Bleeding recurrence risk among hospitalized patients undergoing therapeutic plasma exchange: a multi-center study

Abstract

Background

In hospitalized patients undergoing therapeutic plasma exchange (TPE), it is not known how TPE-associated bleeding risk is impacted by a prior bleeding episode. Therefore, to assess the prevalence and predictors of bleeding recurrence, we analyzed data from the Recipient Epidemiology and Donor Evaluation Study-III (REDS-III).

Materials and methods

Using a retrospective cross-sectional analysis of REDS-III public use files, we identified hospitalized adults who had a major bleeding episode prior to their first TPE procedure. Patients were classified into two cohorts based on bleeding recurrence (no-recurrence vs recurrence). After identifying potential predictors, we used multiple imputation by chained equations to impute variables with <30% missing data. Variable selection was optimized using a 10-fold cross validated least absolute shrinkage and selection operator. Final predictors were identified by fitting a logistic regression model.

Results

In 310 patients with major bleeding prior to TPE initiation, bleeding recurred in 121 (39.0%). We identified the following seven unique predictors: 1) >10 TPE procedures (OR 2.23); 2) intensive care unit stay (OR 1.35); 3) thrombocytopenia (OR 1.26); 4) surgery (OR 1.22); 5) hepatic disease (OR 1.21); 6) 6–10 TPE procedures (OR 1.04); and 7) Asian race (OR 1.01). We also identified the following five interactions: 1) surgery and therapeutic anticoagulation (OR 1.50); 2) 6–10 TPE procedures and therapeutic anticoagulation (OR 1.05); 3) 6–10 TPE procedures and antiplatelets (OR 1.02); 4) >10 TPE procedures and antiplatelets (OR 1.00); and 5) albumin-only TPE and antiplatelets (OR 0.53). When assessed for adjusted performance, the prediction model had a C-statistic of 0.617 (95% CI 0.613–0.619) and Brier Score of 0.342 (95% CI 0.340–0.347).

Discussion

In this study assessing predictors of bleeding recurrence among hospitalized patients undergoing TPE, we identified seven variables and five interactions. These findings should be validated in future studies.

INTRODUCTION

In hospitalized patients undergoing therapeutic plasma exchange (TPE), medical comorbidities and the need for inpatient procedures may contribute to increased bleeding risk. Even prior to initiating TPE, patients may experience a major bleeding event associated with surgery, anticoagulant use, or critical illness^1–3. In these patients, bleeding recurrence may be associated with multiple comorbidities, such as sepsis-associated severe thrombocytopenia and disseminated intravascular coagulation^2,4–6. Bleeding recurrence may also be precipitated or aggravated due to TPE impacts on hemostasis including 1) hemodilution, 2) consumption of coagulation factors in the extracorporeal circuit, and 3) inadvertent removal of platelets and procoagulant proteins⁴.

Patients who have experienced a major bleeding event prior to TPE may be at increased risk for subsequent bleeding episodes. However, the risk of bleeding recurrence following TPE initiation is unknown^2,3. Understanding the factors contributing to bleeding recurrence risk following TPE initiation is critical to guide pre-TPE bleeding mitigation strategies. By developing effective bleeding mitigation strategies, major bleeding outcomes among hospitalized patients undergoing TPE can be improved.

In a prior study of hospitalized patients undergoing TPE in the Recipient Epidemiology and Donor Evaluation Study-III (REDS-III), we identified a 28.5% prevalence of major bleeding events prior to TPE initiation⁷. In this follow-up study, we sought to evaluate the prevalence and predictors of bleeding recurrence. As shown in subsequent sections of this manuscript, bleeding recurrence can occur in up to 39.0% of hospitalized patients who experienced bleeding prior to TPE with several bleeding recurrence predictors identified.

MATERIALS AND METHODS

Data source, study design and participants

The study design is a retrospective, cross-sectional, patient-level analysis of public use data files from the REDS-III study⁷. The REDS-III study was a National Heart, Lung and Blood Institute-sponsored 4-year (2012–2016) multicenter observational study developed to evaluate transfusion outcomes. Using electronic health record data from four United States blood centers and 12 associated hospitals, REDS-III provides data on 6.5 million blood components and 120,290 transfusion recipients⁷. Additionally, during the study period, all non-transfused hospitalized patients (1,285,359 encounters) were included as a comparison group⁷. To protect patient confidentially, all subject identifiers have been replaced by random numbers⁷. Therefore, this study was reviewed by the Duke University Health System Institutional Review Board and determined to be exempt.

