General Articles
Prediction of Intraoperative Transfusion Requirements During Orthotopic Liver Transplantation and the Influence on Postoperative Patient Survival Jacek B. Cywinski, MD,* Joan M. Alster, MS,† Charles Miller, MD,‡ David P. Vogt, MD,‡ and Brian M. Parker, MD* BACKGROUND: Predicting blood product transfusion requirements during orthotopic liver transplantation (OLT) remains difficult. Our primary aim in this study was to determine which patient variables best predict recipient risk for large blood transfusion requirements during OLT. The secondary aim was to determine whether the amount of blood products transfused during OLT impacted patient survival. METHODS: Eight hundred four primary adult OLTs performed during a 9-year period were retrospectively analyzed, and predictive models were developed for blood product usage, usage >20 and usage >30 units of red blood cells (RBCs) plus cell salvage (CS). For survival analysis, potential predictors included all blood products administered during OLT. RESULTS: For analyses of RBC + CS usage, we used several statistical techniques: regression analysis, logistic regression, and classification and regression tree analysis. Several preoperative factors were highly statistically significant predictors of intraoperative blood product usage in each of the analyses, namely lower platelet count and higher Model for End-Stage Liver Disease Score or one or more of its components (creatinine, total bilirubin, international normalized ratio). Despite these highly significant associations, the models were unable to predict reliably that patients might require the largest amount of blood products during OLT. For example, the classification and regression tree analyses were able to predict only 32% and 11% of patients requiring >20 and >30 units of RBC + CS, respectively. Survival analysis demonstrated poorer survival among patients receiving larger amounts of RBC + CS during OLT. CONCLUSION: Prediction of intraoperative blood product requirements based on preoperatively available variables is unreliable; however, there is a strong measurable association between transfusion and postoperative mortality. (Anesth Analg 2014;118:428–37)
P
redicting blood product transfusion requirements during orthotopic liver transplantation (OLT) has remained difficult. Because OLT is so resource intensive, understanding which variables predict transfusion requirements could allow improved blood product allocation and utilization. Data published both before, and after, the use of the Model for End-Stage Liver Disease (MELD) Score have shown select preoperative factors to be only minimally predictive of intraoperative transfusion requirements.1–4 Some of the preoperative variables reported in the literature affecting intraoperative transfusion requirements include: recipient age and weight, etiology of liver disease, starting hemoglobin level, prothrombin international normalized ratio (INR) and partial thromboplastin time, platelet count, serum creatinine (Cr), and presence of ascites. Unfortunately, many of these variables have not
From the Departments of *General Anesthesiology and Transplant Center, †Quantitative Health Sciences, and ‡Hepato-pancreato-biliary and Transplant Surgery, Cleveland Clinic, Cleveland, Ohio. Accepted for publication July 17, 2013. Funding: Supported by internal funds. The authors declare no conflicts of interest. Reprints will not be available from the authors Address correspondence to Brian M. Parker, MD, Cleveland Clinic, 9500 Euclid Ave., Cleveland, OH 44195. Address e-mail to [email protected]. Copyright © 2014 International Anesthesia Research Society DOI: 10.1213/ANE.0b013e3182a76f19
428 www.anesthesia-analgesia.org
consistently been found to be associated with transfusion needs during OLT. Higher transfusion requirements during OLT are associated with worse postoperative outcomes including prolonged length of intensive care unit stay, hospital stay, and increased mortality;5 however, these relationships have not been well characterized.6,7 Furthermore, an association rather than causality was established in these studies with varying strengths of correlation. Interestingly, when survival outcomes are compared among centers with different rates of blood product transfusion, centers with higher blood product administration rates do not necessarily have worse survival rates. This apparent anomaly highlights the fact that the relationship between transfusion and outcomes is complex and most likely dependent on factors unique to the particular transplant center.8 Recognizing that it may be very difficult to develop a single, reliable, and universally applicable model to predict transfusion requirements for patients undergoing OLT, the primary aim of this study was to determine which preoperatively available variables best predict recipient blood transfusion requirements during OLT in our program, with particular emphasis on predicting high transfusion requirements (>20 or >30 units of blood). Such models could identify the patients who consume the majority of the blood bank resources and possibly be modified to fit the practice of other institutions. The secondary aim of this study was to determine whether the amount of blood products transfused impacted survival after OLT. February 2014 • Volume 118 • Number 2
METHODS
After receiving IRB approval which waived the need for informed consent, 835 consecutive primary adult (age ≥18 years) OLTs performed from January 1, 2001 to June 30, 2010 at the Cleveland Clinic were retrospectively analyzed. During this time frame, 47 retransplants were performed; in these cases, we analyzed data from the first transplant only. Patients undergoing combined liver and other organ transplants (liver-kidney n = 36, liver-heart n = 3, liver-lung n = 3, liver-pancreas n = 2, liver-pancreas-small intestine n = 1) as well as split liver graft (n = 13) were included in the analysis. In addition, 24 living donor liver transplants were performed during this period. None of the patients undergoing retransplantation or living donor liver transplants was included in the analyses nor were 7 transplants for which we did not have blood product usage information. This resulted in a final analysis of 804 OLTs. All blood products administered intraoperatively to OLT recipients were analyzed, including allogeneic red blood cells (RBCs), fresh frozen plasma (FFP), cryoprecipitate, platelets, and intraoperative red blood cell salvage (CS) units returned to the patient. A predictive model was developed for RBC + CS (using RBC equivalents) usage during liver transplantation. Stepwise multivariable regression analysis was used to identify recipient, donor, and transplant procedure variables associated with blood product usage. The natural logarithm of (RBC + CS) was used to produce a nearly normally distributed analysis variable with stable error variance. Bootstrap aggregation (bagging) was used for variable selection,9 with P < 0.07 for entry into the models and P < 0.05 for variable retention. Variables appearing in at least 50% of bootstrap analyses were considered reliably statistically significant at P < 0.05 (median rule). Several transformations of continuous predictors (x2, 1/x, 1/x2, sqrt[x], ln[x]) were plotted against ln(RBC + CS) to check whether transformation could result in more nearly linear relationships between the predictors and ln(RBC + CS). Since the transformations did not result in significantly improved linearity, the predictors were analyzed on their original scales. In addition, goodness of fit of the regression model was confirmed by plotting residuals versus predicted values. To predict large transfusion requirements, use of >20 and >30 units of RBC + CS (approximately top 20% and 10% of blood product needs in our study) were modeled
Table 1. Recipient Data Variable Age (y) Gender: female BMI Blood type A B AB O Previous abdominal surgery INR Platelet count (k/mm3) TIPS present MELD Score
n (%) or mean ± SD 54 ± 10 241 (30) 28.9 ± 6.2 301 (37) 112 (14) 37 (4.6) 353 (44) 323 (42) 1.52 ± 0.7 95.1 ± 68.3 85 (11) 21 ± 5.6
BMI = body mass index; INR = prothrombin international normalized ratio; TIPS = transjugular intrahepatic portosystemic shunt; MELD = model for endstage liver disease.
February 2014 • Volume 118 • Number 2
Table 2. Primary Diagnosis of End-Stage Liver Disease in 804 Study Patients Diagnosis
n (%) 213 (26) 187 (23) 110 (14) 78 (9.7) 68 (8.5) 67 (8.3) 28 (3.5) 15 (1.9) 13 (1.6) 1 (0.12)
HCC Viral hepatitis Cholestatic Cryptogenic NASH Laennec’s Metabolic disease Autoimmune Fulminant hepatic failure Neuroendocrine tumor
HCC = hepatocellular carcinoma; NASH = nonalcoholic steatohepatitis.
Table 3. Donor Data Variable Donor age Gender: female BMI Blood type A B AB O Mechanism of death ICH/stroke Blunt trauma Gunshot wound Cardiovascular Drug intoxication Asphyxiation Other Unknown Cold ischemic time (min)
n (%) or mean ± SD 44 ± 17 317 (40) 27.3 ± 10 310 92 27 368
(39) (12) (3.4) (46)
386 (48) 168 (21) 61 (13) 61 (7.6) 21 (2.6) 16 (2.0) 42 (5.2) 7 (0.87) 460 ± 163
BMI = body mass index; ICH = intracranial hemorrhage.
