Why the study?
Does the Hypotension Prediction Index (HPI) algorithm improve the prediction of hypotensive events compared to current mean arterial pressure?
Does the Hypotension Prediction Index (HPI) algorithm improve the prediction of hypotensive events compared to current mean arterial pressure?
Answers are grounded in this paper and cite their sources.
The Hypotension Prediction Index algorithm appears to offer no clinically significant advantage over current mean arterial pressure monitoring for predicting hypotensive events.
To the Editor: We read with interest the recent article by Davies et al.1 in Anesthesiology on the evaluation of Edwards Lifesciences’ (USA) Hypotension Prediction Index (HPI). The HPI algorithm takes a 20-s interval of a blood pressure waveform as input and outputs a unitless value between 0 and 100, where values exceeding 85 predict a hypotensive event (defined as a mean arterial pressure [MAP] of less than 65 mmHg for at least 1 min) within the next 15 min.2 Publications on this algorithm by members and consultants of the company have recently generated controversy3 for two reasons. First, the HPI algorithm was developed and initially evaluated using a “backward analysis,” in which only hypotensive events, but not nonevents, could be preceded by an MAP less than 75 mmHg.4 These event definitions suggest that the performance of the HPI algorithm may have been overestimated in initial studies. Thus, a “forward analysis” is needed to reveal its real-world performance in which nonevents could also be preceded by MAP less than 75 mmHg and are in clinically observed proportions to the far fewer hypotensive events. Second, the HPI algorithm had not been fairly compared to the current MAP (average of the blood pressure waveform over the same 20-s interval) in predicting hypotension.5 Thus, comparisons are needed to reveal whether the HPI algorithm offers added value over the standard of care. Davies et al.1 have now applied both forward and backward analyses to evaluate the HPI algorithm and performed fair comparisons with the current MAP (“MAP algorithm”) in a study of 2,022 patients. They reported that the performance of the HPI algorithm is similar for forward and backward analyses with receiver operating characteristic area under the curve (AUC) values greater than 0.92, consistent with their earlier studies. However, they appear to have applied the forward analysis in a way that overestimates the predictive performance and did not show the results clearly, especially the comparisons between the HPI and MAP algorithms. Figure 2 in the Davies et al.1 article illustrates their methodology for implementing the forward analysis. As can be seen, if HPI is less than 85 for a given sample of data (20-s interval of the blood pressure waveform) and a hypotensive event follows within 5, 10, or 15 min, the sample is not automatically labeled as a false negative—as it should be. Instead, if HPI is greater than 85 “at least 2 min in advance” of that hypotensive event, the sample may be discarded. In other words, a false negative may have been removed, even though the HPI algorithm was unable to predict the hypotensive event at the sample. Figure 4 from the Davies et al.1 article shows that the HPI algorithm yielded virtually the same receiver operating characteristic curves for within 5, 10, and 15 min ahead predictions in the forward analysis. Intuitively, sensitivity should be highest for predictions within 15 min ahead, whereas specificity should be highest for predictions within 5 min ahead. The identical receiver operating characteristic curves in figure 4 may be a result of discarding the false negatives. We thus believe their results for the forward analysis may be overestimated. Figure 3 in the Davies et al.1 article shows all HPI values plotted against the current MAP values. The plot suggests that the two variables differ substantially. However, each blue data point in the plot can represent one to many samples. In other words, a blue outlier data point could represent just one sample of data, whereas a single blue data point near the white trace can represent thousands of data samples. The error bars for the white trace reflect 2 SDs. Use of SD for error bars is inappropriate in the case of outliers. Furthermore, the strength of the relationship between the HPI and current MAP values is reported via an R2 value of 0.77. The R2 value measures the strength of a linear relationship. However, the HPI algorithm employs logistic regression, so the potential relationship is expected to be sigmoidal. The R2 value should have been computed after applying a logistic function to MAP. We thus believe that the HPI and current MAP values may be more tightly related than what is suggested in the article. The HPI and MAP results in predicting hypotension would ideally have been presented side by side in the Davies et al.1 article. Our figure 1 shows the results reported in the article but in a way to facilitate comparison. The receiver operating characteristic AUC values for the HPI and MAP algorithms (reported in the abstract) are virtually identical and about 0.93 for the forward analysis (fig. 1A). The receiver operating characteristic curves for the HPI and MAP algorithms are likely superimposable (see fig. 1B, in which the red trace in the left panel of fig. 4 and the green trace in the right panel of fig. 5 of the Davies et al.1 article are approximately plotted on the same axes using PlotDigitizer6 [Porbital]). The crucial positive predictive values for predictions within 5 min ahead in the forward analysis (reported in supplemental spreadsheets) are less than 30% for the HPI algorithm at a decision threshold of 83 (near the standard 85) and for the MAP algorithm at decision threshold of 70.5 (fig. 1C). Such low positive predictive values may only exacerbate the alarm fatigue problem. Unfortunately, the positive predictive values for the HPI algorithm at the 85 decision threshold were not provided. A study of the HPI and MAP algorithms published in the same issue of Anesthesiology showed somewhat lower receiver operating characteristic AUC values of 0.88 to 0.89 and mildly higher positive predictive values of 32 to 33% for within 5 min ahead prediction in 100 patients.7 Despite the nearly identical results for the HPI and MAP algorithms appearing in the Davies et al.1 article, a net reclassification improvement of 7.6% was claimed to be offered by the HPI algorithm at the 85 decision threshold over the MAP algorithm at an assumed decision threshold of 72 mmHg. This result is tenuous, because reducing the threshold to 71 mmHg would likely abolish the small improvement. We thus believe that the data presented in the Davies et al.1 article indicate that the HPI algorithm and current MAP are essentially equivalent in predicting hypotension.Fig. 1.: Hypotension Prediction Index (HPI; Edwards Lifesciences, USA) versus current mean arterial pressure (MAP) in predicting hypotension reported by Davies et al. 1 but shown here side by side. (A) Receiver operating characteristic area under the curve (ROC AUC) comparisons (see abstract in Davies et al. 1). (B) ROC curve comparisons (see red trace(s) in the left panel of fig. 4 in Davies et al. 1 and green trace in the right panel of fig. 5 in Davies et al. 1). The ROC curves were approximated here via PlotDigitizer.6 (C) Positive predictive value (PPV) comparisons (see supplementary Excel spreadsheet for forward analysis and “ROC” tab accompanying1). The values are means (lower 95% CI to upper 95% CI). All results here are from invasive blood pressure waveforms with intervention data, artifact, and prediction during hypotensive events excluded.Davies et al.1 ask the question: can MAP predict itself (“If it is assumed that MAP can predict itself. . . .”)? They argue that the answer is no by chalking up the relationship between current and future MAP as “mathematical coupling.” To support their argument, they cite an earlier publication by Archie,8 which concerns when two variables contain the same variable (e.g., MAP and systematic vascular resistance, which is calculated as MAP divided by cardiac output, or cardiac output and heart rate, because cardiac output fundamentally equals heart rate times stroke volume); their correlation may appear artificially high. The work of Archie,8 on the contrary, does not support the argument of Davies et al.1 for two reasons. First, the current MAP and MAP within 5 to 15 min ahead do not contain a common variable (e.g., cardiac output and systematic vascular resistance, which are the components of MAP, change over this time scale). Second, the MAP algorithm does not try to determine how well the current and future MAP correlate but rather whether the current MAP can predict future MAP less than 65 mmHg. Such time series, in fact, are not used as examples or even mentioned by Archie.8 It is also important to note that the argument of Davies et al.1 is an academic one and does not offer any practical implication, because the MAP algorithm can be clinically deployed in the exact same way as the HPI algorithm. Davies et al.1 further argue that if MAP can be used to predict hypotension, then inappropriate treatments will result. Because their data show that the HPI algorithm and MAP algorithm at a decision threshold of approximately 71 mmHg are equivalent in predicting hypotension, the same could be said for the HPI algorithm. It should also be mentioned that blanket targeting of higher MAP for treatment has not improved patient outcomes up to now.9 We conclude that Davies et al.1 showed that the HPI algorithm and current MAP are equal in predicting hypotension with overestimated but still low positive predictive value. Competing Interests Dr. Mukkamala is cofounder of and holds equity interests in Retia Medical Systems (White Plains, New York). Dr. Mahajan is cofounder of Sensydia Inc. (Los Angeles, California) and chief medical officer for Pipcare (New York, New York). Dr. Schnetz declares no competing interests.
No takes yet. Share an insight, caveat, or question.
Mukkamala et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: