart of study subject screening and enrolment. CRRT, continuous renal replacement therapy; RCA, regional citrate anticoagulation.

Inclusion criteria

Aged ≥18 years; treated with citrate anticoagulation alone; experienced unplanned early CRRT withdrawal; only data from the first CRRT circuit filter episode during the patient’s first ICU admission were included; subsequent repeated CRRT circuits from the same patient were excluded from analysis; and only first ICU admission data for multiple admissions.

Exclusion criteria

Missing key monitoring data; concurrent CRRT and Extracorporeal Membrane Oxygenation (ECMO) treatment; pregnancy and receipt of systemic anticoagulation, antiplatelet therapy or thrombolysis during CRRT.

This study specifically focused on patients who experienced unplanned early withdrawal of citrate-anticoagulated CRRT. This population was selected because these patients represent the clinically relevant subgroup in which circuit coagulation has already occurred, and understanding the dynamic Ca²⁺ trajectory patterns preceding these events may provide insights for risk stratification and preventive management. However, we acknowledge that this selection means the findings characterise trajectory patterns within a subgroup already enriched for premature interruption events, rather than predicting coagulation risk in the overall population of patients receiving RCA-CRRT.

The target population of this study was ICU patients receiving RCA-CRRT who experienced unplanned early withdrawal. The estimand was the association between intracatheter Ca²⁺ trajectory class membership and (1) coagulation-related versus non-coagulation-related unplanned early withdrawal and (2) extracorporeal circuit lifespan within this population. Patients who completed CRRT as planned were not included because the T3 time point (the most recent measurement before premature shutdown) is only defined in patients who experience unplanned withdrawal, and including them would have required a non-equivalent definition of the final time point.

Although this study is a retrospective cohort study, a sample size calculation was still performed to ensure that the sample met the basic requirements for analysis and to guarantee the accuracy of the results. Based on the longitudinal study sample size estimation table proposed by Barcikowski and Robey,¹⁹ with an intraclass correlation coefficient (I-IB) set at 0.80, data were collected four times for each subject, the average correlation coefficient was taken as 0.50, the effect size was set at 0.05 and the type I error probability was set at 0.05. A total of 142 patients were required to be included. According to the multivariate analysis requirement, the sample size should be 5–10 times the number of independent variables.²⁰ This study involved a total of 58 independent variables for analysis.

$${\displaystyle {\displaystyle \text{CR}=\frac{{\text{Q}}_{\text{UF}}}{{\text{Q}}_{\text{B}}+{\text{Q}}_{\text{R}}^{\text{PRE}}}}}$$

$${\displaystyle {\displaystyle \text{FF}=\frac{{\text{Q}}_{\text{UF-net}}+{\text{Q}}_{\text{Pre}}}{{\text{Q}}_{\text{B}}\times 60\times (1-\text{Hct})+{\text{Q}}_{\text{Pre}}}}}$$

When expanded by 10 times and considering a 20% dropout rate, at least n=696 patients were required to be included. The final sample size included in this study was 732 patients.

Data extraction

The following data were extracted from the patients using a self-designed general information form:

(1) The following data were extracted from the patients regarding the patient-related aspects: gender, age, total CRRT treatment duration, ICU length of stay, total hospitalisation duration, Body Mass Index (BMI), reason for CRRT treatment, Acute Physiology and Chronic Health Evaluation II score within 24 hours of ICU admission, Sequential Organ Failure Assessment (SOFA) score, sedation score (RASS), whether there was a surgery of Grade 2 or above within 1 month (Grade 2 surgery defined according to the ‘Measures for the Administration of Surgery Grading in Medical Institutions’ formulated by the National Health Commission of China),²¹ CRRT catheterisation site, whether anticoagulant drugs were used additionally, whether lipid agents were used, whether blood products were transfused during CRRT, laboratory examination indicators and other disease-related information and filtration fraction (FF). FF is the ratio of ultrafiltration rate to plasma flow rate. A higher FF can improve solute clearance efficiency but will also increase the coagulation risk in the extracorporeal circuit.²² Concentration Ratio (CR) is another indicator for evaluating the coagulation risk of the filter during CVH. The calculation formula for FF²³ and the calculation formula for CR are as follows:

In the FF formula, Q_UF-net represents the net ultrafiltration rate (L/hour), Q_Pre represents the predilution fluid rate (L/hour), Q_B represents the blood flow velocity (L/min) and Hct represents the haematocrit. It is worth noting that the predilution fluid rate is already included in the FF calculation.