Table 4. Intraoperative Transfusion Data Variable RBCs (units) CS (units) RBC + CS (units) FFP (units) Platelets (5 packs pooled)
Mean ± SD 8.49 ± 8.6 5.46 ± 6.3 13.9 ± 13.6 10.3 ± 10.6 3.58 ± 3.0
NB: CS was used for all patients undergoing OLT regardless of liver disease etiology. OLT = orthotopic liver transplant; RBCs = red blood cells; CS = cell salvage; FFP = fresh frozen plasma.
using logistic regression and bootstrap aggregation, as above. Model fit was confirmed using Hosmer-Lemeshow goodness of fit tests. The large transfusion requirement models were further investigated using bootstrap forest partitioning methods with 100 trees in the forest, 16 terms sampled per split, and minimum split size of 5. Twenty percent of the population sample was withheld for validation purposes. The 15 most important classification variables found by the bootstrap forest technique were compared with results from the logistic regression analyses, and the top 10 of these were also input to a classification and regression tree (CART) analysis. Resulting trees for predicting >20 and >30 units of RBC + CS were depicted.10 The CART analyses used 8 steps with
www.anesthesia-analgesia.org
429
Transfusion Requirements During Liver Transplantation
Table 5. Predictors of Units of Red Blood Cell + Cell Salvage Usage Variable
Estimate ± SE
Older recipient age at transplant (y) Recipient diagnosis: HCC + malignant liver tumors Recipient pretransplant labs: Platelet count INR Total bilirubin Creatinine Donor type: expanded criteria donor Surgeon X
P 0.004 20 units]. INR = prothrombin international normalized ratio; MELD = Model for EndStage Liver Disease Score; CMV= cytomegalovirus. a All factors are pretransplant recipient factors except as noted.
430 www.anesthesia-analgesia.org
likelihood ratio χ2 tests used to determine the best splitting variables at each step. No minimum number of cases were required for final branches in the trees. Risk unadjusted survival was estimated nonparametrically by the method of Kaplan and Meier and parametrically by a multiphase hazard decomposition method.11 Nonproportional multiphase, multivariable hazard methodology11 was used to identify recipient, donor, and transplant procedure variables associated with each hazard phase simultaneously. Bootstrap aggregation (bagging) was used for variable selection as described above. Analysis of intraoperative RBC + CS usage included recipient factors (age, gender, race, height, weight, body mass index, body surface area, blood type, Rh factor, primary diagnosis [primary cause of liver failure], previous abdominal surgery [yes/no], cytomegalovirus status, transjugular intrahepatic portosystemic shunt, total bilirubin, serum creatinine, assigned MELD Scorea, pretransplant INR, and The assigned MELD Score is a composite of the biologic MELD Score (which is determined using the patient’s most recent bilirubin, prothrombin INR time and creatinine values) and special case exception points which are assigned for urgent situations including the presence of hepatocellular carcinoma.
a
anesthesia & analgesia
Table 9. Predictors of Red Blood Cell + Cell Salvage Usage >30 Units by Logistic Regression Analysis Variablea MELD Score (assigned) Platelet count Surgeon Y Surgeon Z
Odds ratio (95% CI) 1.07 (1.02–1.11) 0.993 (0.987–0.998) 2.64 (1.35–5.17) 3.86 (1.53–9.71)
P 0.004 0.006 0.004 0.004
Bootstrap % 64 75 91 91
Seventy-five of 804 patients (9.3%) required >30 units. MELD = model for end-stage liver disease score. a All factors are pretransplant recipient factors except as noted. See Appendix for interpretation of odds ratios.
platelets), donor factors (age, gender, race, expanded criteria donor, cause of death, circumstance of death, and mechanism of death), and procedure variables (surgeon, date of transplant, liver-only transplant, whole liver transplant). Survival analysis incorporated the above variables plus donor blood type and Rh factor, donor/recipient matching variables (gender match [male to male, female to female, female to male, male to female], Rh compatibility, and length of surgery [time from the surgical incision to closure]), graft cold ischemic time, and blood products given during transplantation.
RESULTS
Recipient, donor, and transfusion data are presented in Tables 1–4. Cryoprecipitate was not analyzed due the majority of this product being given to 3 patients who received massive transfusions and were later dropped from further survival analysis. Ten surgeons performed the OLTs with minimum 29, and maximum 122 transplants performed by each.
RBC + CS
The following factors were associated with increased RBC and CS administration during OLT (Table 5): older recipient age (P = 0.004); lower platelet count (P < 0.0001); greater recipient pretransplant INR (P < 0.0001), total bilirubin (P = 0.007), and Cr (P = 0.0005); and expanded criteria donor (P = 0.03). Two factors were associated with decreased combined RBC and CS administration: recipient diagnosis of hepatocellular carcinoma (P < 0.0001) and Surgeon X (P < 0.0001). The r2 for the model was small, with these predictors accounting for only 22% of the variance in amount of blood product needed. Analyses of blood product requirements >20 units and >30 units of RBC + CS during OLT (approximate 80th and 90th percentiles of usage) are shown in Tables 6–11. Since the 75 patients with RBC + CS >30 were also among the 156
Table 11. Predictors of Red Blood Cell + Cell Salvage Usage >30 Units by Bootstrap Forest Analysis Importance rank 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Variablea Creatinine Age Platelet count MELD Score (assigned) Total bilirubin Interval 1/1/2001 to transplant Donor age Height INR Weight Body mass index Blood type O Blood type A Body surface area Surgeon Y
G2 960 912 769 754 598 556 418 377 376 371 370 321 310 245 238
Most important 15 factors (of 66 considered) in partitioning data into patients requiring >30 vs ≤30 units RBC + CS [75 of 804 patients (9.3%) required >30 units]. MELD = model for end-stage liver disease score; INR = international normalized ratio. a All factors are pretransplant recipient factors except as noted.
patients with RBC + CS >20, we expected some compatibility in the models developed.
RBC + CS >20
Logistic regression analysis found the following factors to be associated with increased odds of requiring >20 units of RBC + CS (Table 6): higher levels of pretransplant Cr (P < 0.0001) and INR (P = 0.003), lower pretransplant platelet count (P = 0.002), previous abdominal surgery (P = 0.003), blood type A (P = 0.007), and surgeon Y (P = 0.009). Model c (concordance) statistic = 0.70 (95% CI, 0.66–0.75). Table 7 shows univariate descriptive summaries of the logistic regression predictors by RBC + CS >20 units. Table 8 shows the 15 most important variables found by bootstrap forest analysis of RBC + CS >20 units. There is good overlap in the 2 analyses: the logistic regression variables Cr, platelet count, and INR are the top 3 variables used for partitioning in the bootstrap forest analysis, and previous abdominal surgery is the 15th ranked variable. Blood type and surgeon were not important in the bootstrap forest analysis so may be of lesser importance in predicting RBC + CS >20. As the 15th ranked variable, previous abdominal surgery may also be of only marginal importance. The CART analysis yielded the representative tree shown in Figure 1 with the following variables
Table 10. Logistic Regression Predictors by Red Blood Cell + Cell Salvage Usage RBC + CS ≤30 Variablea MELD Score (assigned) Platelet count Surgeon Y Surgeon Z
Available N 559 729 727 727
Valueb 20.9 ± 5.5 97.2 ± 69.5 61 (8.4) 22 (3.0)
RBC + CS >30 Available N 55 75 75 75
Valueb 23.3 ± 6.0 74.1 ± 51.8 13 (17) 7 (9.3)
P 0.003 0.002 0.01 0.005
MELD = model for end-stage liver disease score. a All factors are pretransplant recipient factors except as noted. b Values are mean ± SD or count (%).
February 2014 • Volume 118 • Number 2
www.anesthesia-analgesia.org
431
Transfusion Requirements During Liver Transplantation
Figure 1. Partition tree with 8 splits for red blood cell (RBC) + cell salvage (CS) >20 units. Red indicates patients requiring >20 units. Candidate variables are the top 10 from Table 8.
used in the 8 partitions: Cr, total bilirubin, INR, MELD Score, platelet count, and body weight. After the 8 partitions, 620 of 648 (96%) patients with RBC + CS ≤20 were correctly classified, and 50 of 156 (32%) patients with RBC + CS >20 were correctly classified. Classification improves with more partitions; however, trees with many partitions can be difficult to interpret and likely would not replicate well on new data.
RBC + CS >30
Logistic regression analysis found the following factors to be associated with increased odds of requiring >30 units of RBC + CS (Table 9): higher assigned MELD Score (P = 0.004), lower pretransplant platelet count (P = 0.006), surgeon Y (P = 0.004), and surgeon Z (P = 0.004). Model c (concordance) statistic = 0.67 (95% CI, 0.60–0.73). Table 10 shows univariate descriptive summaries of the logistic regression predictors by RBC + CS >30 units. Table 11 shows the 15 most important variables found by bootstrap forest analysis of RBC + CS >30 units. There is moderate consistency in the 2 analyses: the logistic regression variables platelet count and MELD Score were the third and fourth most important variables used for partitioning in the bootstrap forest analysis; however, the bootstrap forest analysis found Cr and age to be of greater importance than platelet count or MELD Score. The bootstrap forest method identified Surgeon Y as only the 15th most important variable and did not identify Surgeon Z at all. Note that the bootstrap forest method identified the interval from January 1, 2001 to transplant as important. Surgeon Y’s transplants were mostly performed much later than average in the sequence of transplants, so there is confounding
432 www.anesthesia-analgesia.org
between surgeon and date of transplant. It is not surprising that 2 analytic methods might choose differently among the confounded factors. The CART analysis yielded the representative tree shown in Figure 2 with the following variables used in the 8 partitions: Cr, assigned MELD Score, height, platelet count, years since January 1, 2001, and age. After the 8 partitions, 727 of 729 (99.7%) patients with RBC + CS ≤30 were correctly classified; however, only 8 of 75 (11%) patients with RBC + CS >30 were correctly classified. Creating a tree with more splits did not improve prediction appreciably.
Survival Analysis
Mean follow-up time for survivors in the study was 3.6 ± 2.3 years. Because the goal of the analysis was to assess whether the quantity of blood products transfused affected survival, it was important to consider whether high numbers of units may have been transfused at the time of transplant in order to rescue dying patients. If so, this could bias the conclusions toward finding association of transfusions with death. Indeed, we found that among the 10 patients who died within a day of transplantation, 8 had 40 or more units of RBC + CS transfused, placing them in the top 5% of RBC + CS usage among study patients. Cases that resulted in mortality within the first postoperative day (10 patients) were excluded from the primary survival analysis. However, the analysis was repeated with the 10 patients included to assure that the results from the 2 analyses were compatible. Instantaneous risk of death after liver transplantation consisted of an early hazard phase of greatest risk, followed, after approximately 9 months, by a late hazard phase of nearly constant risk (Figs. 3, A and B). The hazard of death
anesthesia & analgesia
Figure 2. Partition tree with 8 splits for red blood cell (RBC) + cell salvage (CS)> 30 units. Red indicates patients requiring >30 Units. Candidate variables are the top 10 from Table 11.
peaked at about 2 weeks posttransplant. The results of the multivariable hazard modeling are shown in Table 12. More units of blood transfused, greater pretransplant INR, older donor age and donor race other than African American or Caucasian were all associated with a greater hazard of death in the early posttransplant period. A primary diagnosis of viral hepatitis and disease etiology of liver cancer were associated with a greater hazard of death in the ensuing late hazard phase. The effects of donor race and units of RBC + CS transfused on predicted 1-year survival are shown graphically in Figure 4. Kaplan-Meier unadjusted survival curves by units of RBC + CS transfused are shown in Figure 5. As demonstrated in Figure 5, patients receiving >15 units of RBC + CS have markedly worse survival after OLT.
DISCUSSION
During the evolution and establishment of OLT as the most effective treatment option for end-stage liver disease, there has been significant interest in the literature in elucidating the predictors of blood transfusion during this surgery.2,3 Intraoperative blood loss and subsequent blood transfusion can influence multiple postoperative outcomes (i.e., need for re-exploration, incidence of sepsis, poorer patient and graft survival); however, cause and effect relationships have not been clearly established. The difficulties in predicting transfusion requirements during OLT are many. Quantification of intraoperative
February 2014 • Volume 118 • Number 2
surgical conditions (intra-abdominal adhesion, venous collaterals, etc.) is considered by some to be very important but may be difficult to anticipate in advance.2 These factors, although intuitively easy to appreciate, do not reliably translate into intraoperative blood loss and transfusion requirements. Previous investigations have attempted to identify preoperative predictors of blood transfusion; however, the predictive value of preoperative variables and patient characteristics were inconsistent and weak at best.2,3 Transfusion requirements depend not only on the intraoperative blood loss but also on the threshold for when transfusions of different products are initiated. Hevesi et al.12 demonstrated that requirements for RBC and FFP can be reduced almost 2- and 3-fold, respectively, if the anesthesia team universally followed protocols including goal-directed transfusion practices. Therefore, comparison of intraoperative transfusion requirements from different transplant centers may be inherently biased by an inability to account for differences in transfusion triggers and clinical practices. Consequently, the predictive models developed in one institution may hardly, if ever, be applicable in the others. For example, Massicotte et al developed a predictive model to estimate the probability of RBC transfusion based on 3 individually weighted risk factors derived from their cohort of OLT patients (transfusion of FFP, inability to perform phlebotomy, and starting hemoglobin). Despite the high c-statistic of that model, its applicability in other centers is questionable because the rate of transfusion in the program where
www.anesthesia-analgesia.org
433
Transfusion Requirements During Liver Transplantation
Figure 4. Predicted 1-year survival by donor race and number of units red blood cell (RBC) + cell salvage (CS) transfused. Long dashed line represents donor race other than African American or Caucasian; solid line represents African American or Caucasian donor races. Corresponding dotted/short-dashed lines represent 68% confidence limits, equivalent to ±1 SE. Plot assumes patient international normalized ratio (INR) of 1.5, disease etiology not liver cancer, primary diagnosis not viral hepatitis, and donor age of 44 years.
Figure 3. Death after deceased donor liver transplantation. A, survival. Each square represents a death, and numbers in parentheses indicate patients remaining at risk. Solid lines within dashed 68% confidence limits represent parametric estimates. B, Instantaneous risk of death (hazard function). Solid lines enclosed within dashed confidence limits represent parametric estimates. Note expanded horizontal scale.
the model was developed was extremely low with only 19.5% of the patients transfused at all.4 In contrast, many other centers report significantly higher transfusion rates. Our results demonstrate that a model created to predict transfusion requirements for patients undergoing OLT is unreliable, and hence, the practical utility remains minimal. Although there was substantial overlap in the sets of variables that predicted >20 and >30 units of RBCs among our models, no set of variables provided good prediction for large transfusion requirements. The CART analysis model
was best in identifying patients with intraoperative RBC + CS requirements 20 units of RBC and CS) proved difficult. We did confirm, however, a strong association between intraoperative transfusion and early mortality, as well as quantified the risk of transfusion on posttransplant survival. The main limitation of this study is its retrospective nature and the fact that we could not account for all the factors influencing transfusion requirements, including transfusion triggers. Also, the results of this study reflect our institutional practice and should not be interpreted February 2014 • Volume 118 • Number 2
outside of this context. It is possible, although unlikely, that a similar analysis conducted on patients from other institutions could produce different results, simply based on differences in practice and the population presenting for OLT. In conclusion, prediction of intraoperative blood product requirements based on preoperatively available variables is unreliable in all transfusion range categories analyzed. We were able to quantify a strong measurable association between intraoperative blood transfusion and recipient mortality after OLT. E
APPENDIX INTERPRETATION OF REGRESSION, LOGISTIC REGRESSION, AND SURVIVAL MODEL EFFECTS Multivariable Regression: Predictors of red blood cell (RBC) + cell salvage (CS) usage (Table 5). Recipient Age at Transplant For each year older a patient is, there is a 0.9% increase in the predicted number of units of RBC + CS required. For every increase of 10 years in recipient age, there is a 9% increase in the estimated number of units of RBC + CS required. Recipient Diagnosis: HCC + Malignant Liver Tumor For patients with this diagnosis, the predicted need for blood products is 30% lower than for patients with other diagnoses. Recipient Pretransplant Platelet Count Each increase of 1k/mm3 in pretransplant platelet count results in a 0.3% decrease in the predicted number of units of RBC + CS required. Each increase of 10k/mm3 in pretransplant platelet count results in a 3% decrease in the predicted number of units of RBC + CS required. Recipient Pretransplant INR Each increase of 1 unit of INR results in a 36% increase in predicted number of units of RBC + CS required.
www.anesthesia-analgesia.org
435
Transfusion Requirements During Liver Transplantation
Recipient Pretransplant Total Bilirubin Each increase of 1 unit of total bilirubin results in a 0.9% increase in predicted number of units of RBC + CS required.
Surgeon Y Having transplant Surgeon Y increases odds of needing >30 units of RBC + CS by 164% (53%–871%).
Recipient Pretransplant Creatinine Each increase of 1 unit of creatinine results in a 9% increase in predicted number of units of RBC + CS required.
Surgeon Z Having transplant Surgeon Z increases odds of needing > 30 units of RBC + CS by 286% (66%-986%).
Expanded Criteria Donor Use of expanded criteria donor results in 26% increase in predicted number of units of RBC + CS required.
Survival Analysis Hazard Model (Table 12) (Effect size percentages are followed by 95% confidence intervals on the effect sizes).
Surgeon X When Surgeon X performs the transplant, the predicted number of units of RBC + CS required is 30% lower than when another surgeon performs the transplant.
RBC + CS Units Transfused Each additional unit transfused increases the hazard of death in the early hazard phase by 4.3% (3.0%–5.6%).
Logistic Regression: Predictors of RBC + CS usage > 20 Units (Table 6) (Effect size percentages are followed by 95% confidence intervals on the effect sizes). Recipient Pretransplant Creatinine Each 1 unit increase in creatinine increases odds of needing >20 units of RBC + CS by 34% (17%–53%). Recipient Pretransplant INR Each 1 unit increase in INR increases odds of needing >20 units of RBC + CS by 81% (23%–167%). Recipient Pretransplant Platelet Count Each decrease of 1k/mm3 in pretransplant platelet count results in a 0.6% (0.2%–0.9%) increase in the odds of needing >20 units of RBC + CS. Each decrease of 10k/mm3 in pretransplant platelet count results in a 5.8% (2.0%–8.6%) increase in the odds of needing >20 units of RBC + CS. Recipient Previous Abdominal Surgery History of previous abdominal surgery increases odds of needing >20 units of RBC + CS by 79% (23%–162%). Recipient Blood Type A Blood type A increases odds of needing >20 units of RBC + CS by 68% (15%–143%). Surgeon Y Having transplant Surgeon Y increases odds of needing >20 units of RBC + CS by 112% (21%–272%). Logistic Regression: Predictors of RBC + CS usage > 30 Units (Table 9) (Effect size percentages are followed by 95% confidence intervals on the effect sizes). Recipient MELD Score Each 1 unit increase in the assigned MELD Score increases odds of needing >30 units of RBC + CS by 7% (2%–11%). Recipient Pretransplant Platelet Count Each decrease of 1k/mm3 in pretransplant platelet count results in a 0.7% (0.2%–1.3%) increase in the odds of needing >30 units of RBC + CS. Each decrease of 10k/mm3 in pretransplant platelet count results in a 6.8% (1.9%–12.3%) increase in the odds of needing >30 units of RBC + CS.
436 www.anesthesia-analgesia.org
INR Each increase of 0.1 in INR is associated with a 3.3% (1.4%– 5.3%) increase in hazard of death in the early hazard phase. Donor Age Each increase of 1 year in donor age is associated with a 3.0% (1.2%–4.9%) increase in hazard of death in the early hazard phase. Each increase of 10 years in donor age is associated with a 34% (12%–61%) increase in hazard of death in the early hazard phase. Donor Race Other than African American or Caucasian Patients whose donor is of race other than African American or Caucasian have 4.4 (1.6–11.7) times the hazard of death in the early hazard phase as patients with African American or Caucasian donors. Diagnosis Viral Hepatitis Patients with a primary diagnosis of viral hepatitis have 3.0 (1.4–6.1) times the hazard of death in the late hazard phase as patients with other diagnoses. Disease Etiology Liver Cancer Patients with disease etiology of liver cancer have 4.2 (2.0– 8.8) times the hazard of death in the late hazard phase as patients with other disease etiologies. DISCLOSURES
Name: Jacek B. Cywinski, MD. Contribution: This author helped design and conduct the study, analyze the data, and write the manuscript. Attestation: Jacek B. Cywinski has seen the original study data, reviewed the analysis of the data, approved the final manuscript, and is the author responsible for archiving the study files. Name: Joan M. Alster, MS. Contribution: This author helped analyze the data and write the manuscript. Attestation: Joan M. Alster has seen the original study data, reviewed the analysis of the data, and approved the final manuscript. Name: Charles Miller, MD. Contribution: This author helped write the manuscript. Attestation: Charles Miller has seen the original study data, reviewed the analysis of the data, and approved the final manuscript.
anesthesia & analgesia
Name: David P. Vogt, MD. Contribution: This author helped design and conduct the study. Attestation: David P. Vogt has seen the original study data, reviewed the analysis of the data, and approved the final manuscript. Name: Brian M. Parker, MD. Contribution: This author helped design and conduct the study, analyze the data, and write the manuscript. Attestation: Brian M. Parker has seen the original study data, reviewed the analysis of the data, and approved the final manuscript. This manuscript was handled by: Edward C. Nemergut, MD. ACKNOWLEDGMENTS
The authors wish to thank Lucy Thuita, MS, for data management and programming and Mrs. Tanya Smith for her editorial assistance. REFERENCES 1. Deakin M, Gunson BK, Dunn JA, McMaster P, Tisone G, Warwick J, Buckels JA. Factors influencing blood transfusion during adult liver transplantation. Ann R Coll Surg Engl 1993;75:339–44 2. Findlay JY, Rettke SR. Poor prediction of blood transfusion requirements in adult liver transplantations from preoperative variables. J Clin Anesth 2000;12:319–23 3. Steib A, Freys G, Lehmann C, Meyer C, Mahoudeau G. Intraoperative blood losses and transfusion requirements during adult liver transplantation remain difficult to predict. Can J Anaesth 2001;48:1075–9 4. Massicotte L, Sassine MP, Lenis S, Roy A. Transfusion predictors in liver transplant. Anesth Analg 2004;98:1245–51, table of contents
February 2014 • Volume 118 • Number 2
5. Hendriks HG, van der Meer J, de Wolf JT, Peeters PM, Porte RJ, de Jong K, Lip H, Post WJ, Slooff MJ. Intraoperative blood transfusion requirement is the main determinant of early surgical re-intervention after orthotopic liver transplantation. Transpl Int 2005;17:673–9 6. Schroeder RA, Johnson LB, Plotkin JS, Kuo PC, Klein AS. Total blood transfusion and mortality after orthotopic liver transplantation. Anesthesiology 1999;91:329–30 7. Massicotte L, Sassine MP, Lenis S, Seal RF, Roy A. Survival rate changes with transfusion of blood products during liver transplantation. Can J Anaesth 2005;52:148–55 8. de Boer MT, Molenaar IQ, Hendriks HG, Slooff MJ, Porte RJ. Minimizing blood loss in liver transplantation: progress through research and evolution of techniques. Dig Surg 2005;22:265–75 9. Breiman L. Bagging predictors. Mach Learn 1996;24:123–40 10. Breiman L. Random forests. Mach Learn 2001;45:5–32 11. Blackstone EH, Naftel DC, Turner ME, Jr. The decomposition of time-varying hazard into phases, each incorporating a separate stream of concomitant information. J Am Stat Assoc 1986;81:615–24 12. Hevesi ZG, Lopukhin SY, Mezrich JD, Andrei AC, Lee M. Designated liver transplant anesthesia team reduces blood transfusion, need for mechanical ventilation, and duration of intensive care. Liver Transpl 2009;15:460–5 13. McCluskey SA, Karkouti K, Wijeysundera DN, Kakizawa K, Ghannam M, Hamdy A, Grant D, Levy G. Derivation of a risk index for the prediction of massive blood transfusion in liver transplantation. Liver Transpl 2006;12:1584–93 1 4. Pereboom IT, de Boer MT, Haagsma EB, Hendriks HG, Lisman T, Porte RJ. Platelet transfusion during liver transplantation is associated with increased postoperative mortality due to acute lung injury. Anesth Analg 2009;108:1083–91 15. Ramos E, Dalmau A, Sabate A, Lama C, Llado L, Figueras J, Jaurrieta E. Intraoperative red blood cell transfusion in liver transplantation: influence on patient outcome, prediction of requirements, and measures to reduce them. Liver Transpl 2003;9:1320–7
www.anesthesia-analgesia.org
437
Prediction of Intraoperative Transfusion Requirements During Orthotopic Liver Transplantation and the Influence on Postoperative Patient Survival Jacek B. Cywinski, MD,* Joan M. Alster, MS,† Charles Miller, MD,‡ David P. Vogt, MD,‡ and Brian M. Parker, MD* BACKGROUND: Predicting blood product transfusion requirements during orthotopic liver transplantation (OLT) remains difficult. Our primary aim in this study was to determine which patient variables best predict recipient risk for large blood transfusion requirements during OLT. The secondary aim was to determine whether the amount of blood products transfused during OLT impacted patient survival. METHODS: Eight hundred four primary adult OLTs performed during a 9-year period were retrospectively analyzed, and predictive models were developed for blood product usage, usage >20 and usage >30 units of red blood cells (RBCs) plus cell salvage (CS). For survival analysis, potential predictors included all blood products administered during OLT. RESULTS: For analyses of RBC + CS usage, we used several statistical techniques: regression analysis, logistic regression, and classification and regression tree analysis. Several preoperative factors were highly statistically significant predictors of intraoperative blood product usage in each of the analyses, namely lower platelet count and higher Model for End-Stage Liver Disease Score or one or more of its components (creatinine, total bilirubin, international normalized ratio). Despite these highly significant associations, the models were unable to predict reliably that patients might require the largest amount of blood products during OLT. For example, the classification and regression tree analyses were able to predict only 32% and 11% of patients requiring >20 and >30 units of RBC + CS, respectively. Survival analysis demonstrated poorer survival among patients receiving larger amounts of RBC + CS during OLT. CONCLUSION: Prediction of intraoperative blood product requirements based on preoperatively available variables is unreliable; however, there is a strong measurable association between transfusion and postoperative mortality. (Anesth Analg 2014;118:428–37)
P
redicting blood product transfusion requirements during orthotopic liver transplantation (OLT) has remained difficult. Because OLT is so resource intensive, understanding which variables predict transfusion requirements could allow improved blood product allocation and utilization. Data published both before, and after, the use of the Model for End-Stage Liver Disease (MELD) Score have shown select preoperative factors to be only minimally predictive of intraoperative transfusion requirements.1–4 Some of the preoperative variables reported in the literature affecting intraoperative transfusion requirements include: recipient age and weight, etiology of liver disease, starting hemoglobin level, prothrombin international normalized ratio (INR) and partial thromboplastin time, platelet count, serum creatinine (Cr), and presence of ascites. Unfortunately, many of these variables have not
From the Departments of *General Anesthesiology and Transplant Center, †Quantitative Health Sciences, and ‡Hepato-pancreato-biliary and Transplant Surgery, Cleveland Clinic, Cleveland, Ohio. Accepted for publication July 17, 2013. Funding: Supported by internal funds. The authors declare no conflicts of interest. Reprints will not be available from the authors Address correspondence to Brian M. Parker, MD, Cleveland Clinic, 9500 Euclid Ave., Cleveland, OH 44195. Address e-mail to [email protected]. Copyright © 2014 International Anesthesia Research Society DOI: 10.1213/ANE.0b013e3182a76f19
428 www.anesthesia-analgesia.org
consistently been found to be associated with transfusion needs during OLT. Higher transfusion requirements during OLT are associated with worse postoperative outcomes including prolonged length of intensive care unit stay, hospital stay, and increased mortality;5 however, these relationships have not been well characterized.6,7 Furthermore, an association rather than causality was established in these studies with varying strengths of correlation. Interestingly, when survival outcomes are compared among centers with different rates of blood product transfusion, centers with higher blood product administration rates do not necessarily have worse survival rates. This apparent anomaly highlights the fact that the relationship between transfusion and outcomes is complex and most likely dependent on factors unique to the particular transplant center.8 Recognizing that it may be very difficult to develop a single, reliable, and universally applicable model to predict transfusion requirements for patients undergoing OLT, the primary aim of this study was to determine which preoperatively available variables best predict recipient blood transfusion requirements during OLT in our program, with particular emphasis on predicting high transfusion requirements (>20 or >30 units of blood). Such models could identify the patients who consume the majority of the blood bank resources and possibly be modified to fit the practice of other institutions. The secondary aim of this study was to determine whether the amount of blood products transfused impacted survival after OLT. February 2014 • Volume 118 • Number 2
METHODS
After receiving IRB approval which waived the need for informed consent, 835 consecutive primary adult (age ≥18 years) OLTs performed from January 1, 2001 to June 30, 2010 at the Cleveland Clinic were retrospectively analyzed. During this time frame, 47 retransplants were performed; in these cases, we analyzed data from the first transplant only. Patients undergoing combined liver and other organ transplants (liver-kidney n = 36, liver-heart n = 3, liver-lung n = 3, liver-pancreas n = 2, liver-pancreas-small intestine n = 1) as well as split liver graft (n = 13) were included in the analysis. In addition, 24 living donor liver transplants were performed during this period. None of the patients undergoing retransplantation or living donor liver transplants was included in the analyses nor were 7 transplants for which we did not have blood product usage information. This resulted in a final analysis of 804 OLTs. All blood products administered intraoperatively to OLT recipients were analyzed, including allogeneic red blood cells (RBCs), fresh frozen plasma (FFP), cryoprecipitate, platelets, and intraoperative red blood cell salvage (CS) units returned to the patient. A predictive model was developed for RBC + CS (using RBC equivalents) usage during liver transplantation. Stepwise multivariable regression analysis was used to identify recipient, donor, and transplant procedure variables associated with blood product usage. The natural logarithm of (RBC + CS) was used to produce a nearly normally distributed analysis variable with stable error variance. Bootstrap aggregation (bagging) was used for variable selection,9 with P < 0.07 for entry into the models and P < 0.05 for variable retention. Variables appearing in at least 50% of bootstrap analyses were considered reliably statistically significant at P < 0.05 (median rule). Several transformations of continuous predictors (x2, 1/x, 1/x2, sqrt[x], ln[x]) were plotted against ln(RBC + CS) to check whether transformation could result in more nearly linear relationships between the predictors and ln(RBC + CS). Since the transformations did not result in significantly improved linearity, the predictors were analyzed on their original scales. In addition, goodness of fit of the regression model was confirmed by plotting residuals versus predicted values. To predict large transfusion requirements, use of >20 and >30 units of RBC + CS (approximately top 20% and 10% of blood product needs in our study) were modeled
Table 1. Recipient Data Variable Age (y) Gender: female BMI Blood type A B AB O Previous abdominal surgery INR Platelet count (k/mm3) TIPS present MELD Score
n (%) or mean ± SD 54 ± 10 241 (30) 28.9 ± 6.2 301 (37) 112 (14) 37 (4.6) 353 (44) 323 (42) 1.52 ± 0.7 95.1 ± 68.3 85 (11) 21 ± 5.6
BMI = body mass index; INR = prothrombin international normalized ratio; TIPS = transjugular intrahepatic portosystemic shunt; MELD = model for endstage liver disease.
February 2014 • Volume 118 • Number 2
Table 2. Primary Diagnosis of End-Stage Liver Disease in 804 Study Patients Diagnosis
n (%) 213 (26) 187 (23) 110 (14) 78 (9.7) 68 (8.5) 67 (8.3) 28 (3.5) 15 (1.9) 13 (1.6) 1 (0.12)
HCC Viral hepatitis Cholestatic Cryptogenic NASH Laennec’s Metabolic disease Autoimmune Fulminant hepatic failure Neuroendocrine tumor
HCC = hepatocellular carcinoma; NASH = nonalcoholic steatohepatitis.
Table 3. Donor Data Variable Donor age Gender: female BMI Blood type A B AB O Mechanism of death ICH/stroke Blunt trauma Gunshot wound Cardiovascular Drug intoxication Asphyxiation Other Unknown Cold ischemic time (min)
n (%) or mean ± SD 44 ± 17 317 (40) 27.3 ± 10 310 92 27 368
(39) (12) (3.4) (46)
386 (48) 168 (21) 61 (13) 61 (7.6) 21 (2.6) 16 (2.0) 42 (5.2) 7 (0.87) 460 ± 163
BMI = body mass index; ICH = intracranial hemorrhage.
