Journal Logo

Clinical and Translational Research

Independent Validation of a Model Predicting the Need for Packed Red Blood Cell Transfusion at Liver Transplantation

Massicotte, Luc1,4; Capitanio, Umberto2; Beaulieu, Danielle1; Roy, Jean-Denis1; Roy, André3; Karakiewicz, Pierre I.2

Author Information
doi: 10.1097/TP.0b013e3181aed477

Abstract

Orthotopic liver transplantation (OLT) may be associated classically with massive blood loss and equally important transfusion requirements (1). Transfusions of blood products have been associated with increased mortality and morbidity (2–7). Massive transfusions “consume” a lot of human and financial resources. Several approaches have been proposed to standardize the transfusion requirements during OLT and to minimize the need for blood products (3, 8–10). These reports have given conflicting results. During the last 20 years, a great and constant decrease of allogenic blood product transfusion requirements has been noted. Despite this decrease, a wide spectrum of intraoperative blood product transfusions still exists from a median of red blood cell (RBC) units transfused per patient of 2 to 10 (11–14). Recently, we have described a novel approach for minimizing the need for packed RBC (PRBC) transfusions, which relies on minimizing the transfusion of plasma and on phlebotomy, with the intent of lowering the central venous pressure, which decreases the need for PRBC transfusion (5). In that report, 80.5% of our patients for an OLT did not receive any blood products with a mean of 0.4±1.0 RBCs units per patient for 200 consecutive OLTs (median 0). Moreover, 16% of this series received 1 or 2 RBC units only, and 3.5% received more RBC units with some plasma or platelets.

On the basis of those observations, we identified three independent predictors of PRBC transfusion (avoidance of plasma transfusion [yes vs. no], phlebotomy [yes vs. no], and starting hemoglobin [Hb] value [coded as a continuous variable]). Those were previously reported as a logistic function: Pr (y=1)=1/(1+exp [−{αj+β1χ1+…+βkχ1}]) (5, 15). Unfortunately, this format lends itself poorly to use in busy clinical practice. Therefore, in this article, we describe a user-friendly version of this model, which consists of a logistic regression-based nomogram and allows to determine the probability of PRBC transfusion in individual patients. This model was developed on 406 OLTs that included the 200 reported in the original article. Moreover, to test the ability of this nomogram in predicting the need for RBC transfusion, we performed a validation in an independent cohort of 109 OLTs that were not used in the previous analysis. Finally, we provide the results of discrimination and calibration tests of the novel nomogram in those 109 prospectively gathered new OLT cases.

MATERIALS AND METHODS

After approval by the Centre hospitalier de l’Université de Montréal ethics board, 515 consecutive liver transplantations on adults from cadaver donors were studied starting in January 1998.

The initial series of 406 OLTs were used for the model development (206 from a retrospective series [1998–2002]) (9) and 200 from a prospective series (2002–2005) (5). The remaining most recently prospectively recorded 109 OLTs (2005–2007) were used as an independent validation cohort.

The anesthesiology protocol used for all 515 OLTs was described previously in detail (9, 16). In brief, all patients received sufentanil, propofol, rocuronium, and isoflurane before 2002 and desflurane after 2002, and they were monitored in a standard fashion (pulse oxymetry, capnography, arterial line, Swan-Ganz catheter). For the first 206 OLTs of this series, the transfusion protocol followed the American Society of Anesthesiology guidelines: the triggering Hb value for PRBC transfusions was between 60 and 100 g/L, when the baseline international normalized ratio (INR) value was more than 1.5, patients were allowed to receive 10 to 15 mL/kg of plasma, and when starting the platelet count was lower than 50×10 platelets/L (9), patients received 5 to 10 units of platelets. An exception was when some of the anesthesiologists would wait for a significant blood loss before transfusing plasma (9).