In the CR formula, Q_B represents blood flow velocity, Q_R^Pre represents pre-dilution fluid flow rate and Q_UF represents ultrafiltration flow rate.

(2) Nurse-related aspects included age, years of work experience, professional title and whether they have a valid CRRT training certificate.

(3) Aspects related to the CRRT machine included machine type, filter lifespan, dialyser coagulation grade, arterial/venous pot coagulation grade, whether blood was returned during CRRT shutdown, the worst ultrafiltration rate during the entire CRRT treatment period, the citrate flow rate when the extracorporeal circulation lifetime stops treatment, arterial pressure, venous pressure and transmembrane pressure.

(4) The time points for collecting Ca²⁺ in the CRRT circuit are as follows:

• T0: the index value monitored immediately after the CRRT machine is started.

• T1: 2 hours after the CRRT machine is started.

• T2: 4 hours after the CRRT machine is started.

• T3: the intracatheter Ca²⁺ value measured at the most recent time point before the premature shutdown of CRRT. Because this measurement is obtained in close temporal proximity to the coagulation event, it may reflect the ongoing process of circuit failure rather than serving exclusively as an early predictive marker.

Quality control

Uniform training was provided to all data collectors to ensure strict adherence to inclusion/exclusion criteria. Data collection forms were completed item-by-item and double-checked to avoid missing values. All data were coded, sorted and double-entered by two independent researchers to ensure accuracy.

Missing value

Variables with a missing value rate greater than 40% were excluded. The missing value rates of triglycerides and calcium gluconate were greater than 40%; therefore, these variables were excluded from the study. The overall missing data rate for all variables was 0.642%. In our data, only Antithrombin III (AT III) and Fibrin Degradation Products (FDP) had relatively high amounts of missing data, with missing rates of 16.7% and 11.1%, respectively.

Missing data were handled using multiple imputation by chained equations, and 20 imputed datasets were generated. For all regression models and survival analyses, the results were pooled across the 20 imputed datasets using Rubin’s rules, which appropriately incorporate both within-imputation and between-imputation variance.

Because the latent class growth modelling (LCGM) procedure in Mplus (V.8.3) does not readily support pooling across multiply imputed datasets in the same manner as standard regression procedures, the LCGM analyses were performed on a single imputed dataset. This dataset was selected as the one with the smallest mean squared error among the 20 generated datasets, consistent with previously described approaches.^{24 25} We acknowledge that this approach does not fully incorporate imputation uncertainty into the trajectory classification, and this limitation is discussed in the Limitations section.

Data analysis

We used SPSS (V.24.0) to express the descriptive results of participants’ baseline characteristics as means with SD or medians with IQRs, depending on the histogram and normal curve, as well as the normality test:Kolmogorov-Smirnov test (K-S) normality test and Shapiro-Wilk test (if the p value was greater than 0.05, the data were considered to be normally distributed). For normally distributed and approximately normally distributed measurement data, it is expressed as 'xˉ' :Sample mean; s:Sample standard deviation (xˉ±s), and comparisons between multiple groups were made using one-way analysis of variance. Pairwise comparisons between groups were performed using Least Significant Difference (LSD) t-tests or Tamhane’s T2 tests. Skewed distribution measurement data are expressed as M (P₂₅, P₇₅), and comparisons between groups were made using non-parametric rank-sum tests. Categorical data are expressed as n (%), and comparisons between groups were made using the χ² test.

Using Mplus (V.8.3) software, a latent class growth model was employed to explore longitudinal trajectory types and characteristics of Ca²⁺ levels in the catheter of patients receiving RCA-CRRT (monitored four times). The premise of LCGM is the homogeneity of the study subjects. Considering the differences among the subjects comprehensively, the LCGM was used to explore the developmental trajectory of Ca²⁺ levels in different categories of patients receiving RCA-CRRT. The model construction started with the baseline model and achieved growth through a single category. As the number of categories increased, various models were fitted and evaluated to determine the optimal model based on these evaluations. The model evaluation criteria included²⁶ degrees of freedom (df (K)), Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC) and sample-size-adjusted BIC (aBIC). The smaller the values of these statistics, the better the model fit. When the value of aBIC is smaller, it indicates a better model fit. Therefore, the model with the smallest BIC is usually selected as the best model. Entropy is used to evaluate classification accuracy, with a general range of 0–1. An entropy value greater than 0.70 indicates good classification accuracy, and the closer the entropy is to 1, the higher the classification accuracy. The model comparison was performed using the Likelihood Ratio Test (Lo-Mendell-Rubin (LMR)) and the Bootstrapped Likelihood Ratio Test (BLRT). Significant LMR p values and BLRT p values (p<0.05) indicate that a k-category model is superior to a k−1 category model. Furthermore, individuals’ category membership was determined based on category probabilities, which should be greater than 5% of the total population.²⁷