The inclusion criteria were adult patients (≥18 years old) in a first inpatient encounter who experienced a major bleeding episode prior to their first TPE. Patients receiving TPE were identified using validated International Classification of Diseases (ICD) and Current Procedural Terminology (CPT) codes⁷. Major bleeding was defined using ISTH (International Society on Thrombosis and Haemostasis) criteria⁷. Bleeding recurrence following TPE was defined as any major bleeding event that occurred within 48 hours of a TPE-procedure⁷. Surgery was determined based on patients being transferred to or from the Operation Room using the “RC_Transfers” Table. All other definitions, including ICD and CPT codes and details of major bleeding diagnosis, have been previously published⁷.

Statistical analysis

The statistical analysis was a patient-level analysis using mean, median, interquartile, range, and standard deviation (SD) or, as appropriate, counts and percentages. Tests of associations between recurrence status and baseline characteristics were examined by Chi-Square for categorical variables and Kruskal-Wallis tests for continuous variables. The analysis used SAS version 9.4 (SAS Institute Inc., Cary, NC, USA). A p-value of <0.05 was considered statistically significant. The exploratory univariate results for characteristics (for groups of variables) are presented without and with adjustment for multiple testing and type 1 error. A Bonferroni correction was applied to correct for a total of 97 comparisons run (albeit only 33 comparisons are reported).

Predicting recurrence risk

We conducted a targeted literature review to identify potential bleeding recurrence predictors (see Online Supplementary Table SI)⁴. Identified variables within the REDS-III database (demographic characteristics, transfusion details, medications, and laboratory values) were included in the analysis⁷. Based on the potential predictors identified, we also explored interactions terms of clinical relevance (see Online Supplementary Table SII).

We first evaluated the skew and distribution of continuous variables. Due to irregular distribution, age and number of TPE procedures were included as categorical variables. For our potential predictors, we excluded from model development any variables with low cell counts (<5)⁸. This strategy improves the model’s reliability by mitigating unstable estimates and inflated Type I error rates.

Due to the high number of identified potential predictors and the low number of patients with bleeding recurrence further variable stratification was not performed. By simplifying and streamlining the modeling process, this strategy reduces overfitting risk in LASSO (Least Absolute Shrinkage and Selection Operator) regression, promotes model interpretability, and enhances computational efficiency⁸.

To address missing data, we used Multiple Imputation by Chained Equations (MICE). We specified, a priori, that MICE would be used for only variables with ≤30% of missing data (with a 1% buffer for key variables of interest)⁹. Any variables with >30% of missing data were excluded from the analysis. For the implementation of MICE, we used the “mice” package in R; and specified a seed = 1001 to create 10 imputations after 15 iterations⁹. To impute ethnicity (Hispanic vs non-Hispanic), thrombocytopenia, anemia, and coagulopathy we used logistic regression. To impute race, we used Bayesian polytomous regression model.

To optimize variables selection, we used regular CV (Cross Validated)-LASSO. Considering our imputation technique and relatively low number of patients with bleeding recurrence, we set a weight to each patient equal to 0.1. To derive the LASSO model coefficients on the weighted stacked set, we used the mean λ (0.045; range 0.039–0.052)¹⁰. To obtain the final performance metric of our model, we calculated the apparent and adjusted performance. The apparent performance involves fitting a model and evaluating the model’s performance on each of the imputed datasets individually. The mean performance across the 10 imputed datasets is then used as the final estimate. The adjusted performance involves bootstrapping on each of the imputed datasets¹⁰. By resampling 200 times and fitting a model on only a portion of the data, less biased results are extracted. To demonstrate the performance of our model, we calculated the C-statistic (AUC-ROC) and the Brier Score. The C-statistic measures the probability that a randomly selected patient who experienced bleeding recurrence had a higher risk score than a patient who had not experienced bleeding recurrence given the predictors in our model¹¹. The Brier Score quantifies how well our model predicts an event like major bleeding in comparison to a null model.