In the most recent 309 OLTs, plasma, platelets, and cryoprecipitate were not used to correct preoperative or intraoperative coagulopathy, unless uncontrollable bleeding was encountered. Baseline central venous pressure (CVP) was decreased by one third by intravascular volume restriction, phlebotomy, or both. Phlebotomy consisted of withdrawing blood (from the introducer of the pulmonary artery catheter), at the beginning of the case, without any crystalloid or colloid volume replacement. Criteria for phlebotomy were an Hb value more than 85 g/L, hemodynamic stability, and a normal renal function. The quantity of blood withdrawn varied according to the patient’s mass: 7 to 10 mL/kg. The phlebotomy was interrupted if the blood pressure dropped by more than 20% despite vasopressors or at the onset of metabolic acidosis. A drop of more than 20% of the arterial pressure at induction was considered a contraindication for using phlebotomy. The purpose of the phlebotomy was to reduce the CVP. Avoiding hemodilution led to a preservation of the coagulation factor levels. The CVP was slowly corrected after unclamping the inferior vena cava, to avoid hepatic congestion. Harvested blood was returned to the patient at the end of surgery or earlier if needed (S-T-elevation, tachycardia, hemodynamic instability). Phenylephrine drips were used to correct a drop of more than 20% of the mean arterial pressure from the baseline, and vasopressin was added as needed during the anhepatic phase. The cutoff for PRBC transfusion consisted of Hb value of 60 g/L. PRBC transfusions were given once the bleeding was controlled. Aprotinin was administered in every case (515 cases) according to the Hammersmith protocol: 2 million units as a bolus and 0.5 million units per hour until portal vein anastomosis was completed (17), and plasma has never been used for volume expansion.

Seven hepato-biliary surgeons performed all the OLTs with two staff surgeons involved in each procedure. Neither veno-venous bypass nor piggyback techniques were used. So, for the hepatectomy there was a classical cross-clamping of the inferior vena cava, supra and infrahepatic. During dissection, bleeding was controlled almost exclusively by means of electrocautery and metallic hemostatic clips. All the livers were harvested from brain dead donors and were attributed by ABO-Rh compatibility. Regular steroid induction, followed by maintenance therapy with tacrolimus was used.

Statistical Analyses

Logistic regression analyses addressed the relationship between the three previously identified predictors of the likelihood of PRBC transfusion and the actual rate of PRBC transfusion. The predictors consisted of plasma transfusion, phlebotomy, and starting Hb value (the immediate preoperative value). The latter was coded as a continuous variable (g/L). Conversely, plasma transfusion and phlebotomy were coded in binary fashion (yes/no). Univariable and multivariable models predicting the probability of PRBC transfusion were fitted. The regression coefficients from the multivariable (three variables) logistic regression model that included all three predictors were used to develop the nomogram predicting the individual probability of PRBC transfusion.

Subsequently, we performed an internal validation of the nomogram to test its ability to correctly discriminate between those patients who did and those who did not receive PRBC. The area under the curve (AUC) was used to quantify the nomogram’s discriminant properties. Two hundred bootstrap resamples were used to correct for overfit bias. Subsequently, we tested the model’s calibration. Here, the loess smoother was used to plot the nomogram predicted probabilities of PRBC transfusion according relative to the observed rate of PRBC transfusion in the development cohort.

Finally, we performed an external validation of the nomogram, and the validation was then repeated in the independent set of 109 OLTs, which were not used for the model development. Discrimination was quantified with the AUC, and the loess smoother was used to graphically explore the nomogram’s calibration in the independent sample.

All statistical tests were performed with S-Plus Professional (Statistical Sciences, Seattle, WA). Statistical significance was set at 0.05.

RESULTS

Table 1 shows the descriptive characteristics of the development (n=406) and validation (n=109) datasets of respectively 406 and 109 OLTs. The number of plasma units that were transfused (P=0.001), the rate of phlebotomy (P=0.001), and the number of transfused PRBC units (P=0.001) significantly differed between the development and the validation datasets (Table 1). No difference was recorded in the starting Hb level.

TABLE 1
TABLE 1:
Clinical characteristics of the study population (n=515) that are stratified between the development (n=406) and the external (n=109) validation cohort

Table 2 shows the univariable and multivariable logistic regression models predicting the probability of PRBC transfusion during OLT. In univariable models, transfusion of plasma (odds ratio [OR] 15.0, P<0.001) increased the rate of PRBC transfusion. Conversely, phlebotomy (OR 0.06, P<0.001) and a high starting Hb level (OR 0.95, P<0.001) had the opposite protective effect. In the multivariable model, which included all three variables, all three variables reached independent predictor status (P<0.001).

TABLE 2
TABLE 2:
Univariable and multivariable analyses predicting the probability of packed red blood cell transfusion after orthotopic liver transplantation (n=406)

Figure 1 shows the nomogram predicting the individual probability of PRBC transfusion. Low starting Hb represents the most powerful indicator of the need for transfusion. For example, Hb of 100 contributed to 100 risk points. Conversely, the transfusion of plasma contributed to 30. Finally, absence of phlebotomy contributed to 18 risk points. The joint effect of starting Hb of 100, transfusion of plasma, and absence of intraoperative phlebotomy resulted in 148 risk points, which on the total point axis translated into a 92% probability of PRBC transfusion. Conversely, a starting Hb of 130, absence of plasma transfusion, and use of phlebotomy resulted in 40 risk points and a 2% probability of PRBC transfusion.