Latent class growth models were fitted with linear and quadratic growth terms for the four time points (T0–T3). Given that only four measurement occasions were available, higher-order polynomial terms were not considered. All models were estimated in Mplus (V.8.3) using maximum likelihood with robust SEs. To avoid convergence on local maxima, 500 random sets of starting values and 50 final-stage optimisations were used. Convergence was confirmed by the replication of the best log-likelihood value across multiple starting value sets. Model selection was based on converging lines of evidence, including AIC, BIC, aBIC, entropy, LMR, BLRT, class size (all classes >5% of the total population) and clinical interpretability. The average posterior class-membership probabilities for the most likely class and class sizes are reported in the Results section.

Moreover, we used multivariate regression to assess the relationship between the trajectory of Ca²⁺ levels in the CRRT circuit and the occurrence of unplanned premature machine withdrawal due to coagulation, as well as the impact of basic characteristics on the trajectory groups of Ca²⁺ levels in the CRRT circuit. The Kaplan-Meier analysis was also used to evaluate the relationship between the classification based on the trajectory of Ca²⁺ levels in the CRRT circuit and the extracorporeal circulation lifespan of CRRT. The time origin was defined as the initiation of CRRT. The event of interest was the first occurrence of extracorporeal circuit coagulation leading to unplanned premature withdrawal, as defined in the Outcomes and definitions section. Patients were followed until the earliest of the event, censoring at CRRT discontinuation for non-coagulation reasons or the end of the observation period. Non-coagulation discontinuation was treated as a censoring event; therefore, the Cox models estimate the cause-specific hazard of coagulation-related circuit failure among patients still at risk.

Patient and public involvement

Patients and/or the public were not involved in the design, conduct, reporting or dissemination plans of this research.

Results

Study population and baseline characteristics

A total of 3145 patients received CRRT treatment between January 2022 and December 2023, among whom 2387 were administered citrate anticoagulation. After strict screening per the inclusion and exclusion criteria (excluding patients with missing data, concurrent ECMO treatment and non-coagulation-related early withdrawal), 732 patients with unplanned early CRRT withdrawal were finally enrolled in the analysis (figure 2). Of these, 627 (85.7%) experienced withdrawal due to extracorporeal circuit coagulation (coagulation withdrawal group), and 105 (14.3%) due to other causes (non-coagulation withdrawal group).

Baseline characteristics of the study population are summarised in online supplemental table S1. The median age of all patients was 63 years, with 70.2% (514/732) being male. The most common indication for CRRT was volume management optimisation (42.3%, 310/732), and the right femoral vein was the primary catheterisation site (53.8%, 394/732). Comparative analysis revealed statistically significant differences between the coagulation and non-coagulation withdrawal groups in total hospital stay, ICU stay, CRRT arterial/venous pot coagulation grade, ultrafiltration rate, citrate flow rate, arterial pressure and transmembrane pressure (all p<0.05). No significant differences were observed in gender, BMI, SOFA score and most laboratory indicators between the two groups (all p>0.05).

Intracatheter Ca²⁺ trajectory classification

A latent class growth model was applied to classify intracatheter Ca²⁺ variation trajectories at four time points (T0–T3) for both the coagulation withdrawal group (n=627) and the overall unplanned withdrawal group (n=732), with model fitting evaluated by AIC, BIC, aBIC, entropy, LMR and BLRT (tables 1 and 2). The three-class trajectory model was identified as the optimal fit for both groups, characterised by the lowest AIC/BIC/aBIC values, significant LMR (p<0.05) and BLRT (p<0.001) results, and high entropy values (0.776 for the coagulation group and 0.747 for overall group), indicating good classification accuracy.

Table 1. Comparison of model fitting parameter indices for different trajectory models in the coagulation withdrawal group (n=627).

