Why the study?
Does the Hypotension Prediction Index accurately predict future hypotension compared to static mean arterial pressure values in surgical patients?
Does the Hypotension Prediction Index accurately predict future hypotension compared to static mean arterial pressure values in surgical patients?
Answers are grounded in this paper and cite their sources.
Evaluating Hypotension Prediction Index alerts rather than individual data points may provide a more clinically relevant assessment of its ability to predict impending hypotension.
To the Editor: We read with interest the editorial by Vistisen et al.1 and the accompanying infographic2 relating to the work by Davies et al.3 and Mulder et al.4 Our work3 was in response to a previous editorial suggesting that the predictive ability of the hypotension prediction index was artificially inflated due to the methodology employed.5 However, using the methodology suggested in the editorial,5 the predictive ability of the index was shown to be similar, with a high receiver operating curve area under the curve for predicting hypotension. In the editorial by Vistisen et al., a new topic of discussion has arisen: exploring the relationship between the index, mean arterial pressure (MAP), positive predictive values, and the ability of a single concurrent value of MAP to predict hypotension. The editorial and associated infographic state that a Hypotension Prediction Index of 85 agrees 97% of the time with an alert based on a MAP threshold of 72 mmHg. Although it may appear to be an overly detailed point, we would like to clarify that the article evaluates predictions when Hypotension Prediction Index is greater than 85, not equal to 85 (table S1 in Mulder et al.). This is important because at the Hypotension Prediction Index treatment threshold of 85, the supplementary material (fig. S1) of the article by Mulder et al.4 shows that MAP was less than 72 mmHg only approximately 50% of the time, which is consistent with data published in the EU HypoProtect study.6 The positive predictive values of both studies showed that 7 out of 10 data points were false positives. While this is a correct interpretation of the published statistical data, it is challenging to know whether interventions were performed based on either Hypotension Prediction Index or MAP values. Both articles acknowledged, which the editorial did not mention, that interventions were not able to be accounted for with sufficient granularity, and “that the resulting performance metrics, especially positive predictive values, may be underestimated.” Given that the time weighted average of hypotension in the work by Mulder et al.4 is low at 0.03, then it is possible that treatment of either the MAP or Hypotension Prediction Index was partly to cause for the low positive predictive values. Future studies would require highly annotated and accurate information on interventions to unravel this. In addition, the optimal way to assess positive predictive values for Hypotension Prediction Index may be to perform the analysis based on the Hypotension Prediction Index alerts, not the individual 20-s Hypotension Prediction Index values, as the alerts are more likely to drive clinician treatment. This approach has been used previously, where hypotension event and nonevent are evaluated in the context of a preceding Hypotension Prediction Index alarm.7–9 This type of approach assesses the positive predictive values based on clinically relevant use of Hypotension Prediction Index alerts and ensures a clinically relevant balance of the classes and reduces the impact of repeated measurements. Using pooled data from previous publications, we analyzed hypotension alerts at Hypotension Prediction Index 85 or greater threshold. Among 1,683 surgical patients approximately 75% of patients will experience hypotension in the next 15 min (data not previously reported). Finally, the use of a single value of MAP to predict future hypotension is contentious, with divergent views expressed. We do not dispute that predicting the trajectory of MAP (the LepMAP methodology10) is useful to clinicians, as suggested in the editorial. What we have attempted to highlight is that when using a single static value of a variable to predict its future value, it generates a high area under the receiver operating characteristics curve regardless of the variable or threshold chosen, and hence we question whether it is has meaningful clinical translation to a therapeutic target. As technology evolves, a consensus will be required on the optimal way to validate the predictive ability of these technologies in a clinically meaningful manner, as well as their impact on clinical outcomes. Competing Interests Dr. Davies notes consultancy fees, speaker fees and funded research from Edwards Lifesciences (Irvine, California). Dr. Mythen was paid consultancy and speaker fees from Edwards Lifesciences, and as of August 1, 2023, is a paid employee of Edwards Lifesciences. Drs. Hatib and Jian are paid employees of Edwards Lifesciences.
No takes yet. Share an insight, caveat, or question.
Davies et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: