Machine Learning Modeling of Intraoperative Hemodynamic Predictors of Postoperative Outcomes

Machine Learning Modeling of Intraoperative Hemodynamic Predictors of 30-day Mortality and Major In-hospital Morbidity After Noncardiac Surgery: a Retrospective Population Cohort Study

With population aging and limited resources, strategies to improve outcomes after surgery are ever more important. There is a limited understanding of what ranges of hemodynamic variables under anesthesia are associated with better outcomes. This retrospective cohort study will analyze how hemodynamic variables during surgeries predict mortality, morbidity, Intensive Care Unit admission, length of hospital stay, and hospital readmission. The use of machine learning in a large, broad surgery population dataset could detect new relationships and strategies that may inform current practice, and generate ideas for future research.

Study Overview

• Status: Completed
• Conditions: Death; Surgery; Anesthesia; Perioperative/Postoperative Complications; Morbidity, Multiple
• Intervention / Treatment: Other: Blood pressure; Other: Heart rate; Other: Use of hemodynamic medications (i.e. special medications for blood pressure); Other: Oxygen saturation by pulse oximetry (SpO2); Other: End-tidal Carbon dioxide (EtCO2)

Detailed Description

Lay Summary

Introduction: The World Health Organization estimates that 270-360 million operations are performed every year worldwide. Death and complications after surgery are a big challenge. In Canada, out of every 1000 major surgeries, 16 patients die in hospital after surgery. In the United States, for every 1000 operations, 67 patients unexpectedly need life support in the Intensive Care Unit. With population aging and limited resources, strategies to improve health after surgery are ever more important.

Vital signs, such as blood pressure and heart rate, show how the body is doing. Vital signs change during surgery because of patient, surgical, and anesthetic factors. Anesthesiologists can change vital signs with medications. However, medical professionals are only starting to understand which, and what ranges of, vital signs under anesthesia are associated with better health. Machine learning is a tool that can provide new ways to understand data. With better understanding, medical professionals can work to improve outcomes after surgery.

Objective: This study will analyze vital signs during surgeries for their links to death, complications (heart, lung, kidney, brain, infection), Intensive Care Unit admission, length of hospital stay, and hospital readmission. This study will determine which, and what levels of, vital signs may be harmful. The investigators predict that blood pressure, heart rate, oxygen level, carbon dioxide level, and the need for medications to change blood pressure will interact to be associated with death after surgery.

Methods: After obtaining Research Ethics Board approval, the investigators will analyze data from all patients who are at least 45 years old and had an operation (with the exception of heart surgery) with an overnight stay at the Queen Elizabeth II health centre (Halifax, Canada) from January 1, 2013 to December 1, 2017. There are approximately eligible 35,000 patients. The investigators will use machine learning to model the data and test how well our model explains outcomes after surgery.

Significance: The use of machine learning in a large, broad surgery population dataset could detect new relationships and strategies that may inform current practice, and generate ideas for future research. A better understanding of the impact of vital signs during surgeries may unveil methods to improve outcomes and resource allocation after surgery. The results may suggest ways to identify high-risk patients who should be monitored more closely after surgery. If the model performs well, it may motivate other researchers to use machine learning in health data research.

Please see full protocol for details.

May 2020 update (prior to dataset aggregation and analysis)

1. Added secondary outcome (days alive and out of hospital at 30 days postoperatively)
2. Improved hemodynamic variable artifact processing algorithm
3. Added sub-study: machine learning for invasive blood pressure artifact removal algorithm

• Study Type: Observational
• Enrollment (Anticipated): 35000

Participation Criteria

Eligibility Criteria

• Ages Eligible for Study: 45 years and older (Adult, Older Adult)
• Accepts Healthy Volunteers: No
• Genders Eligible for Study: All

• Sampling Method: Non-Probability Sample

Study Population

For data analysis in summer 2019, we have access to mortality data up to December 31, 2017. We chose December 1, 2017, as the last surgery date to be included, to allow for a complete data set 30 days after surgery. January 1, 2013 was chosen to obtain a study period of 5 years.

Description

Inclusion Criteria:

• All patients ages ≥ 45 receiving their index (i.e. first) non-cardiac surgery with an overnight stay at the Nova Scotia Health Authority Queen Elizabeth II (QEII) hospitals (Victoria General and Halifax Infirmary) Halifax, Canada, from January 1, 2013 to December 1, 2017.
• For patients who had multiple surgeries, only the first non-cardiac surgery with an overnight stay at QEII will be included to avoid confounding from previous surgical admissions (i.e. one surgical admission per patient).

Exclusion Criteria:

• No intraoperative anesthetic records
• Cardiac surgery patients
• Deceased organ donation

Study Plan

How is the study designed?

Design Details

• Observational Models: Cohort
• Time Perspectives: Retrospective

Number of groups / cohorts

1

Cohorts and Interventions

Group / Cohort

Intervention / Treatment

Cohort; Patients ages ≥ 45 receiving their index (i.e. first) non-cardiac surgery with an overnight stay at the Nova Scotia Health Authority Queen Elizabeth II (QEII) hospitals (Victoria General and Halifax Infirmary) Halifax, Canada, from January 1, 2013 to December 1, 2017 will be included. Patients under going cardiac surgery or deceased organ donation will be excluded. Patients without an electronic anesthetic record during surgery will also be excluded. Preliminary analysis of the intraoperative database estimates approximately 35,000 patients in this cohort.

