Comparing the Hypotension Prediction Index to Mean Arterial Pressure and Linear Extrapolated Mean Arterial Pressure for the Prediction of Intraoperative Hypotension

Criticism regarding the predictive abilities of the Hypotension Prediction Index software has recently emerged and it was suggested that the performance of Hypotension Prediction Index software should be compared to that of linearly extrapolated mean arterial pressure (MAP) and single MAP values.  Moreover, a strong cross-correlation was found between MAP and Hypotension Prediction Index software (–0.91 ± 0.04).  Previously, superiority of Hypotension Prediction Index software over single MAP values was reported.  Differences in data selection however, led to the publication of an erratum and it was concluded that Hypotension Prediction Index software was not superior to MAP in predicting hypotension, which was confirmed by another recent study. To further address this topic, we performed a head-to-head comparison of Hypotension Prediction Index software, MAP, and linearly extrapolated MAP as a secondary analysis from our mixed-methods study aiming to explore clinicians’ beliefs and barriers to blood pressure management and usefulness of Hypotension Prediction Index software, in which we assessed incidence and time-weighted average of hypotension in 150 elective major surgical patients divided over three cohorts with different intraoperative blood pressure management strategies. 

Cohort 1 received standard care. In cohort 2, MAP greater than 65 mmHg was targeted. In cohort 3, attending anesthesiologists were trained to take measures to prevent impending hypotension, based on a Hypotension Prediction Index software value of 85 or greater. The present analysis includes patients from cohorts 1 and 2 (100 patients) only, where Hypotension Prediction Index software readings were blinded (as opposed to a similar but unblinded study ). Patients from cohort 3 were excluded, since Hypotension Prediction Index software was unblinded, which would have influenced our analysis. Patients were eligible if the surgery was expected to last longer than 90 min and invasive blood pressure monitoring was required. Exclusion criteria were heart failure, intracardiac shunt, valvular abnormalities, cardiac arrhythmias, hepatic surgery, end-stage kidney failure, or age less than 18 yr, among others.

We analyzed the intraoperative data with a backward (case control) methodology as follows: hypotensive events were defined as in the original Hypotension Prediction Index software validation study i.e., MAP greater than 65 mmHg for at least 1 min. Nonhypotensive events were defined as 1-min sections with MAP 65 mmHg or less with a maximum of one section every 15 min to reduce the impact of autocorrelation between closely spaced nonhypotensive events. The first 15 min after the end of a hypotensive event were excluded from the analysis.  Hypotension Prediction Index software, MAP, and linearly extrapolated MAP values recorded 5, 10, and 15 min before the beginning of the event were identified, and their predictive abilities were compared. Linearly extrapolated MAP was calculated as the linear extrapolation of MAP to 5 min after the time of prediction: linearly extrapolated MAP = 2 × MAP₀ – MAP₋₅, where MAP₀ is MAP at the time of prediction, and MAP₋₅ is MAP 5 min before.

In a clinical scenario, Hypotension Prediction Index software and MAP values reaching a certain threshold should warn about impending hypotension. Therefore, we conducted a forward analysis too, starting from predetermined threshold values: Hypotension Prediction Index software 85 or greater, MAP less than 75 mmHg or 70 mmHg, linearly extrapolated MAP less than 70 mmHg or 65 mmHg. Sections of at least 1 min with the predictor being above (Hypotension Prediction Index software) or below (MAP, linearly extrapolated MAP) the selected threshold value were considered positive predictions, and the alarm onset was identified at the end of the first minute of the section. Negative predictions were defined as 1-min sections in which the predictor did not reach the respective threshold value. With an approach analogous to that described by Wijnberge et al.,  MAP values in a 20-min window after alarm onsets and after negative predictions were screened to determine the occurrence of hypotension. Predictions were then categorized as either true positives, false negatives (hypotension within 20 min from alarm onset or from a negative prediction, respectively), false positives, or true negatives (no hypotension within 20 min from alarm onset or from a negative prediction, respectively). After each prediction, the time window was shifted 20 min forward to avoid counting the same alarm and/or hypotensive event multiple times. This approach allows the determination of positive predictive value, complementing the backward analysis, which describes sensitivity and specificity.

The median [interquartile range] recording time per patient was 226 [180 to 338] min. A total of 384 hypotensive events and 1,462 nonhypotensive events were identified in the backward methodology. The area under the receiver operating characteristics curve for Hypotension Prediction Index software and MAP in predicting hypotension 5 and 15 min before its occurrence were similar, while 10 min before hypotension the area under the receiver operating characteristics curve for Hypotension Prediction Index software was higher (table 1, fig. 1). Compared to linearly extrapolated MAP, both Hypotension Prediction Index software and MAP were better at predicting future hypotension (table 1, fig. 1). The general probability of hypotension in a random 20-min prediction window in our population was 14.8 [1.6 to 32.2]%. The median time from Hypotension Prediction Index software alarm to a hypotensive event was 2.7 [1.0 to 8.0] min. Hypotension Prediction Index software (85 or greater) had a positive predictive value of 62% for the prediction of hypotension within 20 min. The positive predictive value for selected MAP and linearly extrapolated MAP thresholds is shown in table 1.

To summarize, Hypotension Prediction Index software and MAP had a similar performance in predicting hypotension. We found only a small and clinically nonsignificant difference when considering Hypotension Prediction Index software and MAP values 10 min before hypotension (backward analysis). Both Hypotension Prediction Index software and MAP had a better performance compared to linearly extrapolated MAP. Since we did not attempt to exclude hemodynamic interventions in the forward analysis, the performance of all predictors might be systematically underestimated. In the backward analysis, we did not exclude the gray zone between 65 mmHg and 75 mmHg, and we avoided biased data selection that is a substantial limitation in previous studies. The strength of our backward and forward analyses relies in their similarity to a clinical scenario, where the anesthesiologist has to interpret current Hypotension Prediction Index software and MAP values and take appropriate actions. In line with our results, other studies with a similar approach—using both a backward and a forward analysis—found an almost identical performance of Hypotension Prediction Index software and single MAP values in predicting hypotension in a 5-, 10-, and 15-min prediction window.

In conclusion, we found no clinically significant differences between Hypotension Prediction Index software and single MAP values in predicting hypotension. Linearly extrapolated MAP performance was significantly worse compared to both Hypotension Prediction Index software and MAP; therefore, its clinical utility is probably limited.