Selection Bias in the Hypotension Prediction Index: Reply.

Abstract

We have read the interesting comment by Giustiniano and Nisi1 about our article.2 The authors highlight that the diastolic arterial blood pressure can decrease significantly when the Hypotension Prediction Index (hereafter the Index) reaches approximately 50 and argue for the possible need for treatment when diastolic arterial blood pressure is very low. We appreciate the comment and would like to highlight additional aspects that probably need to be considered in this context.First, Giustiniano and Nisi mention the “gray zone” as the reason for commenting on our publication. We want to stress that the exclusion of “gray-zone” outcomes (mean arterial pressure [MAP] between 65 and 75 mmHg) is not our primary concern with the method used to develop and validate the Index. Our primary concern is that the method also enforces that the prediction data for nonhypotension cannot have a MAP less than 75 mmHg, while the prediction data for hypotension can. The difference between simply implementing a “gray zone” and the biased selection of predictors actually used in development and validation of the Index3 is illustrated in figure 1, B and C. With the biased data selection (fig. 1C), the MAP value at the time of prediction unintendedly carries definitive information about the future outcome: a MAP less than 75 mmHg is always associated with upcoming hypotension. As a consequence, MAP becomes a highly specific predictor of hypotension—similar to what is reported for the Index. We recommend comparing the predictive value of the Index head-to-head with that of MAP.It is certainly possible that diastolic arterial blood pressure has additional value in predicting upcoming hypotension, not least because of the problematic method used to train the Index. Due to the selection bias, the Index algorithm has likely learned the artificially high predictive value of MAP at the cost of underutilizing other relevant predictors of hypotension.2However, we need to understand the selection bias’s effect on the Index’s actual predictive abilities before we can meaningfully discuss the Index’s interplay with other variables.In line with our theoretical discussion of the Index,2 Figure 1, A and B, in Giustiniano and Nisi’s letter and our own observations (fig. 2) demonstrate anecdotally three important aspects:We thank Giustiniano and Nisi for their relevant DAP comment and for sharing data that visualize the dissociation between clinical reality and reported classification results. Before we discuss the interplay of the Index and other physiologic variables, we should address the consequence of the selection bias.Dr. Vistisen is associate editor of the Journal of Clinical Monitoring and Computing. Dr. Vistisen has coauthored one of the published scientific papers aiming to validate the Hypotension Prediction Index. The other authors declare no competing interests.