Table 4. Intraoperative Transfusion Data Variable RBCs (units) CS (units) RBC + CS (units) FFP (units) Platelets (5 packs pooled)
Mean ± SD 8.49 ± 8.6 5.46 ± 6.3 13.9 ± 13.6 10.3 ± 10.6 3.58 ± 3.0
NB: CS was used for all patients undergoing OLT regardless of liver disease etiology. OLT = orthotopic liver transplant; RBCs = red blood cells; CS = cell salvage; FFP = fresh frozen plasma.
using logistic regression and bootstrap aggregation, as above. Model fit was confirmed using Hosmer-Lemeshow goodness of fit tests. The large transfusion requirement models were further investigated using bootstrap forest partitioning methods with 100 trees in the forest, 16 terms sampled per split, and minimum split size of 5. Twenty percent of the population sample was withheld for validation purposes. The 15 most important classification variables found by the bootstrap forest technique were compared with results from the logistic regression analyses, and the top 10 of these were also input to a classification and regression tree (CART) analysis. Resulting trees for predicting >20 and >30 units of RBC + CS were depicted.10 The CART analyses used 8 steps with
www.anesthesia-analgesia.org
429
Transfusion Requirements During Liver Transplantation
Table 5. Predictors of Units of Red Blood Cell + Cell Salvage Usage Variable
Estimate ± SE
Older recipient age at transplant (y) Recipient diagnosis: HCC + malignant liver tumors Recipient pretransplant labs: Platelet count INR Total bilirubin Creatinine Donor type: expanded criteria donor Surgeon X
P 0.004 20 units]. INR = prothrombin international normalized ratio; MELD = Model for EndStage Liver Disease Score; CMV= cytomegalovirus. a All factors are pretransplant recipient factors except as noted.
430 www.anesthesia-analgesia.org
likelihood ratio χ2 tests used to determine the best splitting variables at each step. No minimum number of cases were required for final branches in the trees. Risk unadjusted survival was estimated nonparametrically by the method of Kaplan and Meier and parametrically by a multiphase hazard decomposition method.11 Nonproportional multiphase, multivariable hazard methodology11 was used to identify recipient, donor, and transplant procedure variables associated with each hazard phase simultaneously. Bootstrap aggregation (bagging) was used for variable selection as described above. Analysis of intraoperative RBC + CS usage included recipient factors (age, gender, race, height, weight, body mass index, body surface area, blood type, Rh factor, primary diagnosis [primary cause of liver failure], previous abdominal surgery [yes/no], cytomegalovirus status, transjugular intrahepatic portosystemic shunt, total bilirubin, serum creatinine, assigned MELD Scorea, pretransplant INR, and The assigned MELD Score is a composite of the biologic MELD Score (which is determined using the patient’s most recent bilirubin, prothrombin INR time and creatinine values) and special case exception points which are assigned for urgent situations including the presence of hepatocellular carcinoma.
a
anesthesia & analgesia
Table 9. Predictors of Red Blood Cell + Cell Salvage Usage >30 Units by Logistic Regression Analysis Variablea MELD Score (assigned) Platelet count Surgeon Y Surgeon Z
Odds ratio (95% CI) 1.07 (1.02–1.11) 0.993 (0.987–0.998) 2.64 (1.35–5.17) 3.86 (1.53–9.71)
P 0.004 0.006 0.004 0.004
Bootstrap % 64 75 91 91
Seventy-five of 804 patients (9.3%) required >30 units. MELD = model for end-stage liver disease score. a All factors are pretransplant recipient factors except as noted. See Appendix for interpretation of odds ratios.
platelets), donor factors (age, gender, race, expanded criteria donor, cause of death, circumstance of death, and mechanism of death), and procedure variables (surgeon, date of transplant, liver-only transplant, whole liver transplant). Survival analysis incorporated the above variables plus donor blood type and Rh factor, donor/recipient matching variables (gender match [male to male, female to female, female to male, male to female], Rh compatibility, and length of surgery [time from the surgical incision to closure]), graft cold ischemic time, and blood products given during transplantation.
RESULTS
Recipient, donor, and transfusion data are presented in Tables 1–4. Cryoprecipitate was not analyzed due the majority of this product being given to 3 patients who received massive transfusions and were later dropped from further survival analysis. Ten surgeons performed the OLTs with minimum 29, and maximum 122 transplants performed by each.
RBC + CS
The following factors were associated with increased RBC and CS administration during OLT (Table 5): older recipient age (P = 0.004); lower platelet count (P < 0.0001); greater recipient pretransplant INR (P < 0.0001), total bilirubin (P = 0.007), and Cr (P = 0.0005); and expanded criteria donor (P = 0.03). Two factors were associated with decreased combined RBC and CS administration: recipient diagnosis of hepatocellular carcinoma (P < 0.0001) and Surgeon X (P < 0.0001). The r2 for the model was small, with these predictors accounting for only 22% of the variance in amount of blood product needed. Analyses of blood product requirements >20 units and >30 units of RBC + CS during OLT (approximate 80th and 90th percentiles of usage) are shown in Tables 6–11. Since the 75 patients with RBC + CS >30 were also among the 156
Table 11. Predictors of Red Blood Cell + Cell Salvage Usage >30 Units by Bootstrap Forest Analysis Importance rank 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Variablea Creatinine Age Platelet count MELD Score (assigned) Total bilirubin Interval 1/1/2001 to transplant Donor age Height INR Weight Body mass index Blood type O Blood type A Body surface area Surgeon Y
G2 960 912 769 754 598 556 418 377 376 371 370 321 310 245 238
Most important 15 factors (of 66 considered) in partitioning data into patients requiring >30 vs ≤30 units RBC + CS [75 of 804 patients (9.3%) required >30 units]. MELD = model for end-stage liver disease score; INR = international normalized ratio. a All factors are pretransplant recipient factors except as noted.
patients with RBC + CS >20, we expected some compatibility in the models developed.
RBC + CS >20
Logistic regression analysis found the following factors to be associated with increased odds of requiring >20 units of RBC + CS (Table 6): higher levels of pretransplant Cr (P < 0.0001) and INR (P = 0.003), lower pretransplant platelet count (P = 0.002), previous abdominal surgery (P = 0.003), blood type A (P = 0.007), and surgeon Y (P = 0.009). Model c (concordance) statistic = 0.70 (95% CI, 0.66–0.75). Table 7 shows univariate descriptive summaries of the logistic regression predictors by RBC + CS >20 units. Table 8 shows the 15 most important variables found by bootstrap forest analysis of RBC + CS >20 units. There is good overlap in the 2 analyses: the logistic regression variables Cr, platelet count, and INR are the top 3 variables used for partitioning in the bootstrap forest analysis, and previous abdominal surgery is the 15th ranked variable. Blood type and surgeon were not important in the bootstrap forest analysis so may be of lesser importance in predicting RBC + CS >20. As the 15th ranked variable, previous abdominal surgery may also be of only marginal importance. The CART analysis yielded the representative tree shown in Figure 1 with the following variables
Table 10. Logistic Regression Predictors by Red Blood Cell + Cell Salvage Usage RBC + CS ≤30 Variablea MELD Score (assigned) Platelet count Surgeon Y Surgeon Z
Available N 559 729 727 727
Valueb 20.9 ± 5.5 97.2 ± 69.5 61 (8.4) 22 (3.0)
RBC + CS >30 Available N 55 75 75 75
Valueb 23.3 ± 6.0 74.1 ± 51.8 13 (17) 7 (9.3)
P 0.003 0.002 0.01 0.005
MELD = model for end-stage liver disease score. a All factors are pretransplant recipient factors except as noted. b Values are mean ± SD or count (%).
February 2014 • Volume 118 • Number 2
www.anesthesia-analgesia.org
431
Transfusion Requirements During Liver Transplantation
Figure 1. Partition tree with 8 splits for red blood cell (RBC) + cell salvage (CS) >20 units. Red indicates patients requiring >20 units. Candidate variables are the top 10 from Table 8.
used in the 8 partitions: Cr, total bilirubin, INR, MELD Score, platelet count, and body weight. After the 8 partitions, 620 of 648 (96%) patients with RBC + CS ≤20 were correctly classified, and 50 of 156 (32%) patients with RBC + CS >20 were correctly classified. Classification improves with more partitions; however, trees with many partitions can be difficult to interpret and likely would not replicate well on new data.
