Risk of Perioperative Transfusion in Elective Hepatectomy
We thank Dr Cockbain and colleagues for their interest and we offer some observations and our thoughts about the concerns they outline in their letter.
The transfusion rates found in our data series are indeed higher than reported by others. One explanation is that the exact definition of perioperative may differ between studies. For example, we evaluated transfusions occurring at any time during the index hospital admission; in contrast, Cockbain et al1 refer to transfusions performed within 48 hour of surgery. It is also important to keep in mind that the transfusion rates that we report pertain to liver resections performed between 1998 and 2003. The rates are steadily decreasing; we recently reported a randomized trial where the transfusion rate for patients receiving standard of care was 25%.2
The letter draws attention to the potential predictive value of a few variables that were not included in our nomogram. Redo surgeries represented only 5% of the liver resections performed in our data series, and transfusion rates were very similar in initial resections (45%) and redo resections (48%). Size of the largest tumor resected was indeed associated with the risk of transfusion on the univariate analysis. However, size was strongly correlated with the number of segments resected and, in contrast with the results observed in Cockbain et al,1 lost its predictive power after adjusting for number of segments. Its addition to the prediction model did not improve the prediction accuracy.
Our decision to treat diagnosis as a binary variable (primary liver malignancy vs liver metastasis or benign disease) may seem as an oversimplification, but it came after careful consideration of the possible options. Patients with hilar cholangiocarcinomas had transfusion rates in excess of 70% whereas other patients with primary malignancies were transfused in proportion of about 50%. The prevalence of hilar cases, however, was less than 4% in our training data set, which prevented us from considering it as a good candidate for the model. To fully address this comment, we rebuilt our predictive model using hilar cholangiocarcinoma as a separate category, but this model had identical concordance (0.709) to the original model in the validation data set. Our conclusion is that it does not add to the predictive strength of the model and its exclusion is justified. In this context, it is worth mentioning that a good prediction model will not always reflect in detail all the associations that are clinically evident. In some cases, a few clinical and pathological factors are correlated and one of them captures the common effect so strongly that the inclusion of the others becomes unnecessary. In other situations, a factor that has a strong effect on the outcome is present only in a small proportion of cases, and its effect can only be estimated with considerable uncertainly; consequently, its inclusion in the prediction model will add little, if anything, to the overall prediction.
We expect the nomogram to retain its predictive ability in centers with similar case-mix (similar distributions of the predictive variables) even when the transfusion rates are lower. It is not possible to speculate on what would happen in centers with substantially different case-mix. We would encourage investigators from these centers to validate our model using the data from their centers to answer this question.
The cost-effectiveness of ANH and PADB interventions is difficult to address authoritatively without a full analysis, but we believe that it will strongly depend on transfusion requirements. We are currently conducting a randomized trial to investigate whether a targeted approach (identifying patients at high risk of transfusion ahead of time) will improve the yield of ANH.