Other: Blood pressure; Systolic Blood Pressure (SBP); Maximum change from preoperative SBP, in a) absolute change (mmHg), and b) relative change (%)(emergency and elective cases analyzed separately); Cumulative duration (minutes) >=20% below preoperative SBP; Longest single episode (minutes) below a) 80, b) 90, and c)100 mmHg; Cumulative duration (minutes) below a) 80, b) 90, and c)100 mmHg Mean Arterial Pressure (MAP); Maximum change from preoperative MAP, in a) absolute change (mmHg), and b) relative change (%) (emergency and elective cases analyzed separately); Cumulative duration (minutes) >=20% below preoperative MAP; Longest single episode (minutes) below a) 60, b) 65, c) 70, and d) 80mmHg; Cumulative duration (minutes) below a) 60, b) 65, c) 70, and d) 80mmHg; Other: Heart rate; Maximum change (beats per minute, BPM) from preoperative heart rate (positive and negative); Relative change (%) from preoperative heart rate (positive and negative); Maximum pulse variation (maximum heart rate minus minimum heart rate); Longest single episode (minutes) a) below 60, and b) above 100BPM; Cumulative duration (minutes) a) below 60, and b) above 100BPM; Other: Use of hemodynamic medications (i.e. special medications for blood pressure); Vasopressor/inotrope use (yes vs. no): phenylephrine, norepinephrine, epinephrine, vasopressin, dobutamine, or milrinone; Infusion of any vasopressor/inotropes above (yes vs. no) (identified by unit of weight over time); Phenylephrine/ephedrine bolus (yes vs. no) (identified by unit of weight only); Vasodilator use (yes vs. no): labetalol, esmolol, nitroglycerin, nitroprusside; Infusion of any vasodilator above (yes vs. no) (identified by unit of weight over time); Other: Oxygen saturation by pulse oximetry (SpO2); Longest single episode (minutes) below a) 88, and b) 90%; Cumulative duration (minutes) below a) 88, and b) 90%; Other: End-tidal Carbon dioxide (EtCO2); Longest single episode (minutes) a) below 30, and b) above 45mmHg; Cumulative duration (minutes) a) below 30, and b) above 45mmHg

What is the study measuring?

Primary Outcome Measures

Outcome Measure	Measure Description	Time Frame
Mortality	All-cause postoperative mortality (yes/no)	30 days after date of surgery

Secondary Outcome Measures

Outcome Measure	Measure Description	Time Frame
In-hospital Morbidity: Any	Any complications in terms of cardiac, respiratory, renal, cerebrovascular, delirium, or septic shock (yes/no)	30 days after date of surgery
In-hospital Morbidity: Cardiac	Composite of acute myocardial infarction, cardiac arrest, ventricular tachycardia, congestive heart failure, pulmonary edema, complete heart block, shock excluding septic shock (yes/no)	30 days after date of surgery
In-hospital Morbidity: Respiratory	Composite of pneumonia, pulmonary embolism, acute respiratory failure, respiratory arrest, Mechanical Ventilation >= 96 hours (yes/no)	30 days after date of surgery
In-hospital Morbidity: Acute Kidney Injury	Acute Kidney Injury (yes/no)	30 days after date of surgery
In-hospital Morbidity: Cerebrovascular	Composite of strokes and transient ischemic attacks (yes/no)	30 days after date of surgery
In-hospital Morbidity: Delirium	Delirium (yes/no)	30 days after date of surgery
In-hospital Morbidity: Septic Shock	Septic Shock (yes/no)	30 days after date of surgery
Postoperative ICU admission	ICU admission (yes/no)	30 days after date of surgery
Prolonged Postoperative Length of Stay (LOS)	Greater than vs. less than or equal to Canadian Institute of Health Information Expected Length of Stay (ELOS) as assigned by the Case Mix Grouping	30 days after date of surgery
Hospital readmission	Hospital readmission (yes/no)	30 days after date of surgery
Intraoperative mortality	Intraoperative mortality (yes/no)	30 days after date of surgery
Days alive and out of hospital at 30 days postoperatively	Number of days	30 days after date of surgery

Collaborators and Investigators

• Sponsor: Janny Xue Chen Ke
• Collaborators: University of Toronto; University of Ottawa; Dalhousie University; Harvard University

Publications and helpful links

General Publications

• Ke JXC, McIsaac DI, George RB, Branco P, Cook EF, Beattie WS, Urquhart R, MacDonald DB. Postoperative mortality risk prediction that incorporates intraoperative vital signs: development and internal validation in a historical cohort. Can J Anaesth. 2022 Sep;69(9):1086-1098. doi: 10.1007/s12630-022-02287-0. Epub 2022 Aug 22.

Study record dates

Study Major Dates

• Study Start (Actual): January 1, 2013
• Primary Completion (Actual): December 31, 2017
• Study Completion (Actual): December 31, 2017

Study Registration Dates

• First Submitted: July 7, 2019
• First Submitted That Met QC Criteria: July 7, 2019
• First Posted (Actual): July 10, 2019

Study Record Updates

• Last Update Posted (Actual): June 30, 2020
• Last Update Submitted That Met QC Criteria: June 27, 2020
• Last Verified: June 1, 2020

More Information

Terms related to this study

• Keywords: Perioperative; Anesthesia; Hemodynamics; Machine Learning; Non-cardiac Surgery
• Additional Relevant MeSH Terms: Pathologic Processes; Postoperative Complications
• Other Study ID Numbers: 1024251

Plan for Individual participant data (IPD)

• Plan to Share Individual Participant Data (IPD)?: No

Drug and device information, study documents

• Studies a U.S. FDA-regulated drug product: No
• Studies a U.S. FDA-regulated device product: No