RBC + CS >30
Logistic regression analysis found the following factors to be associated with increased odds of requiring >30 units of RBC + CS (Table 9): higher assigned MELD Score (P = 0.004), lower pretransplant platelet count (P = 0.006), surgeon Y (P = 0.004), and surgeon Z (P = 0.004). Model c (concordance) statistic = 0.67 (95% CI, 0.60–0.73). Table 10 shows univariate descriptive summaries of the logistic regression predictors by RBC + CS >30 units. Table 11 shows the 15 most important variables found by bootstrap forest analysis of RBC + CS >30 units. There is moderate consistency in the 2 analyses: the logistic regression variables platelet count and MELD Score were the third and fourth most important variables used for partitioning in the bootstrap forest analysis; however, the bootstrap forest analysis found Cr and age to be of greater importance than platelet count or MELD Score. The bootstrap forest method identified Surgeon Y as only the 15th most important variable and did not identify Surgeon Z at all. Note that the bootstrap forest method identified the interval from January 1, 2001 to transplant as important. Surgeon Y’s transplants were mostly performed much later than average in the sequence of transplants, so there is confounding
432 www.anesthesia-analgesia.org
between surgeon and date of transplant. It is not surprising that 2 analytic methods might choose differently among the confounded factors. The CART analysis yielded the representative tree shown in Figure 2 with the following variables used in the 8 partitions: Cr, assigned MELD Score, height, platelet count, years since January 1, 2001, and age. After the 8 partitions, 727 of 729 (99.7%) patients with RBC + CS ≤30 were correctly classified; however, only 8 of 75 (11%) patients with RBC + CS >30 were correctly classified. Creating a tree with more splits did not improve prediction appreciably.
Survival Analysis
Mean follow-up time for survivors in the study was 3.6 ± 2.3 years. Because the goal of the analysis was to assess whether the quantity of blood products transfused affected survival, it was important to consider whether high numbers of units may have been transfused at the time of transplant in order to rescue dying patients. If so, this could bias the conclusions toward finding association of transfusions with death. Indeed, we found that among the 10 patients who died within a day of transplantation, 8 had 40 or more units of RBC + CS transfused, placing them in the top 5% of RBC + CS usage among study patients. Cases that resulted in mortality within the first postoperative day (10 patients) were excluded from the primary survival analysis. However, the analysis was repeated with the 10 patients included to assure that the results from the 2 analyses were compatible. Instantaneous risk of death after liver transplantation consisted of an early hazard phase of greatest risk, followed, after approximately 9 months, by a late hazard phase of nearly constant risk (Figs. 3, A and B). The hazard of death
anesthesia & analgesia
Figure 2. Partition tree with 8 splits for red blood cell (RBC) + cell salvage (CS)> 30 units. Red indicates patients requiring >30 Units. Candidate variables are the top 10 from Table 11.
peaked at about 2 weeks posttransplant. The results of the multivariable hazard modeling are shown in Table 12. More units of blood transfused, greater pretransplant INR, older donor age and donor race other than African American or Caucasian were all associated with a greater hazard of death in the early posttransplant period. A primary diagnosis of viral hepatitis and disease etiology of liver cancer were associated with a greater hazard of death in the ensuing late hazard phase. The effects of donor race and units of RBC + CS transfused on predicted 1-year survival are shown graphically in Figure 4. Kaplan-Meier unadjusted survival curves by units of RBC + CS transfused are shown in Figure 5. As demonstrated in Figure 5, patients receiving >15 units of RBC + CS have markedly worse survival after OLT.
DISCUSSION
During the evolution and establishment of OLT as the most effective treatment option for end-stage liver disease, there has been significant interest in the literature in elucidating the predictors of blood transfusion during this surgery.2,3 Intraoperative blood loss and subsequent blood transfusion can influence multiple postoperative outcomes (i.e., need for re-exploration, incidence of sepsis, poorer patient and graft survival); however, cause and effect relationships have not been clearly established. The difficulties in predicting transfusion requirements during OLT are many. Quantification of intraoperative
February 2014 • Volume 118 • Number 2
surgical conditions (intra-abdominal adhesion, venous collaterals, etc.) is considered by some to be very important but may be difficult to anticipate in advance.2 These factors, although intuitively easy to appreciate, do not reliably translate into intraoperative blood loss and transfusion requirements. Previous investigations have attempted to identify preoperative predictors of blood transfusion; however, the predictive value of preoperative variables and patient characteristics were inconsistent and weak at best.2,3 Transfusion requirements depend not only on the intraoperative blood loss but also on the threshold for when transfusions of different products are initiated. Hevesi et al.12 demonstrated that requirements for RBC and FFP can be reduced almost 2- and 3-fold, respectively, if the anesthesia team universally followed protocols including goal-directed transfusion practices. Therefore, comparison of intraoperative transfusion requirements from different transplant centers may be inherently biased by an inability to account for differences in transfusion triggers and clinical practices. Consequently, the predictive models developed in one institution may hardly, if ever, be applicable in the others. For example, Massicotte et al developed a predictive model to estimate the probability of RBC transfusion based on 3 individually weighted risk factors derived from their cohort of OLT patients (transfusion of FFP, inability to perform phlebotomy, and starting hemoglobin). Despite the high c-statistic of that model, its applicability in other centers is questionable because the rate of transfusion in the program where
www.anesthesia-analgesia.org
433
Transfusion Requirements During Liver Transplantation
Figure 4. Predicted 1-year survival by donor race and number of units red blood cell (RBC) + cell salvage (CS) transfused. Long dashed line represents donor race other than African American or Caucasian; solid line represents African American or Caucasian donor races. Corresponding dotted/short-dashed lines represent 68% confidence limits, equivalent to ±1 SE. Plot assumes patient international normalized ratio (INR) of 1.5, disease etiology not liver cancer, primary diagnosis not viral hepatitis, and donor age of 44 years.
Figure 3. Death after deceased donor liver transplantation. A, survival. Each square represents a death, and numbers in parentheses indicate patients remaining at risk. Solid lines within dashed 68% confidence limits represent parametric estimates. B, Instantaneous risk of death (hazard function). Solid lines enclosed within dashed confidence limits represent parametric estimates. Note expanded horizontal scale.
the model was developed was extremely low with only 19.5% of the patients transfused at all.4 In contrast, many other centers report significantly higher transfusion rates. Our results demonstrate that a model created to predict transfusion requirements for patients undergoing OLT is unreliable, and hence, the practical utility remains minimal. Although there was substantial overlap in the sets of variables that predicted >20 and >30 units of RBCs among our models, no set of variables provided good prediction for large transfusion requirements. The CART analysis model
was best in identifying patients with intraoperative RBC + CS requirements 20 units of RBC and CS) proved difficult. We did confirm, however, a strong association between intraoperative transfusion and early mortality, as well as quantified the risk of transfusion on posttransplant survival. The main limitation of this study is its retrospective nature and the fact that we could not account for all the factors influencing transfusion requirements, including transfusion triggers. Also, the results of this study reflect our institutional practice and should not be interpreted February 2014 • Volume 118 • Number 2
outside of this context. It is possible, although unlikely, that a similar analysis conducted on patients from other institutions could produce different results, simply based on differences in practice and the population presenting for OLT. In conclusion, prediction of intraoperative blood product requirements based on preoperatively available variables is unreliable in all transfusion range categories analyzed. We were able to quantify a strong measurable association between intraoperative blood transfusion and recipient mortality after OLT. E
APPENDIX INTERPRETATION OF REGRESSION, LOGISTIC REGRESSION, AND SURVIVAL MODEL EFFECTS Multivariable Regression: Predictors of red blood cell (RBC) + cell salvage (CS) usage (Table 5). Recipient Age at Transplant For each year older a patient is, there is a 0.9% increase in the predicted number of units of RBC + CS required. For every increase of 10 years in recipient age, there is a 9% increase in the estimated number of units of RBC + CS required. Recipient Diagnosis: HCC + Malignant Liver Tumor For patients with this diagnosis, the predicted need for blood products is 30% lower than for patients with other diagnoses. Recipient Pretransplant Platelet Count Each increase of 1k/mm3 in pretransplant platelet count results in a 0.3% decrease in the predicted number of units of RBC + CS required. Each increase of 10k/mm3 in pretransplant platelet count results in a 3% decrease in the predicted number of units of RBC + CS required. Recipient Pretransplant INR Each increase of 1 unit of INR results in a 36% increase in predicted number of units of RBC + CS required.
www.anesthesia-analgesia.org
435
Transfusion Requirements During Liver Transplantation
Recipient Pretransplant Total Bilirubin Each increase of 1 unit of total bilirubin results in a 0.9% increase in predicted number of units of RBC + CS required.
Surgeon Y Having transplant Surgeon Y increases odds of needing >30 units of RBC + CS by 164% (53%–871%).
Recipient Pretransplant Creatinine Each increase of 1 unit of creatinine results in a 9% increase in predicted number of units of RBC + CS required.
Surgeon Z Having transplant Surgeon Z increases odds of needing > 30 units of RBC + CS by 286% (66%-986%).
Expanded Criteria Donor Use of expanded criteria donor results in 26% increase in predicted number of units of RBC + CS required.
Survival Analysis Hazard Model (Table 12) (Effect size percentages are followed by 95% confidence intervals on the effect sizes).
Surgeon X When Surgeon X performs the transplant, the predicted number of units of RBC + CS required is 30% lower than when another surgeon performs the transplant.
RBC + CS Units Transfused Each additional unit transfused increases the hazard of death in the early hazard phase by 4.3% (3.0%–5.6%).
Logistic Regression: Predictors of RBC + CS usage > 20 Units (Table 6) (Effect size percentages are followed by 95% confidence intervals on the effect sizes). Recipient Pretransplant Creatinine Each 1 unit increase in creatinine increases odds of needing >20 units of RBC + CS by 34% (17%–53%). Recipient Pretransplant INR Each 1 unit increase in INR increases odds of needing >20 units of RBC + CS by 81% (23%–167%). Recipient Pretransplant Platelet Count Each decrease of 1k/mm3 in pretransplant platelet count results in a 0.6% (0.2%–0.9%) increase in the odds of needing >20 units of RBC + CS. Each decrease of 10k/mm3 in pretransplant platelet count results in a 5.8% (2.0%–8.6%) increase in the odds of needing >20 units of RBC + CS. Recipient Previous Abdominal Surgery History of previous abdominal surgery increases odds of needing >20 units of RBC + CS by 79% (23%–162%). Recipient Blood Type A Blood type A increases odds of needing >20 units of RBC + CS by 68% (15%–143%). Surgeon Y Having transplant Surgeon Y increases odds of needing >20 units of RBC + CS by 112% (21%–272%). Logistic Regression: Predictors of RBC + CS usage > 30 Units (Table 9) (Effect size percentages are followed by 95% confidence intervals on the effect sizes). Recipient MELD Score Each 1 unit increase in the assigned MELD Score increases odds of needing >30 units of RBC + CS by 7% (2%–11%). Recipient Pretransplant Platelet Count Each decrease of 1k/mm3 in pretransplant platelet count results in a 0.7% (0.2%–1.3%) increase in the odds of needing >30 units of RBC + CS. Each decrease of 10k/mm3 in pretransplant platelet count results in a 6.8% (1.9%–12.3%) increase in the odds of needing >30 units of RBC + CS.
436 www.anesthesia-analgesia.org
INR Each increase of 0.1 in INR is associated with a 3.3% (1.4%– 5.3%) increase in hazard of death in the early hazard phase. Donor Age Each increase of 1 year in donor age is associated with a 3.0% (1.2%–4.9%) increase in hazard of death in the early hazard phase. Each increase of 10 years in donor age is associated with a 34% (12%–61%) increase in hazard of death in the early hazard phase. Donor Race Other than African American or Caucasian Patients whose donor is of race other than African American or Caucasian have 4.4 (1.6–11.7) times the hazard of death in the early hazard phase as patients with African American or Caucasian donors. Diagnosis Viral Hepatitis Patients with a primary diagnosis of viral hepatitis have 3.0 (1.4–6.1) times the hazard of death in the late hazard phase as patients with other diagnoses. Disease Etiology Liver Cancer Patients with disease etiology of liver cancer have 4.2 (2.0– 8.8) times the hazard of death in the late hazard phase as patients with other disease etiologies. DISCLOSURES
Name: Jacek B. Cywinski, MD. Contribution: This author helped design and conduct the study, analyze the data, and write the manuscript. Attestation: Jacek B. Cywinski has seen the original study data, reviewed the analysis of the data, approved the final manuscript, and is the author responsible for archiving the study files. Name: Joan M. Alster, MS. Contribution: This author helped analyze the data and write the manuscript. Attestation: Joan M. Alster has seen the original study data, reviewed the analysis of the data, and approved the final manuscript. Name: Charles Miller, MD. Contribution: This author helped write the manuscript. Attestation: Charles Miller has seen the original study data, reviewed the analysis of the data, and approved the final manuscript.
anesthesia & analgesia
Name: David P. Vogt, MD. Contribution: This author helped design and conduct the study. Attestation: David P. Vogt has seen the original study data, reviewed the analysis of the data, and approved the final manuscript. Name: Brian M. Parker, MD. Contribution: This author helped design and conduct the study, analyze the data, and write the manuscript. Attestation: Brian M. Parker has seen the original study data, reviewed the analysis of the data, and approved the final manuscript. This manuscript was handled by: Edward C. Nemergut, MD. ACKNOWLEDGMENTS
The authors wish to thank Lucy Thuita, MS, for data management and programming and Mrs. Tanya Smith for her editorial assistance. REFERENCES 1. Deakin M, Gunson BK, Dunn JA, McMaster P, Tisone G, Warwick J, Buckels JA. Factors influencing blood transfusion during adult liver transplantation. Ann R Coll Surg Engl 1993;75:339–44 2. Findlay JY, Rettke SR. Poor prediction of blood transfusion requirements in adult liver transplantations from preoperative variables. J Clin Anesth 2000;12:319–23 3. Steib A, Freys G, Lehmann C, Meyer C, Mahoudeau G. Intraoperative blood losses and transfusion requirements during adult liver transplantation remain difficult to predict. Can J Anaesth 2001;48:1075–9 4. Massicotte L, Sassine MP, Lenis S, Roy A. Transfusion predictors in liver transplant. Anesth Analg 2004;98:1245–51, table of contents
February 2014 • Volume 118 • Number 2
5. Hendriks HG, van der Meer J, de Wolf JT, Peeters PM, Porte RJ, de Jong K, Lip H, Post WJ, Slooff MJ. Intraoperative blood transfusion requirement is the main determinant of early surgical re-intervention after orthotopic liver transplantation. Transpl Int 2005;17:673–9 6. Schroeder RA, Johnson LB, Plotkin JS, Kuo PC, Klein AS. Total blood transfusion and mortality after orthotopic liver transplantation. Anesthesiology 1999;91:329–30 7. Massicotte L, Sassine MP, Lenis S, Seal RF, Roy A. Survival rate changes with transfusion of blood products during liver transplantation. Can J Anaesth 2005;52:148–55 8. de Boer MT, Molenaar IQ, Hendriks HG, Slooff MJ, Porte RJ. Minimizing blood loss in liver transplantation: progress through research and evolution of techniques. Dig Surg 2005;22:265–75 9. Breiman L. Bagging predictors. Mach Learn 1996;24:123–40 10. Breiman L. Random forests. Mach Learn 2001;45:5–32 11. Blackstone EH, Naftel DC, Turner ME, Jr. The decomposition of time-varying hazard into phases, each incorporating a separate stream of concomitant information. J Am Stat Assoc 1986;81:615–24 12. Hevesi ZG, Lopukhin SY, Mezrich JD, Andrei AC, Lee M. Designated liver transplant anesthesia team reduces blood transfusion, need for mechanical ventilation, and duration of intensive care. Liver Transpl 2009;15:460–5 13. McCluskey SA, Karkouti K, Wijeysundera DN, Kakizawa K, Ghannam M, Hamdy A, Grant D, Levy G. Derivation of a risk index for the prediction of massive blood transfusion in liver transplantation. Liver Transpl 2006;12:1584–93 1 4. Pereboom IT, de Boer MT, Haagsma EB, Hendriks HG, Lisman T, Porte RJ. Platelet transfusion during liver transplantation is associated with increased postoperative mortality due to acute lung injury. Anesth Analg 2009;108:1083–91 15. Ramos E, Dalmau A, Sabate A, Lama C, Llado L, Figueras J, Jaurrieta E. Intraoperative red blood cell transfusion in liver transplantation: influence on patient outcome, prediction of requirements, and measures to reduce them. Liver Transpl 2003;9:1320–7
www.anesthesia-analgesia.org
437
