A reduction in Type and Screen: preoperative prediction of RBC transfusions in surgery procedures with intermediate transfusion risks

W. A. van Klei1,2, K. G. M. Moons1,2, A. T. Rheineck Leyssius3, J. T. A. Knape1, C. L. G. Rutten4 and D. E. Grobbee2

1Department of Peri-operative Care, Anesthesia and Pain Management, 2Julius Centre for General Practice and Patient Oriented Research, University Medical Centre Utrecht, PO Box 85500, 3508 GA Utrecht, The Netherlands. 3Department of Anesthesiology and Intensive Care Medicine, Twenteborg Hospital, Zilvermeeuw 1, 7609 PP Almelo, The Netherlands. 4Department of Anesthesiology, Isala Clinics, Weezenlanden Hospital, PO Box 10500, 8000 GM Zwolle, The Netherlands*Corresponding author

{dagger}An abstract about this study was accepted for presentation at the 2000 annual meeting of the American Society of Anesthesiologists (Presentation at San Franscisco, California, USA, October 18, 2000).

Accepted for publication: March 3, 2001


    Abstract
 Top
 Abstract
 Introduction
 Methods
 Results
 Discussion
 References
 
In many patients, a ‘type and screen’ procedure is routinely performed before surgery. However, most patients are not transfused after all. Can we predict, which surgical patients will and will not be transfused, to reduce the number of these investigations? We studied 1482 consecutive surgical patients with intermediate risk for transfusion. Multivariate logistic regression modelling and the area under the Receiver Operating Characteristic curve (ROC area) were used to quantify how well age, gender, surgical procedure, emergency or elective surgery and anaesthetic technique predicted transfusion, and whether the preoperative haemoglobin concentration had added predictive value. Gender, age >=70 yr, and type of surgery were independent predictors of transfusion, with a ROC area of 0.75 (95% CI: 0.72–0.79). Validating this model with an easily used prediction rule in a second patient population yielded a ROC area of 0.70 (95% CI: 0.63–0.77). With this rule type and screen could correctly be withheld in 35% of these patients. In the remaining 65% of the patients, a further reduction in type and screen investigations of 15% could be achieved using the preoperative haemoglobin concentration. Using a simple prediction rule, preoperative type and screen investigations in patients who have to undergo surgery procedures with intermediate transfusion risk can be avoided in about 50%. This may reduce patient burden and hospital costs (on average: 3 million US$ per 100 000 procedures).

Br J Anaesth 2001; 87: 250–7

Keywords: blood, transfusion; model, transfusion prediction; anaesthesia, audit


    Introduction
 Top
 Abstract
 Introduction
 Methods
 Results
 Discussion
 References
 
Transfusion of blood (red blood cells or RBC) is sometimes necessary in patients having surgery. A ‘type and screen’ is done preoperatively to prevent complications from incompatibility between donor and recipient or the existence of irregular antibodies. This procedure is much cheaper than full cross matching, and gives the same immuno-haematological safety.1–4 Generally, physicians type and screen patients who might need a perioperative transfusion (commonly based on past experience with the surgical procedure as single predictor). However, most patients who are typed and screened before surgery will not require a transfusion, which means unnecessary patient burden and costs. It would be efficient to further classify patients according to their risk of transfusion using objective and easy obtainable information. Various prediction rules have been developed (especially in orthopaedic surgery), but a laboratory value (preoperative haemoglobin concentration or haematocrit) was always included.5 However, it would be even more efficient if the same predictive accuracy could be obtained without the need for laboratory tests.69

We developed and validated a rule based on patient and surgery characteristics, to predict blood transfusion in patients undergoing surgery with intermediate transfusion risk (1–30%). Subsequently, we evaluated how knowing the preoperative haemoglobin concentration could increase the predictive accuracy of this prediction rule.


