Utilization of non-invasive ventilation before prehospital emergency anesthesia in trauma - a cohort analysis with machine learning.

Abstract

Background: For preoxygenation, German guidelines consider non-invasive ventilation (NIV) as a possible method in prehospital trauma care in the absence of aspiration, severe head or face injuries, unconsciousness, or patient non-compliance. As data on the utilization and characteristics of patients receiving NIV are lacking, this study aims to identify predictors of NIV usage in trauma patients using machine learning and compare these findings with the current national guideline.

Methods: A cross-regional registry of prehospital emergency services in southwestern Germany was searched for cases of emergency anesthesia in multiply injured patients in the period from 2018 to 2020. Initial vital signs, oxygen saturation, respiratory rate, heart rate, systolic blood pressure, Glasgow Coma Scale (GCS), injury pattern, shock index and age were examined using logistic regression. A decision tree algorithm was then applied in parallel to reduce the number of attributes, which were subsequently tested in several machine learning algorithms to predict the usage of NIV before the induction of anesthesia.

Results: Of 992 patients with emergency anesthesia, 333 received NIV (34%). Attributes with a statistically significant influence (p < 0.05) in favour of NIV were bronchial spasm (odds ratio (OR) 119.75), dyspnea/cyanosis (OR 2.28), moderate and severe head injury (both OR 3.37) and the respiratory rate (OR 1.07). Main splitting points in the initial decision tree included auscultation (rhonchus and bronchial spasm), respiratory rate, heart rate, age, oxygen saturation and head injury with moderate head injury being more frequent in the NIV group (23% vs. 12%, p < 0.01). The rates of aspiration and the level of consciousness were equal in both groups (0.01% and median GCS 15, both p > 0.05). The prediction accuracy for NIV usage was high for all algorithms, except for multilayer perceptron and logistic regression. For instance, a Bayes Network yielded an AUC-ROC of 0.96 (95% CI, 0.95-0.96) and PRC-areas of 0.96 [0.96-0.96] for predicting and 0.95 [0.95-0.96] for excluding NIV usage.

Conclusions: Machine learning demonstrated an excellent categorizability of the cohort using only a few selected attributes. Injured patients without severe head injury who presented with dyspnea, cyanosis, or bronchial spasm were regularly preoxygenated with NIV, indicating a common prehospital practice. This usage appears to be in accordance with current German clinical guidelines. Further research should focus on other aspects of the decision making like airway anatomy and investigate the impact of preoxygenation with NIV in prehospital trauma care on relevant outcome parameters, as the current evidence level is limited.