RESULTS

Of 310 patients with major bleeding prior to TPE initiation, major bleeding recurred in 121 patients (39.0%; see Figure 1).

Figure 1.

Study flow diagram

There were a total of 310 patients who experienced major bleeding prior to TPE. Of this number, 189 (61.0%) had no recurrence and 121 (39.0%) had recurrence.

Variable selection

Fifty-six variables were excluded due to low cell counts (see Online Supplementary Table SIII). Three additional variables excluded for missing data were body mass index (43.23%), obesity (39.35%), and hypofibrinogenemia (63.55%). Imputed variables were thrombocytopenia (4.84%), anemia (2.58%), race (2.90%), and ethnicity (2.90%). Because its 30.97% missingness was close to our cutoff, we also imputed coagulopathy. Therefore, our model included a final list of 33 potential predictors and 30 interaction terms (see Online Supplementary Table SIV).

Patient characteristics by bleeding recurrence status

In patients with bleeding recurrence, the average age was 54.4 years old (see Table I). Most patients were male (No.=65; 53.7%), white race (No.=85; 72.0%), and non-Hispanic (No.=109; 90.1%). From the univariate analysis, bleeding recurrence differed only for surgery (27.3 vs 11.6%; p=0.0388). The most prevalent American Society for Apheresis TPE-indications showed no statistically significant findings (summarized in Table I)⁷.

Table I.

Baseline and procedure characteristics by recurrence status

Characteristics	No-recurrence (No.=189)	Recurrence (No.=121)	Total (No.=310)	p-value	Bonferroni adjusted p-values
	Demographics
Age ¶				0.83581	>0.9999
Mean (SD)	54.5 (17.6)	54.4 (15.8)	54.5 (16.9)
Median	57.0	55.0	56.0
Range	(18.0–90.0)	(21.0–87.0)	(18.0–90.0)
Interquartile range (Q1–Q3)	40.0–67.0	43.0–66.0	42.0–67.0
Age Categorized				0.5026	>0.9999
18–44 yrs	60 (31.7%)	32 (26.4%)	92 (29.7%)
45–64 yrs	67 (35.4%)	53 (43.8%)	120 (38.7%)
65–79 yrs	50 (26.5%)	30 (24.8%)	80 (25.8%)
80+ yrs	12 (6.3%)	6 (5.0%)	18 (5.8%)
Gender				0.15802	>0.9999
Female	103 (54.5%)	56 (46.3%)	159 (51.3%)
Male	86 (45.5%)	65 (53.7%)	151 (48.7%)
Race				0.11852	>0.9999
White	132 (72.1%)	85 (72.0%)	217 (72.1%)
Black or African American	29 (15.8%)	21 (17.8%)	50 (16.6%)
Asian	4 (2.2%)	7 (5.9%)	11 (3.7%)
Other	18 (9.8%)	5 (4.2%)	23 (7.6%)
Not reported	6	3	9
Ethnicity				0.86552	>0.9999
Hispanic	15 (7.9%)	8 (6.6%)	23 (7.4%)
Non-hispanic	169 (89.4%)	109 (90.1%)	278 (89.7%)
Not specified	5 (2.6%)	4 (3.3%)	9 (2.9%)
	TPE procedure details
Number of TPE Procedures ¶				0.02261	>0.9999
Mean (SD)	2.2 (1.9)	3.1 (3.4)	2.5 (2.6)
Median	1.0	2.0	1.0
Range	(1.0–9.0)	(1.0–22.0)	(1.0–22.0)
Interquartile Range (Q1–Q3)	1.0–3.0	1.0–4.0	1.0–3.0
Number of TPE procedures categorized				0.0019	>0.9999
1–2 procedures	137 (72.5%)	79 (65.3%)	216 (69.7%)
3–5 procedures	41 (21.7%)	20 (16.5%)	61 (19.7%)
6–10 procedures	11 (5.8%)	17 (14.0%)	28 (9.0%)
>10 procedures	0 (0.0%)	5 (4.1%)	5 (1.6%)
	Type of replacement solution
100% Plasma	45 (23.8%)	28 (23.1%)	73 (23.5%)	0.89232	>0.9999
	Comorbidities
Antiplatelet therapy	165 (87.3%)	100 (82.6%)	265 (85.5%)	0.25622	>0.9999
Therapeutic anticoagulation	50 (26.5%)	42 (34.7%)	92 (29.7%)	0.12062	>0.9999
Arterial thrombosis	14 (7.4%)	11 (9.1%)	25 (8.1%)	0.59542	>0.9999
Venous thrombosis	23 (12.2%)	14 (11.6%)	37 (11.9%)	0.87392	>0.9999
Central venous catheter	168 (88.9%)	110 (90.9%)	278 (89.7%)	0.56852	>0.9999
Hypertension	46 (24.3%)	29 (24.0%)	75 (24.2%)	0.94062	>0.9999
Ischemic heart disease	11 (5.8%)	13 (10.7%)	24 (7.7%)	0.11362	>0.9999
Heart failure	15 (7.9%)	17 (14.0%)	32 (10.3%)	0.08442	>0.9999
Atrial fibrillation	10 (5.3%)	4 (3.3%)	14 (4.5%)	0.41162	>0.9999
Diabetes	22 (11.6%)	13 (10.7%)	35 (11.3%)	0.80782	>0.9999
Smoker	38 (20.1%)	24 (19.8%)	62 (20.0%)	0.95362	>0.9999
Chronic obstructive pulmonary disease	25 (13.2%)	16 (13.2%)	41 (13.2%)	0.99912	>0.9999
Renal disease	70 (37.0%)	52 (43.0%)	122 (39.4%)	0.29652	>0.9999
Hepatic disease	17 (9.0%)	22 (18.2%)	39 (12.6%)	0.01732	>0.9999
Cancer	30 (15.9%)	14 (11.6%)	44 (14.2%)	0.28962	>0.9999
Mechanical invasive ventilation	17 (9.0%)	14 (11.6%)	31 (10.0%)	0.46092	>0.9999
Invasive procedures	64 (33.9%)	36 (29.8%)	100 (32.3%)	0.45012	>0.9999
Intensive care unit stay	111 (58.7%)	94 (77.7%)	205 (66.1%)	0.00062	0.0582
Surgery	22 (11.6%)	33 (27.3%)	55 (17.7%)	0.00042	0.0388
Thyroid disease	25 (13.2%)	14 (11.6%)	39 (12.6%)	0.6678	>0.9999
Anemia				0.19492	>0.9999
Missing	4	4	8
No-anemia	17 (9.2%)	6 (5.1%)	23 (7.6%)
Yes-anemia	168 (90.8%)	111 (94.9%)	279 (92.4%)
Coagulopathy				0.94962	>0.9999
Missing	67	29	96
No-coagulopathy	34 (27.9%)	26 (28.3%)	60 (28.0%)
Yes-coagulopathy	88 (72.1%)	66 (71.7%)	154 (72.0%)
Thrombocytopenia				0.00092	0.0873
Missing	11	4	15
No-thrombocytopenia	83 (46.6%)	32 (27.4%)	115 (39.0%)
Yes-thrombocytopenia	95 (53.4%)	85 (72.6%)	180 (61.0%)
	TPE-Indication **
Hypertriglyceridemic and/or acute pancreatitis	32 (16.9%)	21 (17.4%)	53 (17.1%)	0.92292	>0.9999
Sepsis	12 (6.3%)	16 (13.2%)	28 (9.0%)	0.03942	>0.9999
Thrombotic thrombocytopenic purpura	19 (10.1%)	12 (9.9%)	31 (10.0%)	0.96902	>0.9999
Renal transplant	13 (6.9%)	6 (5.0%)	19 (6.1%)	0.49182	>0.9999

^*No.=98 in No. recurrence, No.=78 in yes-recurrence; No.=176 in total;

^**overlap of patients across indications may occur; ¶: not used in modeling;

¹Kruskal-Wallis;

²Chi-square.