    Methods
 Top
 Abstract
 Introduction
 Methods
 Results
 Discussion
 References
 
Patients
We studied 1482 patients (aged 18–98 yr) with intermediate transfusion risk (‘type and screen patients’), undergoing surgery in the Twenteborg Hospital in The Netherlands, in 1998. This hospital is a 638-bed non-university hospital in which neurosurgery and cardiac surgery are not performed. The classification of type and screen patients was based on the current transfusion guide. This divides patients into three surgical groups according to expert opinion. Group A patients have low expected risk for transfusion (0–1%; e.g. arthroscopy or ear surgery), group B patients have intermediate risk for transfusion (1% to approximately 30%; e.g. cholescystectomy or hysterectomy) and group C are high risk patients (more than 30%; e.g. aortic surgery). In group A patients, type and screen is never done (78% of all patients). Patients belonging to group B are always typed and screened, but blood is not stored (16%). Group C patients are always typed and screened and blood is stored (6%). Of all patients in group A, nearly 2% received transfusions. In group B and C the incidence of transfusion was 19% and 43%, respectively. This study evaluates only group B patients (‘type and screen patients’). None of the 1482 patients donated autologous blood preoperatively.

Outcome
The outcome was defined as any allogeneic RBC transfusion (defined as transfusion of one or more units packed cells) on the day of surgery or the first postoperative day. The transfusion decision was made by individual clinicians (anaesthesiologists and surgeons). A rigid protocol was not in use, but in general blood was given when the haemoglobin level was below 10 g dl–1 (6.0 mmol litre–1).

Potential predictor variables
Age, gender, surgery procedures, whether it was an emergency operation (yes/no), the anaesthetic technique and the preoperative haemoglobin level were evaluated as potential predictors. As 39 different surgical procedures were used, they were allocated into 5 categories based on actual risk (occurrence) of transfusion: Group 1 contained only laparoscopic cholecystectomy (transfusion incidence <5%); Group 2 mastectomy and transurethral resection of tumour (TURT) or prostate (TURP) (transfusion incidence 5–9%); Group 3 open cholescystectomy, vaginal hysterectomy, Caesarean section, surgery for urine incontinence or vaginal prolapse (10–19%); Group 4 non-cardiac thoracic surgery (e.g. lobectomy), vascular (arterial) surgery (e.g. femoro-popliteal bypass), prostate enucleation and endometrial cancer surgery (20–29%); Group 5 abdominal and vaginal hysterectomy, hip fracture surgery, revision knee prosthesis, leg amputation, gastro-enterostomy, colon-resection and radical abdominal hysterectomy (30% or more). Anaesthetic technique was defined as a dichotomous variable: a single form of anaesthesia (general, regional or local) compared with a combination anaesthesia (general anaesthesia combined with epidural analgesia). Although in principle a potential predictor, we decided not to include the identity of the surgeon and anaesthesiologist in the model, as they are hard to extrapolate to other hospitals and the aim was to derive an easy and widely applicable prediction rule.

Data collection
The hospital ethics committee approved the study. All data were collected retrospectively from the hospital information system. There were no missing data on any of the predictor or outcome variables, except for the haemoglobin concentration. In 152 patients (10%) it was not measured preoperatively.

Analysis
In the present study, two data sets were randomly selected from all 1482 patients: a derivation set of approximately 75% (1151 patients) and a validation set of approximately 25% (331 patients). SPSS release 9.0 for Windows was used in the analysis (Windows NT 4.0, DELL computer). In the derivation set the association between each predictor and transfusion outcome was quantified using univariable logistic regression modelling. This type of analysis is an alternative to using chi-square tests and gives similar results. In the analysis, surgery was included as four indicator variables (groups 2 to 5) with group 1 as the reference. As the incidence of transfusion in patients aged 18 to 69 was between 10% and 20% in each decade, whereas in patients aged over 70 the incidence increased more rapidly, age was included in the model after dichotomization at 70. After univariate analyses, multivariable logistic regression modelling was applied in order to obtain a prediction model including the independent predictors of transfusion outcome only. This was done by a two-step approach. As age, gender, type of surgery (again included as four indicators), elective surgery and anaesthetic procedure are much easier to obtain, we first evaluated whether these had independent value in the prediction of perioperative transfusion. In this, the interaction between type of surgery and anaesthetic technique was evaluated as well, since both are closely related (regional anaesthesia may reduce blood loss). Subsequently, the added predictive value of the preoperative haemoglobin concentration was evaluated. The full model was reduced by manually (i.e. not automatically) deleting non-significant variables. Predictors with odds ratios that differed significantly from one, defined as odds ratio with P-value <0.10 using log likelihood ratio testing, were considered as independent predictors and retained in the final model. This is commonly done in prognostic research.10

To obtain an easily applicable prediction or scoring rule, the regression coefficients (=ln(OR)) of the predictors in the final model were divided by the smallest coefficient and rounded to the nearest integer. For each subject a score was estimated by assigning points for each variable present and adding the results. The reliability of our prediction rule (goodness of fit) was quantified by using the Hosmer and Lemeshow test. This test is used to compare observed probabilities with predicted probabilities. A high P-value of this test (>0.20) indicates that there is no difference between both probabilities, which means good fit of a model.11 The ability of the model to discriminate between patients with and without transfusion was quantified by using the area under the Receiver Operating Characteristic curve (ROC area).1012 The ROC area can range from 0.5 (no more predictive than a coin flip) to 1.0 (perfect discrimination). A value over 0.7 can be interpreted as reasonable or fair, and over 0.8 as good.13 Differences in ROC area were used to quantify the difference in discriminative ability between full and reduced models taking into account the correlation between the models as they were based on the same cases.14

The performance of the rule was tested in the validation set and the resulting ROC area was compared with the derivation set. A ROC area reflects the overall added value of a model and does not directly indicate its clinical value.15 16 Therefore, in the validation set, we estimated the absolute number of correctly predicted transfused and not transfused patients for the various risk scores of the rule.


    Results
 Top
 Abstract
 Introduction
 Methods
 Results
 Discussion
 References
 
Table 1 shows the comparison of patient characteristics of the derivation and validation set. There were no major differences between the two sets. The transfusion rates for the derivation and validation set were 18.1% (n=208) and 20.8% (n=69), respectively.


View this table:
[in this window]
[in a new window]
 
Table 1 Patient characteristics of derivation and validation set. Values are numbers and percentages between parentheses. # Values are mean and standard deviation between parentheses. * Procedures are listed in the text
 
In the univariate analysis (Table 2) all variables were significantly associated with transfusion. The odds ratios of the four indicators for surgery (groups 2 to 5) indicate the relative risk of transfusion for that group, compared to the reference group 1 (e.g. procedures within group 3 have a 4.1 times higher risk of transfusion than those of group 1). After entering age, gender, surgery procedure and emergency surgery into a multivariate logistic model, all were independently associated with transfusion (Table 3), except emergency surgery (OR 1.26; 95% CI: 0.84–1.88). The ROC area of this first model was 0.75 (95% CI: 0.71–0.79). As further exclusion of variables from this model significantly reduced the ROC area, the model with dichotomized age, gender and surgery procedure was defined as the final prediction model. Addition of anaesthetic technique (including the interaction terms with surgery procedure) to this model showed no added value in the prediction of transfusion: the ROC area remained 0.75. (For the estimation of the added value of the preoperative haemoglobin concentration see below.) Its ROC area in the validation set was 0.71 (95% CI: 0.64–0.78). The model’s estimated risks of transfusion were comparable to the observed risks, which indicated a good model fit (the P-value of the Hosmer and Lemeshow test was 0.98).


View this table:
[in this window]
[in a new window]
 
Table 2 Association of each variable with the incidence of transfusion. # Reference category. * Surgery procedures were included as four indicator variables with group 1 as the reference category. mean. {dagger} OR per g dl–1 increase in haemoglobin concentration. OR=odds ratio; 95% CI=95% confidence interval; LLR=log likelihood ratio test
 

View this table:
[in this window]
[in a new window]
 
Table 3 Association of each variable with the incidence of transfusion in the multivariate logistic model. * Procedures per group are listed in the text; group 1 is the reference group. # The score of each predictor was obtained by dividing the corresponding regression coefficient by the smallest coefficient (0.524) and rounded to the nearest integer. 95% CI=95% confidence interval; OR=odds ratio
 
This final model was transformed into an easily used scoring rule by dividing each regression coefficient by the smallest coefficient (0.524) and rounded to the nearest integer (last column of Table 3): 1*gender + 1*age >=70 + (1, 2, 4 or 5)*surgery procedure. Being a woman counts for 1 point, age >=70 for 1 point, and surgery procedure for 1, 2, 4 or 5 points (for groups 2, 3, 4 and 5, respectively). Such a scoring rule can be considered as one overall predictor test, including several predictor variables. The score can be considered as its (test) result and can be estimated for each patient by assigning the points for each predictor present and adding these points. For instance, a 72-year-old man who will undergo a colon-resection, receives a score of 6 (0 + 1 + 5). In both datasets the score ranged from 0 to 7. The ROC area of the transformed prediction rule was 0.75 (95% CI: 0.71–0.78) and 0.70 (95% CI: 0.63–0.77) in the derivation and validation set, respectively (Fig. 1).



View larger version (16K):
[in this window]
[in a new window]
 
Fig 1 ROC curve of the transformed prediction rule (Table 3). Each bullet indicates a score threshold from 0 (upper-right) to 7 (bottom-left). For example, a score threshold of >2 gives a sensitivity of 0.84 and a specificity of 0.40 (1–0.60).

 
This prediction rule can be used preoperatively to distinguish patients who will and will not be transfused and therefore should and should not be typed and screened. Table 4 shows the actual number of transfused and not transfused patients across score categories (and across corresponding risk of transfusion as estimated by the untransformed model, i.e. second column of Table 3), after the rule was applied to the validation set.


View this table:
[in this window]
[in a new window]
 
Table 4 Distribution of transfused and not transfused patients in the validation set, according to the score of the rule (and to the corresponding risk of transfusion). Values are presented as absolute numbers and as percentages of the ‘Total’ column between parentheses. # Categories of the score estimated from the (transformed) scoring rule (Table 3). *Risk or probability of transfusion estimated by the untransformed prediction model (Table 3) that correspond to the score from the first row: Risk=1/(1 + exp –(–3.701 + 0.629*gender +0.546*age >=70 + 0.524*group2 + 1.291*group3 + 2.287*group4 + 2.386*group5)). n=number of subjects per score (risk) category.
 
From Table 4 one can directly obtain the predictive value for transfusion per score category (reading the table vertically). For example, of all 115 patients with score <=2 (or risk of transfusion <=10%), 104 patients were indeed not transfused, yielding a negative predictive value of 90%. In the group of patients with score >=5, 39 of the 111 were indeed transfused; a positive predictive value of 35%. Table 4 also enables to estimate the sensitivity and specificity at different score thresholds (reading the table horizontally). For example, introducing a threshold at 2, a score <=2 will be considered as test negative and a score >2 will be considered as test positive. This means that, according to the rule, a test negative patient will not be transfused and does not need to be typed and screened, whether a test positive patient will be transfused and needs to be typed and screened. Using this threshold of <=2 (or transfusion risk <=10%), the specificity was 40% (104/262) with 60% unnecessary type and screen procedures, whereas the sensitivity was 84% (19+39/69) with 11 (16%) missed transfused patients.

This 16% of patients who needed transfusion and who would not have been tested, was only 2% less than using a model with type of surgery as a single predictor. Because age and gender are easily obtainable, we decided to leave them in the model. The sensitivity and specificity of all possible score thresholds can be obtained from the ROC curve in Figure 1.

We wished to reduce the number of unnecessary type and screen procedures (i.e. to obtain a high specificity). We tested whether the preoperative haemoglobin concentration (preopHb), when added to the former prediction model (Table 3), contributed useful information. Adding preopHb to the previous model of Table 3, the ROC area increased from 0.71 to 0.80 (95% CI: 0.74–0.86) in the validation set. In absolute numbers the percentage of missed transfused patients decreased from 16% to 12%. We reasoned that it would therefore be inefficient to include preopHb in the initial prediction model, as it led to only a small decrease (4%) in missed transfusions at the expense of a haemoglobin measurement in all patients. Nevertheless, using the preopHb additionally after the application of the rule, i.e. only measuring the preoperative haemoglobin level among those patients with score >2, a further reduction in the number of unnecessary type and screen procedures was achievable. Of the 216 patients with score >2 (Table 4), 31 were excluded due to missing values on preopHb, leaving 185. Although we evaluated preopHb as a continuous as well as a dichotomous predictor variable, we decided to use the dichotomised form (at 14 g dl–1) to enhance applicability, as there was no difference in predictive accuracy. Table 5 shows the results and can be read in the same way as Table 4. By withholding type and screen procedures in all patients with a preopHb level >=14.0 g dl–1, a further reduction in type and screen investigations of 24% could be achieved at the expense of another five missed transfused patients. Other haemoglobin thresholds yielded worse results.


