Hypotension prediction index to predict hemodynamic instability related to renal replacement therapy in critically ill patients: A prospective observational study

Fluid removal during continuous renal replacement therapy (CRRT) is a frequent cause of hemodynamic instability (HI) in critically ill patients, yet clinicians lack reliable tools to anticipate intolerance to ultrafiltration. We aimed to determine the performance and clinical utility of the hypotension prediction index (HPI) to predict HI in critically ill patients undergoing fluid removal with CRRT. We conducted a prospective observational study including consecutive patients receiving CRRT with fluid removal in a tertiary intensive care unit (ICU) of Centre Hospitalier Universitaire Vaudois, Lausanne, Switzerland, between March 2022 and August 2023. Arterial pressure waveforms were recorded using FloTrac and HemoSphere Advanced Monitor and HPI was computed every 20 s (offline, for blinding). An HPI alarm was defined as three consecutive HPI values ≥85; an HI event was defined as three consecutive mean arterial pressure readings <65 mmHg, initiation or escalation of vasopressor therapy, or clinical signs of hemodynamic intolerance. We evaluated the performance of HPI for predicting HI within the following 15-min window. To assess HPI clinical utility at bedside, we computed alarm-to-event delays, the proportion of HPI alarms followed by HI and the proportion of HI preceded by at least one HPI alarm. We repeated analyses considering only “new” alarms and “new” HI, defined respectively as alarms and HI events preceded by a 60-min event-free interval. We recorded 223,660 observations corresponding to 1243 h of monitoring. HPI was ≥85 in 33.4% of observations and MAP <65 mmHg in 11.0%. Overall, HPI demonstrated good discriminative performance for predicting HI (area under the curve of the receiver operating characteristic curve=0.89), with high sensitivity (99.7%) and specificity (81.6%). However, positive predictive value (PPV) was low, as only 54.7% of alarms were followed by HI, with a mean alarm-to-event delay of 103 s. Bootstrapped analyses performed to adjust for varying importance of individual patients confirmed these findings. When considering new alarms ( n =147) and events ( n =174), only 33 (22.4%) alarms were followed by an HI, and 34 (19.4%) HI were preceded by an alarm. Sensitivity analyses considering a lower cutoff for HPI and a shorter event-free period to define a new event confirmed these findings. Although HPI shows excellent statistical performance for predicting HI during CRRT with fluid removal, its clinical utility is limited by short time-to-event delay, frequent false-positive alarms, and prolonged time with high HPI values. These findings underscore the gap between predictive accuracy and bedside utility of monitoring tools in critical care.