FIGURE 1.
FIGURE 1.:
Nomogram predicting the need of packed red blood cell (PRBC) transfusion after orthotopic liver transplantation. Instruction: to obtain nomogram predicted probability PRBC transfusion, locate patient values at each axis. Draw a vertical line to the “Point” axis to determine how many points are attributed for each variable value. Sum the points for all variables. Locate the sum on the “Total Points” line to be able to assess the individual probability of PRBC transfusion after orthotopic liver transplantation on the “P(RBC Transfus.)” line.

The internal validation of the nomogram was performed in the cohort of 406 OLTs that were used for model development. Two hundred bootstrap resamples were used to control for overfit bias that might be associated with the development and testing of a predictive model within the same cohort. The bootstrap-adjusted AUC of the model was 89.8%, where 50% is synonymous with a flip of a coin and 100% represents perfect predictions. The calibration shown in Figure 2A demonstrated virtually perfect agreement between nomogram-predicted probabilities and the observed rate of PRBC, within the development cohort. In that figure, the x-axis represents the nomogram predicted probability of PRBC transfusion. The latter can range from 0.1% to 99.7% (Fig. 1). The y-axis represents the observed rate of PRBC transfusion. Vertical axes represent the frequency distribution of the population and indicate that it was distributed predominantly in the lower predicted risk values. The 45° line represents the ideal relationship between predicted and observed rate of PRBC transfusion. A model that invariably predicts the observed rate would correspond to that line. The broken line corresponds to the model performing where the predictions are at times below the 45° line (overprediction) and at times above it (underprediction). Overall, the departures from ideal prediction are minimal and are distributed evenly between the lowest and the highest predicted probabilities.

FIGURE 2.
FIGURE 2.:
Calibration plots of the predictions of packed red blood cell (PRBC) transfusion in the internal (A) and in the external (B) validation.

The nomogram was then subjected to an independent validation within a cohort of 109 OLTs that were not used for its development. The nomogram’s ability to discriminate between those with and without transfusion resulted in an AUC of 89.8%. The calibration in the external validation is shown in Figure 2B. The same principles apply to the external validation cohort except that the population now represent a fully independent cohort. As expected, the relationship between nomogram predicted probabilities of PRBC transfusion and the observed rate is not as close as in the internal validation. In the low range (0%–40%) of predicted probabilities, the nomogram tends to overpredict the observed rate of PRBC transfusion. Conversely, in the range from 41% to 97.7%, the nomogram tends to underestimate. Greater departures from ideal predictions are expected in an external and independent cohort. Specifically, in the current case, when the nomogram predicted a 20% risk, the actual rate was closer to 10%. Conversely, in patients with higher nomogram- predicted risk of PRBC transfusion, the nomogram tended to underestimate the true rate of PRBC transfusion. For example, when the nomogram-predicted probability was 60% the observed rate was closer to 65%.

DISCUSSION

In this article, our goal was to provide the clinician with a tool that combines the most important, significant, and informative predictors of the likelihood of PRBC transfusion. Moreover, our objective was to devise a tool that quantifies the risk of PRBC transfusion on an individual basis, instead of providing ORs that might be difficult to interpret. The nomogram format represents an excellent graphical display of the effects of several risk factors on the outcome of interest. It also provides the individual probability of the risk of PRBC transfusion, which is expressed on a 0% to 100% scale. Such format is substantially more intuitive to virtually all clinicians than an ORs (18).

The risk axes of the current nomogram clearly depict the effect of each of the three risk variables on the risk of PRBC transfusion. Close assessment of the nomogram axes reveals that the use of plasma transfusion has the same effect as the decrease in starting Hb from 145 to 100 g/L. The avoidance of phlebotomy is associated with the equivalent of a drop of starting Hb from 155 to 145 g/L. The combined effect of no intraoperative phlebotomy and of plasma transfusion is synonymous with a drop of starting Hb from 145 to 70 g/L, which was close to our transfusion threshold of 60 g/L.

Therefore, the nomogram format represents a useful clinical aid, which can demonstrate the importance of phlebotomy and the danger of liberal plasma transfusions. In addition, the nomogram can be used to determine the effect of phlebotomy and of plasma transfusion status on an individual patient’s probability of receiving one or several PRBC transfusion. For example, a patient with starting Hb of 110 g/L, who receives no plasma transfusion and is subjected to phlebotomy, has a 5% chance of needing PRBC transfusions. Conversely, without the use of phlebotomy and with plasma transfusions, the same patient has an 82% chance of needing PRBCs.

We have not only demonstrated that the use of the nomogram is useful in clinical practice but also confirmed its accuracy in internal and independent data samples. The discriminant properties of the nomogram were virtually 90% in both settings, which implies that its predictions are correct in 9 of 10 cases. Nonetheless, the calibration plot expectedly demonstrated greater departures from ideal predictions in the external validation set than in the internal validation cohort.

Our model represents the first multivariable model capable of predicting the individual risk of PRBC transfusion at OLT. Although our results are encouraging, further external validations are needed to corroborate or refute the robustness of our model in other populations. For example, our model may perform equally well in other high-volume tertiary care institutions. It might perform less well in lower volume, secondary care centers, due to differences in surgical expertise and other variables that are surgeon and hospital volume dependent.

Although our model is accurate, it is not perfectly accurate. However, it is possible that even better AUC could be obtained if an interaction term was included. Assessment of interactions within the nomogram would have required a substantially larger sample size. Moreover, our model cannot be used as evidence for causality. Only statistically significant associations have been shown. These limitations are shared with most of the predictive and prognostic models currently available. In an attempt to improve its discrimination and calibration, additional variables could be added. In consequence, our model could be used as a stepping stone toward the development of future models.

Previous studies of variables associated with poor outcomes in OLT have agreed on a few points. Intraoperative transfusion of blood products is recognized as a poor prognostic factor (2–4, 7, 19, 20). This finding may be caused by direct negative effects of blood transfusion such as transfusion reaction, infectious contamination of blood products, immune modulation of the transfused patient, or inflammatory process. Many authors have tried to predict bleeding and blood product transfusions during OLT with conflicting results. Many reasons could explain these differences. These studies have taken place during a 20-year period (1988–2007) (8–14, 16, 19–21).

Surgical techniques have been mastered with time and experience. Whatever the method selected, different results are to be expected. Finally, all series published included no more than a few hundred patients (8–10, 21–25). Could our actual results of prediction be compared with the results of our retrospective series (9) where transfusion of plasma and starting INR value were the variables with the strongest link with transfusion of RBC? The main difference between the two studies was the transfusion of plasma according to American Society of Anesthesiology guidelines (10–15 mg/kg if INR≥1.5) in the retrospective series. We did not find any benefit trying to correct coagulation defects with plasma (9). Furthermore, we found a significant link between transfusion of plasma and decreased survival rate. Therefore, we have stopped transfusing plasma according to the biochemical values. In the prospective series, there were two variables linked to bloodless OLT: phlebotomy and starting Hb value. Phlebotomy is not an end by itself, it is a tool to decrease CVP and blood volume. Phlebotomy and Hb are data linked together. When the starting Hb value is low, the patient will not benefit from phlebotomy. When his starting Hb value is more than or equal to 85 g/L, he has more chance to be phlebotomized and fare through his surgery without any transfusion.

The data of the first 406 OLTs of this series indicate that this novel practice does not result in any deleterious results. There was no cardiac complication (ischemia, infarct, acute pulmonary edema, or sudden death). A patient with symptomatic coronary disease does not qualify for OLT. The disease is investigated with transthoracic echography, electrocardiogram (EKG)-stress test, methyl isobutyl isonitrile (MIBI) dipyridamole-persantin, or coronarography. If a patient has lesions that are not treatable with stent or surgery, or is unsuccessfully manageable with medication, he will not be a candidate for OLT. There was no new neurological deficits or intraoperative death (5, 16). The phlebotomy seemed to have no deleterious effects on the creatinine value and survival rate up to 1 year postoperatively (5).

Our predictions could help the transplant team to identify the patients at higher risk of transfusions. These patients could benefit from preoperative erythropoietin alpha or the use of recombinant factor VIIa (26) or prothrombin complex concentrates (27).

CONCLUSIONS

Our model represents the first multivariable model capable of predicting the individual risk of PRBC transfusion at OLT. Its accuracy is highly promising. Nonetheless, it requires further validation in independent external cohorts.

ACKNOWLEDGMENTS

The authors thank Mrs. Denise Bois for her secretarial work and Mr. Robert Boileau for data processing.