View this table:
[in this window]
[in a new window]
 
Table 5 Distribution of transfused and not transfused patients according to the preoperative haemoglobin concentration in the patients from Table 4 with score >2. Values are presented as absolute numbers and as percentages of the ‘Total’ column between parentheses. * Preoperative haemoglobin concentration in g dl–1. n=number of subjects per haemoglobin category
 
Figure 2 summarizes the results. Using the scoring rule with a threshold of <=2 and subsequently the preopHb level at a threshold of >=14.0 g dl–1, type and screen would be withheld in about 50% of all patients undergoing surgery procedures with intermediate transfusion risk (35% plus 24% of 65%), with 16 (23%) missed transfusions. We investigated the characteristics of the missed transfused patients in the total population of 1482 patients (derivation and validation set together). The prediction rule and subsequent use of the preoperative haemoglobin concentration would miss 55 (20%) transfused patients (Table 6). On average, they required 2.5 units RBC per subject (95% CI: 2.1–2.9) and 82% of them (45 subjects) required no more than 2 units. Two patients required 8 and 10 units. They were emergency patients who were re-operated due to postoperative haemorrhage.



View larger version (11K):
[in this window]
[in a new window]
 
Fig 2 Flow chart of the results of this study after using the scoring rule at a threshold of score >2 and a preoperative haemoglobin concentration of 14.0 g dl–1. #Preoperative haemoglobin concentration. *LAB: patients have to go to the laboratory, and two blood samples have to be taken; one to measure the preoperative Hb and one to investigate eventually the blood group.

 

View this table:
[in this window]
[in a new window]
 
Table 6 Surgery and transfusion characteristics of missed transfused patients (n=55) after application of the scoring rule and the preoperative haemoglobin concentration (derivation and validation set combined).Between parentheses the individual number of units RBC of the patients who required >2 units. #TUR=transurethral resection of prostate or tumour. *PreopHb= preoperative haemoglobin concentration
 

    Discussion
 Top
 Abstract
 Introduction
 Methods
 Results
 Discussion
 References
 
To reduce the number of unnecessary preoperative type and screen procedures, we defined an easily applicable scoring rule containing three simple variables (gender, age, and surgery procedure) to predict transfusion in patients undergoing surgical procedures with intermediate transfusion risk.

Some comments are necessary. First, the prediction rules were based on data from one particular hospital. It is known that there are large differences in blood use between hospitals.1723 Although we have tried to show the robustness of the rule by testing it in a second dataset of our hospital, further research has to be done to validate the rule in other hospitals.

Second, the transfusion trigger was a haemoglobin level of 10 g dl–1 as was recommended formerly.2426 Currently, fewer patients are transfused as RBC transfusion is now based on the patient’s risk for developing inadequate tissue oxygenation, which decreases the transfusion trigger to a haemoglobin level between 6 and 10 g dl–1.2735 This means that when the proposed prediction rule (including the subsequent preoperative haemoglobin measurement) is applied in current practice, the yearly number of (missed) transfusions will be lower. Rehm et al., for example, found a 26% decline in the number of RBC transfusions when the modern recommendations for transfusions were used.4 This can be inferred from Table 6: the large majority of patients received only 2 units packed cells or less, and would likely not now be transfused. However, although the number of wrongly predicted transfusions will likely be lower, the question about the acceptability of the remaining missed transfusions still exists. As can be seen in Table 6, an emergency transfusion seemed in general not probable (except for two patients who needed re-operation). If a patient is not typed and screened and massive haemorrhage occurs, colloids must be administered and the patient typed and screened. If, however, the patients’ blood group is not available on time, 0 blood can always be administered even though the presence or absence of irregular erythrocyte antibodies is not yet known.36 Furthermore, the risk of adverse reactions would be low anyway, given the low prevalence of these antibodies in the general population (2.5%).3740 We estimated that in only 0.1% of all transfusions for procedures with intermediate transfusion risk irregular antibodies would be a problem. (The 10 patients in Table 6 who required more than 2 units packed cells count for 3.6% of all transfusions; 2.5%*3.6% = 0.1%.)

