An Algorithm Developed by Machine Learning Put to the Test - the Compensatory Reserve Index and General Anesthesia

Author: Guy Zahavi

Background - The compensatory reverse index (CRI) is an algorithm designed to prospectively identify ongoing loss of blood and estimate the progression toward hemodynamic decompensation and the onset of shock in a bleeding patient. The CRI value is from 0 (total hemodynamic decompensation) to 1 (hemodynamic stability) and is based on the signal from a simple pulse oximeter. The algorithm was developed with machine learning and unsupervised feature extraction methods. This way the algorithm itself must group the data during algorithm development and make sense of it without human supervision. The CRI has been shown to correlate with central hypovolemia induced in human subjects under laboratory conditions by applying negative pressure to the lower body, and controlled low-volume blood loss. The hemodynamic compensation of trauma patients is negatively influenced by the effects of general anesthesia. We found it important to assess whether the CRI reflects the hemodynamic depression commonly induced by anesthetics. The aim of the present preliminary study was to compare changes in CRI to changes in mean blood pressure and heart rate induced by the induction of general anesthesia in elective healthy patients. Patients - Following ethics committee approval and personal informed consent 23 healthy patients ages 20-40 years old, scheduled for elective minor elective operations were included. Protocol - Patients were connected to routine anesthesia monitoring (electrocardiogram, non-invasive blood pressure, pulse oximetry and capnography) and the CRI was monitored using a device named CipherOx (Flashback technologies). Anesthesia induction included midazolam 1-2mg, fentanyl 2 mcg/kg and propofol 2 mg/kg. Parameters including heart rate, blood pressure and CRI were collected before the induction of anesthesia and then every minute for five minutes after the induction. Results - Heart rate was 79±18, 72±15 and 77±17; Mean blood pressure was 100±12mmHg, 75±9mmHg, and 76±10mmHg; CRI was 0.81±0.18, 0.85±0.13 and 0.81±0.18; before induction, during lowest blood pressure measurement after induction and five minutes following induction, respectively. Mean blood pressure but not heart rate or CRI changed following the induction of anesthesia (p<0.0001). Discussion - The induction of general anesthesia was followed by an expected decrease in blood pressure, reflecting hemodynamic decompensation induced by general anesthetics. Surprisingly, in most patients the compensatory reserve index increased despite a decrease in blood pressure, wrongly implying better hemodynamic compensation. Finding the cause of this unexpected increase in the compensatory reserve index would require going through the features extracted by machine learning during algorithm development. This information raises concerns regarding the use of an algorithm developed by machine learning in laboratory conditions in situations that are less sterile and more complex.

Co Author/Co-Investigator Names/Professional Title: Zahavi Guy, MD Ivry Michal, MD Orkin Dina, MD Berkenstadt Haim, MD

Funding Acknowledgement (If Applicable): No external funding and no competeing interests declared.