The use of preoperative risk score to predict postoperative mortality following liver resection

Abstract

To the editor, With great interest, we read the recent article by Dasari et al developing and validating a preoperative risk score for prediction of 90‐day mortality after liver resection. By using univariable and multivariable analyses to identify the predictors of postoperative mortality, they developed a preoperative risk score, with the areas under the receiver operating characteristic curve (AUROC) of 0.778. Furthermore, this risk score remained significant in a validation cohort of 788 patients (AUROC, 0.703; P < 0.001). Although the findings of this study may provide significant insights into the safety improvement of patients undergoing liver resection, several important issues need to be clarified and discussed. First, an important question was why only age, diabetes mellitus, preoperative sodium levels, and type of surgery were used to develop their preoperative risk score. Besides the above four variables, the results of this study showed that hypertension, ischemic heart disease, preoperative neutrophil/lymphocyte ratio, and preoperative levels of hemoglobin, potassium, calcium, albumin, creatinine, were also significantly associated with an increased risk of postoperative 90‐day mortality. Most important, it was unclear how the scoring criteria of four preoperative variables were determined. Especially, normal preoperative sodium levels of 135 to 140 and 141 to 145 mmol/L were scored as 2.5 and 1, respectively. However, preoperative hypernatremia of sodium levels more than 145 mmol/L was scored as 0. As their risk score was calculated as the sum of scores from four preoperative variables, inappropriate scoring criteria of preoperative variables may have decreased the predictive ability of the risk score. Second, in this study, the associations of intraoperative risk factors with postoperative mortality were assessed and the postoperative complications were observed, but these intraoperative and postoperative factors were not taken into the account when the risk score was established. It must be emphasized that the preoperative risk factors commonly used for prediction model cannot completely explain postoperative outcomes. Other than health status and comorbidities of patients, both surgical burden and postoperative complications are the most important determinants for postoperative short‐term mortality. Just like this study had shown, some intraoperative factors including the extent of liver resection, concurrent intra‐abdominal operation, intraoperative massive blood loss and blood transfusion, prolonged duration of operation, have been significantly associated with an increased risk of short‐term mortality after liver resection. Furthermore, postoperative liver dysfunction and multiple complications, such as postoperative transfusion over five units, unexpected intubation, renal failure, cardiac events, septic shock, and pneumonia, have been identified as the independent risk factors of short‐term mortality after liver resection. Thus, we argue that the risk score established by authors based on some preoperative factors in this study may only be used as a simple preoperative screening tool for short‐term mortality after liver resection, but it is not an effective risk prediction tool of short‐term mortality. Finally, in this study, the ROC curve analysis was performed to assess the predictive accuracy of the risk score in both the derivation and validation cohorts. To obtain an overall assessment on the predictive accuracy of a risk score, only providing the AUROCs in both the derivation and validation cohorts is insufficient. By the ROC curve analysis, the authors should further provide the sensitivity, specificity, positive, and negative predictive values of their risk score for postoperative mortality in the derivation and validation cohorts, as performed in the previous study. By providing the predicted probabilities and observed frequencies for postoperative mortality based on the risk score, the readers can estimate whether there is a good overall agreement between predicted probabilities and observed frequencies in the derivation and the validation cohorts.