Anesthesia, You Run Fast!

To the Editor We read with great interest your recent publication by Yan et al.1 The article reports an excellent example of machine learning (ML) application for risk prediction. There are several strengths and interesting results that we would like to emphasize. Contrary to expectations, the addition of intraoperative variables has not been shown to improve the performance of the predictive mortality model performance, compared to that of intraoperative data alone. In our opinion, data transformation, especially of the vital parameters measured continuously, is probably one of the possible causes. Dynamic response to treatments, such as the reaction to a hypothetical bolus of fluids, is difficult to interpret only with discrete data. In addition, some of the monitors for minimally invasive monitoring, which may have been used intraoperatively, today are not easy to connect with the monitors routinely present in the operating room. It should also be emphasized the lack of automatic detection of some parameters, such as diuresis, which could instead provide more detailed and accurate information than the sporadic, albeit carefully programmed, detection. In our opinion, the need for an integrated monitor that automatically collects all the parameters is becoming increasingly important. With the current alternative of having mixed, partially automatic data collections associated with manual introduction, instead, not only the quality of the data drops, but it is possible to lose the complexity of the connections between the variables analyzed too. However, we agree with the authors that, despite the results of this interesting study, intraoperative data have to be considered a valuable resource. Menzenbach et al2 have clearly demonstrated that the preoperative parameters are not sufficient to accurately identify patients who could benefit from postoperative intensive care unit (ICU) admission. In a recent paper, in 5 postoperative complications, the best models were the combined ones, ie, those that unified intraoperative data with preoperative ones.3 Similar results were obtained in the case of The MySurgeryRisk PostOp Extension, with a classification improvement between 2.4% and 10.0% for complications and 11.2% for inhospital mortality exploiting intraoperative data.4 However, what remains fundamental, as the authors say, is to understand what is the best way to use these data, on which models, and in which contexts. Valentina Bellini, MDFrancesco Saturno, MDElena Bignami, MDDepartment of Medicine and SurgeryAnesthesiology, Critical Care, and Pain Medicine DivisionUniversity of ParmaParma, Italy[email protected]