Predictors and model performance

Our prediction model identified seven unique bleeding risk factors as follows: 1) >10 TPE procedures (OR 2.23); 2) ICU stay (OR 1.35); 3) thrombocytopenia (OR 1.26); 4) surgery (OR 1.22); 5) hepatic disease (OR 1.21); 6) 6–10 TPE procedures (OR 1.04); and 7) Asian race (OR 1.01). We also identified the following five interactions: 1) surgery and therapeutic anticoagulation (OR 1.50); 2) 6–10 TPE procedures and therapeutic anticoagulation (OR 1.05); 3) 6–10 TPE procedures and antiplatelets (OR 1.02); 4) >10 TPE procedures and antiplatelets (OR 1.00); and 5) albumin-only TPE & antiplatelets (OR 0.53).

Analysis of our model showed an acceptable apparent performance: C-statistic of 0.729 (95% CI 0.728–0.731) and Brier Score of 0.2072 (95% CI 0.2063–0.2077)^12–17. Our model also had an acceptable adjusted performance: C-statistic of 0.617 (95% CI 0.613–0.619) and Brier Score of 0.342 (95% CI 0.340–0.347).

DISCUSSION

In our cross-sectional analysis of hospitalized patients who experienced major bleeding prior to TPE initiation, we found that bleeding recurred in 39.0%. This high prevalence is corroborated by prior studies demonstrating that patients receiving TPE are often critically ill, and may be at increased bleeding risk^18–20. A Nationwide Inpatient Sample (NIS) study found non-significantly increased bleeding events among immune thrombocytopenia cases treated with TPE (p=0.725). In this study, TPE-treated cases were in the highest disease severity class (34.6%)²¹. The same pattern is suggested by our findings that bleeding recurrence is increased among critically ill patients (increased ICU stay). Therefore, future studies should evaluate effective bleeding risk mitigation strategies in critically ill patients with pre-existing bleeding risk factors prior to the TPE procedure.

Among bleeding recurrence predictors identified in this study is surgery. Surgery was selected both as a unique variable and when associated with therapeutic anticoagulation. Surgery often precedes TPE in patients with acute solid organ transplant rejection²². Surgery can also precede TPE in patients with myasthenia gravis, heparin-induced thrombocytopenia, and thrombotic thrombocytopenic purpura^23,24. Despite limited evidence in patients undergoing TPE, surgery is considered to be a potential bleeding predictor by apheresis providers. This bleeding risk may be further exacerbated when coupled with therapeutic anticoagulation^23,25,26. For this reason, in patients undergoing surgery, apheresis providers typically recommend bleeding prophylaxis strategies²⁴. These prophylactic strategies may involve substitution of the replacement solution with full or partial plasma^23,24,26. Plasma replacement alone, however, may not be sufficient to account for all the hemostatic deficits induced by surgery. Indeed, surgery may precipitate major bleeding through several mechanisms i.e., tissue defects, hyperfibrinolysis, dilutional coagulopathy, and acquired platelet dysfunction^27,28. In addition to the surgery type, surgery-associated bleeding risk may be further exacerbated by TPE-associated transient bleeding risk factors such as thrombocytopenia, coagulation factor depletion, and dilutional coagulopathy^4,5,29. Therefore, future studies are needed to better characterize the complex interplay of surgical and TPE-associated bleeding risk. These studies are critical to ensuring that post-surgical TPE bleeding prophylaxis interventions are effective.

Our analysis also identified hepatic disease (both acute and chronic). Patients with hepatic disease have hemostatic deficits that increase both thrombotic and bleeding complications³⁴. Indeed, prior studies of patients with hepatic disease have reported bleeding rates as high as 35%^35–38. In hepatic disease, potential contributors to bleeding include platelet impairment, portal hypertension, thrombocytopenia, hypofibrinogenemia, and decreased clotting factor synthesis^35–38. When patients with hepatic disease experience TPE-associated coagulation factor depletion, these mechanisms may be further exacerbated by defective rebound synthesis⁴. For these patients, prophylactic strategies targeted to the specific deficits may be needed. Currently, apheresis providers do not consistently use bleeding prophylaxis strategies in this population^1,2. In a national survey of apheresis providers presented with a clinical scenario of hepatic disease associated with a mildly elevated INR (1.5–2.0), only 63% (54/87) considered the replacement solution as a strategy to reduce bleeding risk². In another study of 85 patients who underwent 225 TPE procedures, those with hepatic dysfunction were considered to be at high bleeding¹. In this study, however, bleeding prophylaxis was based, not on the replacement solution, but on the type of circuit anticoagulant (regional citrate anticoagulation vs none). Therefore, future studies assessing the efficacy of bleeding prophylactic strategies in patients with hepatic disease are desired.

