Deployable machine learning-based decision support system for tracheostomy in acute burn patients

Abstract

Background Airway obstruction is a common emergency in acute burns with high mortality. Tracheostomy is the most effective method to keep patency of airway and start mechanical ventilation. However, the indication of tracheostomy is challenging and controversial. We aimed to develop and validate a deployable machine learning (ML)-based decision support system to predict the necessity of tracheostomy for acute burn patients. Methods We enrolled 1011 burn patients from Southwest Hospital (2018–2020) for model development and feature selection. The final model was validated on an independent internal cross-temporal cohort (2021, n = 274) and an external cross-institutional cohort (Second Affiliated Hospital of Zhejiang University School of Medicine 2020–2021, n = 376). To improve the model’s deployment and interpretability, an ML-based nomogram, an online calculator, and an abbreviated scale were constructed and validated. Results The optimal model was the eXtreme Gradient Boosting classifier (XGB), which achieved an AUROC of 0.973 and AUPRC of 0.879 in training dataset, and AUROCs of greater than 0.95 in both cross-temporal and cross-institutional validation. Moreover, it kept stable discriminatory ability in validation subgroups stratified by sex, age, burn area, and inhalation injury (AUROC ranging 0.903–0.990). The analysis of calibration curve, decision curve, and score distribution proved the feasibility and reliability of the ML-based nomogram, abbreviated scale, and online calculator. Conclusions The developed system has strong predictive ability and generalizability in cross-temporal and cross-institutional evaluations. The nomogram, online calculator, and abbreviated scale based on machine learning show comparable prediction performance and can be deployed in broader application scenarios, especially in resource-limited clinical environments.