Number of classes	k	AIC	BIC	aBIC	Entropy	P value		Category ratio
						LMR	BLRT
1	8	−6818.902	−6783.374	−6808.773	—	—	—	—
2	13	−7158.461	−7100.729	−7142.002	0.665	＜0.001	＜0.001	0.620/0.380
3	18	−7264.489	−7184.552	−7241.699	0.776	0.0342	＜0.001	0.257/0.643/0.100
4	23	−7297.638	−7195.496	−7268.518	0.668	0.1369	＜0.001	0.034/0.217/0.479/0.270
5	28	−7330.465	−7206.118	−7295.014	0.722	0.0019	＜0.001	0.002/0.470/0.035/0.277/0.217

- Indicates that there is no data for this item

P value of less than 0.05 is considered statistically significant.

Entropy refers to information entropy.

k the number of latent classes (latent categories) in the fitted latent class model.

aBIC, sample-size-adjusted Bayesian Information Criterion; AIC, Akaike Information Criterion; BIC, Bayesian Information Criterion; BLRT, Bootstrapped Likelihood Ratio Test; LMR, Lo-Mendell-Rubin.

Table 2. Comparison of model fitting parameter indices for different trajectory models in the overall unplanned premature withdrawal group (n=732).

Number of classes	k	AIC	BIC	aBIC	Entropy	P value		Category ratio
						LMR	BLRT
1	8	−7970.934	−7934.167	−7959.570	—	—	—	—
2	13	−8368.989	−8309.244	−8350.523	0.653	＜0.001	＜0.001	0.596/0.404
3	18	−8502.014	−8419.290	−8476.446	0.747	0.0131	＜0.001	0.653/0.253/0.094
4	23	−8535.820	−8430.117	−8503.150	0.672	0.1867	＜0.001	0.027/0.320/0.454/0.199
5	28	−8573.539	−8444.857	−8533.766	0.716	0.0795	＜0.001	0.315/0.438/0.028/0.028/0.191

— Indicates that there is no data for this item.

P value of less than 0.05 is considered statistically significant.

Entropy refers to information entropy.

k the number of latent classes (latent categories) in the fitted latent class model.

aBIC, sample-size-adjusted Bayesian Information Criterion; AIC, Akaike Information Criterion; BIC, Bayesian Information Criterion; BLRT, Bootstrapped Likelihood Ratio Test; LMR, Lo-Mendell-Rubin.

The class solution was selected based on converging lines of evidence rather than a single information criterion: (1) statistical fit, as the three-class model had lower AIC, BIC and aBIC values than the two-class model, with significant LMR (p<0.05) and BLRT (p<0.001) results; (2) classification quality, with entropy values of 0.776 (coagulation group) and 0.747 (overall group) and (3) class size, with all three classes exceeding 5% of the total population and the smallest class comprising 9.4% of the overall group.

The three trajectory types were uniformly defined for both groups based on baseline values and variation trends: low baseline with stable development: 25.7% (coagulation group) versus 25.3% (overall group), with intracatheter Ca²⁺ maintained at 0.296–0.312 mmol/L throughout CRRT; moderate baseline with slow decline: 64.3% (coagulation group) versus 65.3% (overall group), the most common type, with Ca²⁺ declining from 0.379/0.377 mmol/L to 0.353/0.349 mmol/L and high baseline with rapid decline: 10.0% (coagulation group) versus 9.4% (overall group), with Ca²⁺ dropping sharply from 0.489/0.458 mmol/L to 0.382/0.396 mmol/L. The median interval between T3 and circuit termination was 1.0 hour (IQR 0.5 to 2.0 hours).

Table 2 The dynamic variation trends of the three Ca²⁺ trajectory types for the coagulation withdrawal group and the overall group are visualised in figures 3 and 4, respectively.

Figure 3. Trajectories of changes in Ca²⁺ levels in the coagulation withdrawal group pipeline. Low starting point with stable development type (average estimated baseline value range: 0.294–0.312 mmol/L); middle starting point with slow decline type (average estimated baseline value range: 0.349–0.377 mmol/L) and high starting point with rapid decline type (average estimated baseline value range: 0.396–0.458 mmol/L).

Figure 4. Trajectories of changes in Ca²⁺ levels in the overall unplanned premature withdrawal group pipeline. Low starting point with stable development type (average estimated baseline value range: 0.294–0.312 mmol/L); middle starting point with slow decline type (average estimated baseline value range: 0.349–0.377 mmol/L); and high starting point with rapid decline type (average estimated baseline value range: 0.396–0.458 mmol/L).

Baseline differences among Ca²⁺ trajectory groups

Comparative analysis of baseline data across the three Ca²⁺ trajectory groups (online supplemental tables S2 and S3) revealed significant intergroup differences. In the coagulation withdrawal group, statistically significant differences were observed in the incidence of Grade 2 or above surgery within 1 month, fibrinogen (FIB) levels, lactic acid (LAC) levels and the professional title of CRRT-operating nurses (all p<0.05); in the overall unplanned withdrawal group, FIB levels, LAC levels and nurses’ professional titles showed significant intergroup differences (all p<0.05).

Multivariate linear stepwise regression analysis identified the key factors influencing Ca²⁺ trajectory classification (table 3). In the coagulation withdrawal group, nurses’ professional titles (β=−0.228, p=0.004), FIB levels (β=0.000, p=0.014) and patient gender (β=0.118, p=0.042) were independent associated factors; in the overall withdrawal group, nurses’ professional titles (β=−0.215, p=0.004), blood product transfusion (β=−0.113, p=0.022) and serum creatinine (CREA) levels (β=0.000, p=0.040) were the main factors associated with it.

Table 3. Analysis of regression factors associated with trajectory groups.

Variable	Regression coefficient B value	T value	P value	95% CI		Multicollinearity statistics
				Lower limit	Upper limit	Tolerance	VIF
Anticoagulation withdrawal team (n=627)*
(Constant)	1.406	10.250	0.000	1.137	1.676
Professional title of the nurse performing CRRT	−0.228	−2.869	0.004	−0.384	−0.072	0.999	1.001
FIB	0.000	2.462	0.014	0.000	0.001	0.994	1.006
Patient gender	0.118	2.041	0.042	0.004	0.232	0.995	1.005
Overall unplanned crew withdrawal (n=732)†
(Constant)	1.781	15.262	0.000	1.552	2.010
Professional title of the nurse performing CRRT	−0.215	−2.897	0.004	−0.361	−0.069	0.992	1.008
Whether to transfuse blood products	−0.113	−2.294	0.022	−0.209	−0.016	0.990	1.010
CREA	0.000	2.056	0.040	0.000	0.000	0.984	1.016

Tolerance>0.2

VIF<10 indicates no collinearity issue among the variables.

B = Unstandardized regression coefficient.T = t statistic

^*F=4.167, p=0.042, R2=0.027 and adjusted R2=0.023.

^†F=4.228, p=0.040, R2=0.023 and adjusted R2=0.019.

95% CI, 95.0% Confidence Interval; CREA, serum creatinine; CRRT, continuous renal replacement therapy; FIB, fibrinogen; VIF, variance inflation factor.

No multicollinearity was observed among all variables (tolerance>0.2 and variance inflation factor<10).

Survival analysis of filter lifespan and risk stratification

Kaplan-Meier survival analysis was performed to plot filter lifespan curves for the three Ca²⁺ trajectory groups (figure 5), with a log-rank test for intergroup comparison. The results showed that the overall intergroup difference in coagulation withdrawal risk was not statistically significant (χ²=5.974, p=0.05). Descriptive analysis indicated that the high baseline with rapid decline group had the shortest median filter lifespan and the highest coagulation withdrawal risk, followed by the low baseline with stable development group, while the moderate baseline with slow decline group had the longest filter lifespan and the lowest risk.

Figure 5. Survival curve of filter lifespan in trajectory groups of the overall unplanned early withdrawal group.

The survival time of the filter and whether it was withdrawn due to coagulation in the overall unplanned early withdrawal group were taken as the dependent variables. The patients’ baseline characteristics and the classification results of the trajectory groups were taken as covariates and included in the multivariate Cox proportional hazards regression model. Dummy variables were assigned to the categorical variables among the covariates. The specific assignment is shown in online supplemental material 1, online supplemental table S4. Based on the Cox proportional hazards regression analysis results (online supplemental table S5), the three Ca²⁺ trajectory groups were stratified by the risk of coagulation-induced filter lifespan reduction:

• High-risk group: high baseline with rapid decline (reference group, HR=1.0).