Finally, ASA-classification and body mass index are predictors of transfusion.57 9 41 Unfortunately, these variables were not available for most of our patients. This limits the results of the study, although body mass index alone is an objective parameter, but the ASA-classification is proven to be subjective.42 43

The results of this study support previous work. All predictors for transfusion in surgery found in this study (gender, age over 70, surgery procedure and preoperative haemoglobin concentration) were also found by others.59 20 41 4446 However, direct comparison of our rule with other prediction models is difficult as most studies evaluated a particular type of surgery. One study of different types of surgery showed that complexity of surgery procedure, age and preoperative haemoglobin concentration significantly determined the need for perioperative RBC transfusion.6 Most studies included at least one laboratory value in their initial prediction model (preoperative haemoglobin concentration or haematocrit).59 41 Only one study described a rule without a laboratory value, which included similar predictors (e.g. gender) as we found.47 This study only assessed total hip replacement and the outcome was blood loss rather than transfusion. We are the first to construct a prediction rule for surgical transfusion in procedures with intermediate transfusion risk without a laboratory value.

We suggest using our prediction rule at a threshold score of <2 or an estimated risk <10%, as was done by Weber in his model for preoperative prediction of transfusion in cancer surgery.48 Sensitivity and specificity corresponding to this threshold (84% and 40%, respectively) are obtained in a prognostic setting and should not be confused with (usually much higher) estimates obtained in a diagnostic setting. Subsequently, we used the preoperative haemoglobin level at a threshold of >=14.0 g dl–1. These threshold choices are arbitrary. One could use other thresholds in the scoring rule of Table 4 and the haemoglobin level, though leading to other percentages of misclassifications (Fig. 1).

We assumed that the average direct costs of type and screen are about US$ 80.4850 Using the prediction rule, 35% of all type and screen investigations in intermediate risk surgery procedures could be avoided, giving a reduction in costs of 3 million dollars per 100 000 of these procedures (35 000*$80). When the preoperative haemoglobin concentration is used additionally, a further reduction in costs seems achievable, although the measurement costs of the haemoglobin concentration have to be taken into account. Further cost-effectiveness analyses, including the ‘costs’ of the missed transfusions, should be done and is topic for further research.

In conclusion, we believe that prediction of blood transfusion in patients having surgery with intermediate transfusion risks is feasible using the rule we have developed together with the preoperative haemoglobin concentration. Using these predictors, the number of preoperative type and screen investigations will be reduced by about 50% leading to a considerable reduction in costs.


    References
 Top
 Abstract
 Introduction
 Methods
 Results
 Discussion
 References
 
1 Atrah HI, Galea G, Urbaniak SJ. The sustained impact of a group and screen and maximum surgical blood ordering schedule policy on the transfusion practice in gynaecology and obstetrics. Clin Lab Haem 1995; 17: 177–81[ISI][Medline]

2 Riedler GF. "Type and screen": die immunhamatologische Sicherheit ist gewahrleistet und die Kosten werden gesenkt. [English summary] Schweiz Med Wochenschr 1996; 126: 1946–51[ISI][Medline]

3 Ng SP. Blood transfusion requirements for abdominal hysterectomy: 3-year experience in a district hospital (1993–1995). Aust N Z J Obstet Gynaecol 1997; 37: 452–7[ISI][Medline]

4 Rehm JP, Otto PS, West WW, et al. Hospital-wide educational program decreases red blood cell transfusions. J Surg Res 1998; 75: 183–6[ISI][Medline]

5 Nuttall GA, Santrach PJ, Oliver-WC J, et al. The predictors of red cell transfusions in total hip arthroplasties. Transfusion 1996; 36: 144–9[ISI][Medline]