Another bleeding recurrence predictor identified was thrombocytopenia. A bleeding risk factor commonly reported in the literature, thrombocytopenia has been demonstrated in prior studies of patients initiating TPE with lower baseline platelet counts due to infection, systemic illness, and medication effect^1,4,39–41. Pre-existing thrombocytopenia can be exacerbated by the TPE procedure when platelets are lost in the plasma bag or clot in the centrifuge bowl⁴. Therefore, TPE can transiently exacerbate pre-existing thrombocytopenia and lead to recurrent bleeding events. Nevertheless, the threshold platelet count associated with bleeding recurrence following TPE is still unknown. This uncertainty is reflected in a survey of 83 apheresis providers who did not agree on the threshold platelet count needed to guide prophylactic strategies. While most providers agreed that bleeding prophylaxis was needed for a platelet count of <50×10⁹/L (54.2%), others opted for higher intervention thresholds of <100×10⁹/L (9.6%) and <150×10⁹/L (3.6%)³. Recurrent bleeding events may also be affected by thrombocytopenia etiology³⁹. For this reason, future studies are needed to assess the interplay of thrombocytopenia etiology and the optimal platelet threshold.

Although identified as a predictor, ICU stay is likely a surrogate for both critical illness and the patient complexity⁴². Among critically ill patients, up 14% will develop bleeding complications through several distinct mechanisms, including renal failure, extracorporeal circuits, disseminated intravascular coagulation and coagulopathy associated with multi-organ failure^44–46. This complexity and the associated critical illness may contribute to increased bleeding risk^5,18,47,48. In a prior Nationwide Inpatient Sample database study including 6,014 discharges with a myasthenia gravis diagnosis, TPE (and intravenous immunoglobulin) use was more likely in patients with severe presentation (36.72 vs 28.98%, p<0.0001)⁴⁷. Among 698 patients with a severe presentation, TPE use was more likely than intravenous immune globulin to be associated with acute respiratory failure (34.22 vs 12.43%, p<0.0001) and endotracheal intubation (27.98 vs 12.43%, p<0.001)⁴⁷. Similar findings have been reported for several TPE indications, including heparin induced thrombocytopenia, systemic lupus erythematosus, antiphospholipid syndrome, and immune thrombocytopenia^18,21,49,50. Despite critical illness being a potential bleeding predictor in our analysis, it is important to consider that each one of these TPE indications may have different underlying bleeding risk. Therefore, these findings warrant further investigation.

The strengths of this study include the assessment of bleeding recurrence prevalence and predictors using a large electronic health record multicenter study (REDS-III). REDS-III study data is particularly advantageous because it includes variables that have been previously identified in the literature as potential bleeding predictors (especially laboratory and transfusion data). Study limitations include the inability to determine whether active bleeding was present at the time of TPE initiation, the inability to provide further variable stratification due to low cell counts, and the inability to assess the temporal nature of bleeding recurrence. Additionally, we cannot accurately determine whether TPE or major bleeding ICD and CPT codes were coded on the same date as the actual event. Discrepancy between the event date and its subsequent coding could affect our bleeding recurrence prevalence analysis. Furthermore, although we used rigorous statistical methods, the prediction model’s adjusted performance demonstrated a C-statistic of 0.617. This number is somewhat lower than the performance of other bleeding predictive models currently used in clinical practice (range 0.67–0.74)^12–17. Finally, due to missing data beyond the pre-determined threshold for multiple imputation, fibrinogen could not incorporated into our model. Notwithstanding with these limitations, to our knowledge, our study is the first developed specifically to assess bleeding recurrence prevalence and predictors in hospitalized patients undergoing TPE.

CONCLUSIONS

In a cohort of hospitalized patients who experienced a major bleeding episode prior to TPE, we identified a high prevalence of bleeding recurrence. Identified bleeding recurrence predictors included several important predictors that should be validated in future prospective studies.