• Intermediate-risk group: low baseline with stable development (HR=0.534 (95% CI 0.289 to 0.986), p=0.045).

• Low-risk group: moderate baseline with slow decline (HR=0.633 (95% CI 0.425 to 0.942), p=0.024).

Discussion

This retrospective cohort study of 732 ICU patients with unplanned early withdrawal of citrate-anticoagulated CRRT identified three heterogeneous intracatheter Ca²⁺ trajectories—low baseline with stable development, moderate baseline with slow decline and high baseline with rapid decline—and confirmed their association with extracorporeal circuit coagulation and filter lifespan. These findings align with the core premise that dynamic Ca²⁺ changes, rather than single-time-point concentrations, better reflect the efficacy of citrate anticoagulation,¹⁶ and provide novel insights for risk stratification and personalised management of patients receiving CRRT.

The high baseline with rapid decline trajectory was classified as the high-risk group for coagulation-induced filter lifespan reduction (HR=1.0), consistent with previous observations that excessive fluctuations in intracatheter Ca²⁺ are linked to increased coagulation risk.¹⁷ This trajectory may reflect inadequate citrate chelation of Ca²⁺ in the early stage of CRRT, followed by abrupt adjustments that fail to maintain stable anticoagulant effects, leading to accelerated circuit coagulation. In contrast, the moderate baseline with slow decline trajectory (HR=0.633 (95% CI 0.425 to 0.942)) was the low-risk group, suggesting that a gradual reduction in Ca²⁺ within a moderate range (0.349–0.377 mmol/L) is associated with anticoagulation without inducing coagulation or citrate accumulation. This is supported by studies showing that a flexible Ca²⁺ target range (0.35–0.50 mmol/L) maintains good anticoagulant effects,^{14 15} and our findings further narrow this range to 0.349–0.377 mmol/L for clinical practice. The low baseline with a stable development trajectory (HR=0.534 (95% CI 0.289 to 0.986)) was the intermediate-risk group, indicating that even stable Ca²⁺ within the traditional target range (0.294–0.312 mmol/L)⁷ does not eliminate coagulation risk, possibly due to individual differences in citrate metabolism or underlying hypercoagulable states.

The ‘high baseline with rapid decline’ trajectory was associated with the highest risk of coagulation-induced filter lifespan reduction. However, because the T3 time point was obtained immediately before premature shutdown, this trajectory may partially capture the acute deterioration of circuit function rather than representing a purely pre-event predictive pattern. The T0 and T1 values, which reflect the early phase of CRRT, may be more suitable for early risk stratification, whereas the full trajectory incorporating T3 may be more useful for confirming impending circuit failure and guiding immediate clinical decisions. These considerations highlight the need for future prospective studies with predefined, standardised sampling intervals to validate the predictive utility of these trajectories at clinically actionable time points.

Our multivariable analysis identified factors associated with circuit coagulation, including BMI, SOFA score, RASS score, lipid drug administration and blood-product transfusion, consistent with prior published reports. It should be emphasised that some variables captured during CRRT (for example transmembrane pressure) may be altered as filter clotting develops and thus may serve as markers of progressive circuit dysfunction rather than causal exposures. Higher BMI is associated with increased inflammation and vascular endothelial injury,²⁸ while elevated SOFA scores reflect multiorgan failure and systemic inflammatory responses that activate coagulation pathways.^{29 30} A positive RASS score indicates patient restlessness, which can cause catheter displacement and circuit instability,³¹ and lipid drugs increase blood viscosity and filter membrane adsorption,^{32 33} all contributing to coagulation. Blood product transfusion, particularly of platelets and fresh frozen plasma, introduces additional coagulation factors,³⁴ consistent with Wang et al’s findings.³² Notably, a lack of valid CRRT operation qualification was a risk factor, highlighting the critical role of nursing expertise in maintaining circuit patency, as previously emphasised.³⁵

Factors associated with lower risk—no grade 2+surgery within 1 month, higher citrate flow rate and lower transmembrane pressure—are also clinically meaningful. Surgery-induced inflammatory responses and hypercoagulable states increase coagulation risk,³⁶ so avoiding recent major surgery reduces this burden. Higher citrate flow rates enhance Ca²⁺ chelation,¹⁶ while lower transmembrane pressure reflects unobstructed blood flow and reduced filter resistance,^37–39 consistent with studies linking transmembrane pressure stability to prolonged filter lifespan.⁴⁰ However, our finding that lower transmembrane pressure is protective differs from reports of elevated pressure as a risk factor,^{41 42} likely due to our sample’s median transmembrane pressure of 60 mm Hg (within a safe range), emphasising the importance of context-dependent pressure monitoring.

