Performance of a Machine Learning Algorithm to Predict Hypotension in Spontaneously Breathing Non-Ventilated Post-Anesthesia and ICU Patients

Johan T M Tol; Lotte E Terwindt; Santino R Rellum; Marije Wijnberge; Björn J P van der Ster; Eline Kho; Markus W Hollmann; Alexander P J Vlaar; Denise P Veelo; Jimmy Schenk

doi:10.3390/jpm14020210

Performance of a Machine Learning Algorithm to Predict Hypotension in Spontaneously Breathing Non-Ventilated Post-Anesthesia and ICU Patients

J Pers Med. 2024 Feb 15;14(2):210. doi: 10.3390/jpm14020210.

Authors

Johan T M Tol¹, Lotte E Terwindt¹, Santino R Rellum^{1

2}, Marije Wijnberge¹, Björn J P van der Ster¹, Eline Kho^{1

2}, Markus W Hollmann^{1

3}, Alexander P J Vlaar^{2

3}, Denise P Veelo¹, Jimmy Schenk^{1

2

4}

Affiliations

¹ Department of Anesthesiology, University of Amsterdam, Amsterdam UMC, Meibergdreef 9, 1105AZ Amsterdam, The Netherlands.
² Department of Intensive Care Medicine, University of Amsterdam, Amsterdam UMC, Meibergdreef 9, 1105AZ Amsterdam, The Netherlands.
³ Laboratory of Experimental Intensive Care and Anesthesiology, Academic Medical Center Location, University of Amsterdam, Amsterdam UMC, Meibergdreef 9, 1105AZ Amsterdam, The Netherlands.
⁴ Department of Epidemiology and Data Science, University of Amsterdam, Amsterdam UMC, Meibergdreef 9, 1105AZ Amsterdam, The Netherlands.

Abstract

Background: Hypotension is common in the post-anesthesia care unit (PACU) and intensive care unit (ICU), and is associated with adverse patient outcomes. The Hypotension Prediction Index (HPI) algorithm has been shown to accurately predict hypotension in mechanically ventilated patients in the OR and ICU and to reduce intraoperative hypotension (IOH). Since positive pressure ventilation significantly affects patient hemodynamics, we performed this validation study to examine the performance of the HPI algorithm in a non-ventilated PACU and ICU population. Materials & Methods: The performance of the HPI algorithm was assessed using prospectively collected blood pressure (BP) and HPI data from a PACU and a mixed ICU population. Recordings with sufficient time (≥3 h) spent without mechanical ventilation were selected using data from the electronic medical record. All HPI values were evaluated for sensitivity, specificity, predictive value, and time-to-event, and a receiver operating characteristic (ROC) curve was constructed. Results: BP and HPI data from 282 patients were eligible for analysis, of which 242 (86%) were ICU patients. The mean age (standard deviation) was 63 (13.5) years, and 186 (66%) of the patients were male. Overall, the HPI predicted hypotension accurately, with an area under the ROC curve of 0.94. The most used HPI threshold cutoff in research and clinical use, 85, showed a sensitivity of 1.00, specificity of 0.79, median time-to-event of 160 s [60-380], PPV of 0.85, and NPV of 1.00. Conclusion: The absence of positive pressure ventilation and the influence thereof on patient hemodynamics does not negatively affect the performance of the HPI algorithm in predicting hypotension in the PACU and ICU. Future research should evaluate the feasibility and influence on hypotension and outcomes following HPI implementation in non-ventilated patients at risk of hypotension.

Keywords: ICU; PACU; critical illness; hypotension; machine learning; prediction.

Grants and funding

Edwards Lifesciences provided funding for the HYPE trial and PHYSIC study after their design. Edwards Lifesciences was not involved in the design or execution of the present study, nor in the drafting or approval of the manuscript.