REFERENCES

1. Butler P, Israel L, Nusbacher J, et al. Blood transfusion in liver transplantation. Transfusion 1985; 25: 120.
2. Cacciarelli TV, Keeffe EB, Moore DH, et al. Effect of intraoperative blood transfusion on patient outcome in hepatic transplantation. Arch Surg 1999; 134: 25.
3. Ramos E, Dalmau A, Sabaté A, et al. 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.
4. Massicotte L, Sassine M-P, Lenis S, et al. Survival rate changes with transfusion of blood products during liver transplantation. Can J Anaesth 2005; 52: 148.
5. Massicotte L, Beaulieu D, Thibeault L, et al. Coagulation defects do not predict blood product requirements during liver transplantation. Transplantation 2008; 85: 956.
6. Hendriks HG, van der Merr J, de Wolf JT, et al. Intraoperative blood transfusion requirement is the main determinant of early surgical re-intervention after orthotopic liver transplantation. Transpl Int 2005; 17: 673.
7. Boyd SD, Stenard F, Lee DKK, et al. Alloimmunization to red blood cell antigens affects clinical outcomes in liver transplant patients. Liver Transpl 2007; 13: 1654.
8. Steib A, Freys G, Lehmann C, et al. Intraoperative blood losses and transfusion requirements during adult liver transplantation remain difficult to predict. Can J Anaesth 2001; 48: 1075.
9. Massicotte L, Sassine M-P, Lenis S, et al. Transfusion predictors in liver transplant. Anesth Analg 2004; 98: 1245.
10. Frasco PE, Poterack KA, Hentz JG, et al. A comparison of transfusion requirements between living donation and cadaveric donation liver transplantation: Relationship to model of end-stage liver disease score and baseline coagulation status. Anesth Analg 2005; 101: 30.
11. De Boer MT, Molenar IQ, Hendriks HG, et al. Minimizing blood loss in liver transplantation: Progress through research and evolution of techniques. Dig Surg 2005; 22: 265.
12. Lodge JP, Jonas S, Jones RM, et al. Efficacy and safety of repeated perioperative doses of recombinant factor VIIa in liver transplantation. Liver Transpl 2005; 11: 973.
13. Planinsic RM, van der Meer J, Tests G, et al. Safety and efficacy of a single bolus administration of recombinant factor VIIa in liver transplantation due to chronic liver disease. Liver Transpl 2005; 11: 895.
14. Ozier Y, Pessione F, Samain E, et al; French Study Group on Blood Transfusion in Liver Transplantation. Institutional variability in transfusion practice for liver transplantation. Anesth Analg 2003; 97: 671.
15. Glantz SA, Slinker BK. Primer of applied regression and analyses of variance. New York, McGraw-Hill, Inc. 1990, pp 514.
16. Massicotte L, Lenis S, Thibeault L, et al. Effect of low central venous pressure and phlebotomy on blood product transfusion requirements during liver transplantation. Liver Transpl 2006; 12: 117.
17. Royston D, Bidstrup BP, Taylor KM, et al. Effect of aprotinin on need for blood transfusion after repeat open-heart surgery. Lancet 1987; 2: 1289.
18. Capitanio U, Jeldres C, Shariat SF, et al. Clinicians are most familiar with nomograms and rate their clinical usefulness highest, look-up tables are second best. Eur Urol 2008; 54: 958.
19. Matinlauri IH, Nurminen MM, Hockerstedt KA, et al. Risk factors predicting survival of liver transplantation. Transplant Proc 2005; 37: 155.
20. Lee DS, Gil WH, Lee HH, et al. Factors affecting graft survival after living donor liver transplantation. Transplant Proc 2004; 36: 2255.
21. Massicotte L, Thibeault, Beaulieu D. Low CVP during liver transplantation. J Cardiothorac Vasc Anesth 2008; 22: 315.
22. Brown RS, Kumar KS, Russo MW, et al. Model for end-stage liver disease and Child-Turcotte Pugh score as predictors of pretransplantation disease severity, posttransplantation outcome, and resource utilization in united network for organ sharing status 2A patients. Liver Transpl 2002; 8: 278.
23. Shan WLP, Barkun J, Metrakos P, et al. Blood product use during orthotopic liver transplantation. Can J Anaesth 2004; 51: 1045.
24. Schroeder RA, Collins BH, Tuttle-Newhall E, et al. Intraoperative fluid management during orthotopic liver transplantation. J Cardiothorac Vasc Anesth 2004; 18: 438.
25. Costa MG, Chiarandini P, Della Rocca G. Hemodynamics during liver transplantation. Transplant Proc 2007; 39: 1871.
26. Pugliese F, Ruberto F, Summonti D, et al. Activated recombinant factor VII in orthotopic liver transplantation. Transplant Proc 2007; 39: 1883.
27. Levy JH, Tanaka KA, Dietrich W. Perioperative hemostatic management of patients treated with vitamin k antagonists. Anesthesiology 2008; 109: 918.
Keywords:

Liver transplantation; Transfusion; Prediction

© 2009 Lippincott Williams & Wilkins, Inc.