The three Ca²⁺ trajectories were influenced by nurses’ professional titles, FIB and patient gender in the coagulation withdrawal group, and by nurses’ professional titles, blood product transfusion and CREA in the overall group. Nurses’ professional titles may reflect differences in circuit management and citrate dose adjustment proficiency,³⁵ while FIB and CREA are markers of coagulation activation and renal function,³⁰ respectively, underscoring the interplay between clinical factors and Ca²⁺ dynamics.

Although intracatheter Ca²⁺ is the direct mechanistic target of citrate anticoagulation, and an association between abnormal Ca²⁺ patterns and coagulation is therefore partly expected, the clinical value of this study lies in demonstrating that dynamic trajectory patterns may provide information beyond conventional single-point monitoring. Current practice typically relies on intermittent Ca²⁺ measurements interpreted against a fixed target range, which implicitly assumes that patients with similar single-point values carry a similar coagulation risk. Our findings suggest that this assumption may not fully capture the heterogeneity of anticoagulant response: patients with comparable isolated Ca²⁺ values followed fundamentally different trajectories and experienced markedly different risks of coagulation-induced filter lifespan reduction. Notably, the high baseline with rapid decline trajectory carried the highest risk, whereas the moderate baseline with slow decline trajectory carried the lowest, even though both may transiently include values within or near the conventional target range. This suggests that the pattern of change, rather than any single value, may be more informative for risk stratification.

Trajectory-based monitoring may also facilitate earlier identification of high-risk patients. In the high-risk group, early Ca²⁺ values were often above the conventional target range before declining rapidly; reliance on single-point measurements and reactive adjustment may delay recognition until circuit failure is already established. In addition, the lowest-risk trajectory was characterised by a Ca²⁺ range (approximately 0.349–0.377 mmol/L) somewhat higher than the traditionally recommended 0.25–0.35 mmol/L, raising the possibility that the conventional target may be unnecessarily low for some patients and may predispose to excessive citrate delivery, citrate accumulation and increased workload. Finally, the association between nurses’ professional titles and trajectory classification suggests that recognising and responding to dynamic Ca²⁺ patterns, rather than isolated out-of-range values, may be an important component of CRRT nursing expertise and a potential target for training and quality improvement. These considerations support the hypothesis that trajectory-guided management may improve circuit longevity and safety compared with single-point-guided management, although this requires prospective validation. Notably, these trajectory classes and observed intracatheter Ca²⁺ ranges were derived from a single-centre cohort with local institutional CRRT protocols, specific devices, filter configurations and nursing practices. External validation in independent cohorts adopting different CRRT platforms, alternative RCA protocols and varied clinical settings is essential before these trajectory patterns can be applied more broadly.

Our findings have practical implications for clinical practice. First, dynamic monitoring of intracatheter Ca²⁺ trajectories (T0–T3) enables early identification of high-risk patients (eg, high baseline with rapid decline) who require more frequent circuit assessments and citrate dose adjustments. Second, targeting modifiable risk factors—such as optimising sedation (RASS score), avoiding unnecessary lipid drug use or blood transfusion and ensuring CRRT operators are qualified—may be associated with reduced coagulation risk. Third, the recommended intracatheter Ca²⁺ range (0.349–0.377 mmol/L) for low-risk patients provides a more precise target than the KDIGO guideline’s 0.25–0.35 mmol/L,⁷ balancing anticoagulant efficacy and safety.

Limitations

This study has several limitations that should be acknowledged when interpreting the results. First, and most importantly, the study population was restricted to patients who experienced unplanned early CRRT withdrawal due to circuit coagulation or other causes. This means that the study does not evaluate predictors of circuit coagulation in the overall population of patients undergoing RCA-CRRT but rather characterises intracatheter Ca²⁺ trajectory patterns within a selected subgroup already enriched for premature interruption events. Consequently, the trajectory classifications and risk stratification derived from this cohort may not be directly generalisable to all patients initiating RCA-CRRT, and the findings should be interpreted as describing patterns within this specific subgroup rather than as predictive tools for the general CRRT population. Future studies including the full spectrum of patients receiving RCA-CRRT, including those who complete treatment without coagulation events, are needed to validate these trajectories as predictive biomarkers.

