We have read the interesting comment by Giustiniano and Nisi  about our article.  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 Index 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.

Fig. 1.
Visualization of the selection problem in development and validation of Hypotension Prediction Index on simulated data. (A –C) Different data selection strategies. Upper panels show simulated pairs of mean arterial pressure (MAP). Values on the ordinate represent the outcome to predict; it is categorized into hypotension and nonhypotension. Values on the abscissa represent MAP in the prediction data 5 min before the event. The lower panels are the area under the receiver operating characteristics curve (AUC) showing MAP’s ability to discriminate hypotensive events from nonhypotensive events. The data are generated with a moderate correlation (r = 0.5) to make MAP a modest predictor of hypotension in the full analysis (A). The simulation is not an attempt to produce realistic data.

Visualization of the selection problem in development and validation of Hypotension Prediction Index on simulated data. (A –C) Different data selection strategies. Upper panels show simulated pairs of mean arterial pressure (MAP). Values on the ordinate represent the outcome to predict; it is categorized into hypotension and nonhypotension. Values on the abscissa represent MAP in the prediction data 5 min before the event. The lower panels are the area under the receiver operating characteristics curve (AUC) showing MAP’s ability to discriminate hypotensive events from nonhypotensive events. The data are generated with a moderate correlation (r = 0.5) to make MAP a modest predictor of hypotension in the full analysis (A). The simulation is not an attempt to produce realistic data.

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. 

However, 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,  Figure 1, A and B, in Giustiniano and Nisi’s letter and our own observations (fig. 2) demonstrate anecdotally three important aspects:

Fig. 2.
A 7-h postsurgical recording of simultaneous mean arterial pressure (MAP; A) and Hypotension Prediction Index values (B); one pair every 20 s. (C) Scatter plot of simultaneous MAP and Hypotension Prediction Index values. Spearman’s rho is the ranked correlation coefficient. For this patient, MAP was to be kept above 65 mmHg. The Hypotension Prediction Index was not presented for clinicians on the bedside monitor.

A 7-h postsurgical recording of simultaneous mean arterial pressure (MAP; A) and Hypotension Prediction Index values (B); one pair every 20 s. (C) Scatter plot of simultaneous MAP and Hypotension Prediction Index values. Spearman’s rho is the ranked correlation coefficient. For this patient, MAP was to be kept above 65 mmHg. The Hypotension Prediction Index was not presented for clinicians on the bedside monitor.

  1. The Index appears to almost be a mirror of MAP. The relation is not linear, but the two variables reflect almost the same information.
  2. When the Index reaches 85 (the default alarm level), MAP is approximately 72 to 76 mmHg. Using an Index value of 85 as a treatment trigger may not be different from using a MAP treatment trigger of approximately 72 to 76 mmHg. The latter approach could also reduce time spent in hypotension.
  3. Hypotension (MAP less than 65 mmHg for at least 1 min) does not occur in Giustiniano and Nisi’s figure 1A for the last 50 min; nevertheless, the Index is above or toggles around 85. This stable period has no reported interventions. Such data segments containing MAP less than 75 mmHg and associated with nonhypotension are excluded in the validation studies using the biased selection, giving an unreasonably high positive predictive value.

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.