학술논문
A Machine Learning Approach for Predicting Real-time Risk of Intraoperative Hypotension in Traumatic Brain Injury.
Document Type
Academic Journal
Author
Feld SI; Anesthesiology and Pain Medicine, University of Washington.; Hippe DS; The Mountain-Whisper-Light: Statistics & Data Science, Seattle, WA.; Miljacic L; The Mountain-Whisper-Light: Statistics & Data Science, Seattle, WA.; Polissar NL; The Mountain-Whisper-Light: Statistics & Data Science, Seattle, WA.; Newman SF; Anesthesiology and Pain Medicine, University of Washington.; Nair BG; Anesthesiology and Pain Medicine, University of Washington.; Vavilala MS; Anesthesiology and Pain Medicine, University of Washington.
Source
Publisher: Lippincott Williams & Wilkins Country of Publication: United States NLM ID: 8910749 Publication Model: Print-Electronic Cited Medium: Internet ISSN: 1537-1921 (Electronic) Linking ISSN: 08984921 NLM ISO Abbreviation: J Neurosurg Anesthesiol Subsets: MEDLINE
Subject
Language
English
Abstract
Background: Traumatic brain injury (TBI) is a major cause of death and disability. Episodes of hypotension are associated with worse TBI outcomes. Our aim was to model the real-time risk of intraoperative hypotension in TBI patients, compare machine learning and traditional modeling techniques, and identify key contributory features from the patient monitor and medical record for the prediction of intraoperative hypotension.
Methods: The data included neurosurgical procedures in 1005 TBI patients at an academic level 1 trauma center. The clinical event was intraoperative hypotension, defined as mean arterial pressure <65 mm Hg for 5 or more consecutive minutes. Two types of models were developed: one based on preoperative patient-level predictors and one based on intraoperative predictors measured per minute. For each of these models, we took 2 approaches to predict the occurrence of a hypotensive event: a logistic regression model and a gradient boosting tree model.
Results: The area under the receiver operating characteristic curve for the intraoperative logistic regression model was 0.80 (95% confidence interval [CI]: 0.78-0.83), and for the gradient boosting model was 0.83 (95% CI: 0.81-0.85). The area under the precision-recall curve for the intraoperative logistic regression model was 0.16 (95% CI: 0.12-0.20), and for the gradient boosting model was 0.19 (95% CI: 0.14-0.24). Model performance based on preoperative predictors was poor. Features derived from the recent trend of mean arterial pressure emerged as dominantly predictive in both intraoperative models.
Conclusions: This study developed a model for real-time prediction of intraoperative hypotension in TBI patients, which can use computationally efficient machine learning techniques and a streamlined feature-set derived from patient monitor data.
Competing Interests: Unrelated to this study, B.G.N. holds equity in Perimatics LLC and is its Chief Solution Architect. D.S.H. reports research grants from GE Healthcare, Philips Healthcare, Canon America Medical Systems, and Siemens Healthineers, outside this study. The remaining authors have no conflicts of interest to declare.
(Copyright © 2021 Wolters Kluwer Health, Inc. All rights reserved.)
Methods: The data included neurosurgical procedures in 1005 TBI patients at an academic level 1 trauma center. The clinical event was intraoperative hypotension, defined as mean arterial pressure <65 mm Hg for 5 or more consecutive minutes. Two types of models were developed: one based on preoperative patient-level predictors and one based on intraoperative predictors measured per minute. For each of these models, we took 2 approaches to predict the occurrence of a hypotensive event: a logistic regression model and a gradient boosting tree model.
Results: The area under the receiver operating characteristic curve for the intraoperative logistic regression model was 0.80 (95% confidence interval [CI]: 0.78-0.83), and for the gradient boosting model was 0.83 (95% CI: 0.81-0.85). The area under the precision-recall curve for the intraoperative logistic regression model was 0.16 (95% CI: 0.12-0.20), and for the gradient boosting model was 0.19 (95% CI: 0.14-0.24). Model performance based on preoperative predictors was poor. Features derived from the recent trend of mean arterial pressure emerged as dominantly predictive in both intraoperative models.
Conclusions: This study developed a model for real-time prediction of intraoperative hypotension in TBI patients, which can use computationally efficient machine learning techniques and a streamlined feature-set derived from patient monitor data.
Competing Interests: Unrelated to this study, B.G.N. holds equity in Perimatics LLC and is its Chief Solution Architect. D.S.H. reports research grants from GE Healthcare, Philips Healthcare, Canon America Medical Systems, and Siemens Healthineers, outside this study. The remaining authors have no conflicts of interest to declare.
(Copyright © 2021 Wolters Kluwer Health, Inc. All rights reserved.)