6 Ong AH, Sim KM, Boey SK. Preoperative prediction of intra and postoperative red blood cell transfusion in surgical patients. Ann Acad Med Singapore 1997; 26: 430–4[Medline]

7 Keating EM, Meding JB, Faris PM, Ritter MA. Predictors of transfusion risk in elective knee surgery. Clin Orthop 1998; 50–9

8 Larocque BJ, Gilbert K, Brien WF. Prospective validation of a point score system for predicting blood transfusion following hip or knee replacement. Transfusion 1998; 38: 932–7[ISI][Medline]

9 Bierbaum BE, Callaghan JJ, Galante JO, Rubash HE, Tooms RE, Welch RB. An analysis of blood management in patients having a total hip or knee arthroplasty. J Bone Joint Surg Am 1999; 81: 2–10[Abstract/Free Full Text]

10 Harrel FE, Lee KL, Mark DB. Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat Med 1996; 15: 361–87[ISI][Medline]

11 Hosmer DW, Lemeshow S. Applied Logistic Regression. New York: J.Wiley and sons, 1989; 140–5

12 Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic curve. Radiology 1982; 143: 29–36[Abstract]

13 Weinstein MC, Fineberg HV. Clinical Decision Analysis. Philadelphia: WB Saunders, 1980

14 Hanley JA, McNeil BJ. A method of comparing the areas under receiver operating characteristic curves derived from the same cases. Radiology 1983; 148: 839–43[Abstract]

15 Moons KGM, Stijnen T, Michel BC, et al. Application of treatment thresholds to Diagnostic-test Evaluation: An alternative to the comparison of areas under the Receiver Operating Characteristic Curves. Med Decis Making 1997; 17: 447–54[ISI][Medline]

16 Moons KGM, van Es G, Michel BC, Buller HR, Habbema JDF, Grobbee DE. Redundancy of Single Diagnostic Test Evaluation. Epidemiology 1999; 10: 276–81[ISI][Medline]

17 Carson JL, Spence RK, Poses RM, Bonavita G. Severity of anaemia and operative mortality and morbidity. Lancet 1988; 340: 727–9

18 The Sanguis study group. Use of blood products for elective surgery in 43 European hospitals. Trans Med 1994; 4: 251–68[ISI][Medline]

19 Carson JL, Duff A., Poses RM, et al. Effect of anaemia and cardiovascular disease on surgical mortality and morbidity. Lancet 1996; 348: 1055–60[ISI][Medline]

20 Baele P, Beguin C, Waterloos H, et al. The Belgium BIOMED Study about transfusion for surgery. Acta Anaesthesiol Belg 1998; 49: 243–303[Medline]

21 Capraro L. Transfusion practices in primary total joint replacements in Finland. Vox Sang 1998; 75: 1–6[ISI][Medline]

22 Surgenor DM, Churchill WH, Wallace EL, et al. The specific hospital significantly affects red cell and component transfusion practice in coronary artery bypass graft surgery: a study of five hospitals. Transfusion 1998; 38: 122–33[ISI][Medline]

23 Carson JL, Duff A, Berlin JA, et al. Perioperative blood transfusion and postoperative mortality. JAMA 1998; 279: 199–205[Abstract/Free Full Text]

24 Adams RC, Lundy JS. Anesthesia in cases of poor surgical risk. Some suggestions for decreasing the risk. Surg Gynecol Obstet 1942; 74: 1011–19

25 Kowalyshyn TJ, Prager D, Young J. A review of the present status of preoperative hemoglobin requirements. Anesth Analg 1972; 51: 75

26 Finch CA, Lenfant C. Oxygen transport in man. N Engl J Med 1972; 286: 407–15[ISI][Medline]

27 National Institute of Health (Bernstein M). Perioperative red blood cell transfusion. JAMA 1988; 250: 2700–3

28 Stehling L, Simon TL. The red blood cell transfusion trigger. Arch Pathol Lab Med 1994; 118: 429–34[ISI][Medline]

29 Task force on blood component therapy. Practice guidelines for blood component therapy. A report by the American Society of Anesthesiologists task force on blood component therapy. Anesthesiology 1996; 84: 732–47[ISI][Medline]