Second, the temporal structure of the trajectory analysis warrants careful interpretation. T3 was defined as the most recent intracatheter Ca²⁺ measurement before premature shutdown, with a median interval of 1.0 hour (IQR 0.5 to 2.0 hours) between T3 and circuit termination. Because this measurement was obtained in close proximity to the coagulation event, the full T0–T3 trajectory incorporates information obtained close to the outcome and may partially reflect ongoing circuit failure rather than representing a true early predictive marker. This is particularly relevant for the ‘high baseline with rapid decline’ pattern, and the observed association between this trajectory and coagulation-induced filter lifespan reduction should therefore be interpreted with caution. The T0–T3 trajectory should be regarded as describing the pattern of Ca²⁺ changes leading up to circuit failure, rather than as a baseline predictor of subsequent filter failure. The T0 and T1 values, which were obtained earlier during CRRT, may be more informative for early risk stratification, whereas the T3 value may be more indicative of established circuit failure. Because the trajectory classification incorporates measurements obtained near the time of the event, reverse causation cannot be excluded, and the findings should be regarded as hypothesis-generating. Future studies with more frequent and standardised sampling intervals, including measurements well before the anticipated coagulation event, are needed to distinguish early predictive trajectories from those that simply reflect impending or ongoing circuit failures.

Third, the Cox proportional hazards analysis requires careful interpretation. Trajectory class was determined using measurements collected during the CRRT episode, including the final measurement before premature discontinuation; it is therefore a post-baseline, time-dependent exposure rather than a baseline covariate. Treating the trajectory class as though it were known at the beginning of follow-up may introduce time-related bias, and the resulting HRs should be interpreted as describing associations with the pattern of Ca²⁺ changes preceding circuit failure, rather than as predictive effects of a baseline characteristic. In addition, non-coagulation discontinuation was treated as a censoring event; therefore, the models estimate the cause-specific hazard rather than the cumulative incidence of coagulation-related filter failure. Our findings should not be interpreted as absolute risks or cumulative incidences.

Fourth, the definition of ‘unplanned early withdrawal’ was set at 72 hours of CRRT treatment, which may exclude patients who experienced coagulation-related filter failure beyond this timeframe, potentially leading to incomplete capture of outcome events. Fifth, although multiple imputation was used to handle missing data, the LCGM was performed on a single imputed dataset rather than pooling across all 20 imputed datasets, because the LCGM procedure in Mplus (V.8.3) does not readily support standard pooling. This means that the trajectory classification does not fully incorporate imputation uncertainty. Nevertheless, the relatively high missing rates of AT III (16.7%) and FDP (11.1%) may have introduced subtle biases. Finally, in the LCGM, participants were assigned to their most likely latent class, and this classification was treated as observed without error in subsequent Cox proportional hazards and regression models. Classification uncertainty was therefore not propagated into these analyses, which may lead to underestimated SEs and biased effect estimates. Although the entropy values (0.776 and 0.747) and average posterior class-membership probabilities (≥0.80 in all classes) indicated good classification accuracy, the findings should be interpreted with this limitation in mind. Future studies using approaches that incorporate posterior class-membership probabilities, such as the Bolck–Croon–Hagenaars (BCH) or Maximum Likelihood (ML) three-step methods, or joint latent class survival models, would provide more robust estimates.

Conclusions

In ICU patients with unplanned early withdrawal of citrate-anticoagulated CRRT, three heterogeneous intracatheter Ca²⁺ trajectories were identified, with the high baseline with rapid decline pattern independently associated with coagulation-induced filter lifespan reduction. The Ca²⁺ range observed in the low-risk trajectory subgroup is purely exploratory and hypothesis-generating and cannot serve as a clinical therapeutic target without prospective testing. While these findings should be regarded as hypothesis-generating given the selected population and the proximity of T3 to the coagulation event, they suggest that trajectory-based monitoring may offer incremental value over single-point measurements for risk stratification and personalised anticoagulation management. Prospective multicentre external validation in independent populations with different CRRT hardware, RCA protocols and ICU clinical settings is required to determine whether these trajectories can serve as clinically actionable tools and to evaluate whether the observed Ca²⁺ interval can guide clinical anticoagulation practice.

Data availability statement

Data are available upon reasonable request.