30 Spence RK. Emerging trends in surgical blood transfusion. Semin Hematol 1997; 34: 48–53[ISI][Medline]

31 Loo S, Low TC. Perioperative transfusion strategies: a national survey among anaesthetists. Ann Acad Med Singapore 1997; 26: 193–9[Medline]

32 Spence RK. Anemia in the patient undergoing surgery and the transfusion decision. A review. Clin Orthop 1998; 19–29

33 Weiskopf RB. Do we know when to transfuse red cells to treat acute anemia? Transfusion 1998; 38: 517–21[ISI][Medline]

34 Carson JL, Terrin ML, Barton FB, et al. A pilot randomized trial comparing symptomatic vs. hemoglobin-level-driven red blood cell transfusions following hip fracture. Transfusion 1998; 38: 522–9[ISI][Medline]

35 McFarland JG. Perioperative blood transfusions: indications and options. Chest 1999; 115: 113S–121S

36 Fiebig E. Safety of blood supply. Clin Orthop Res 1998; 357: 6–18

37 Wittkopf D, Grundmann A, Sibrowski W, von Eisenhart RB, Kuhnl P. Analyse irregularer Antikorper in der Abteilung fur Transfusionsmedizin am Universitats-Krankenhaus Hamburg-Eppendorf von 1984–1988. [English summary] Infusionstherapie 1990; 17: 280–2[ISI][Medline]

38 Burin-des RN, Nasr O. Recherche des anticorps irreguliers erythrocytaires par la methode du gel-test. Analyse de 35,882 echantillons. [English summary] Rev Fr Transfus Hemobiol 1993; 36: 391–9[ISI][Medline]

39 Trophilme C, Becker AM, Jaffard HP, Jaulin P. Recherche d’anticorps irreguliers. Etude comparative sur 2 series de 20,000 sujets, l’une avec identification en technique tube, l’autre en gel-test. [English summary] Transfus Clin Biol 1995; 2: 189–99[ISI][Medline]

40 Heddle NM, Soutar RL, O’Hoski PL, et al. A prospective study to determine the frequency and clinical significance of alloimmunization post-transfusion. Br J Haematol 1995; 91: 1000–5[ISI][Medline]

41 Magovern JA, Sakert T, Benckart DH, et al. A model for predicting transfusion after coronary artery bypass grafting. Ann Thorac Surg 1996; 61: 27–32[Abstract/Free Full Text]

42 Owens WD, Felts JA, Spitznagel EL. ASA Physical status classifications: a study of consistency for rating. Anesthesiology 1978; 49: 239–43[ISI][Medline]

43 Wilson ME, Williams NB, Baskett PJF, et al. Assessment of fittness for surgical procedures and the variability of anaesthesists’ judgement. Br Med J 1980; 280: 509[ISI][Medline]

44 Fawcus S, McIntyre J, Jewkes RK, et al. Management of incomplete abortions at South African public hospitals. National Incomplete Abortion Study Reference Group. S Afr Med J 1997; 87: 438–42

45 Warner DO, Warner MA, Schroeder DR, Offord KP, Maxson P, Santrach P. Changing transfusion practices in hip and knee arthroplasty. Transfusion 1998; 38: 738–44[ISI][Medline]

46 Sculco TP, Gallina J. Blood management experience: relationship between autologous blood donation and transfusion in orthopedic surgery. Orthopedics 1999; 22: s129–34[ISI][Medline]

47 Grosflam JM, Wright EA, Cleary PD, Katz JN. Predictors of blood loss during total hip replacement surgery. Arthritis Care Res 1995; 8: 167–73[Medline]

48 Weber RS. A model for predicting transfusion requirements in head and neck surgery. Laryngoscope 1995; 105: 1–17[Medline]

49 Ransom SB, McNeeley SG, Malone-JM J. A cost-effectiveness evaluation of preoperative type-and-screen testing for vaginal hysterectomy. Am J Obstet Gynecol 1996; 175: 1201–3[ISI][Medline]

50 Ziekenfondsraad. Diagnostisch kompas 1999/2000. Ziekenfondsraad, The Netherlands, 1999. [Dutch]