Background:
The pandemic outbreak of coronavirus disease (COVID-19) provokes a tangible impact on bronchoscopy examinations for the admitted burn patients due to isolation and triage measures. A machine learning approach was undertaken to discover risk factors for distinguishing mild and severe inhalation injury grade in burn patients.

Methods:
A retrospective study on 20-year (January 2000-June 2021) single-center dataset of 341 burn patients who had undergone bronchoscopy examination was established. The first admission day medical data and bronchoscopy diagnosed inhalation injury grade (ground truth) were compiled using gradient boosting-based machine learning to create an inhalation injury grade prediction model.

Results:
When grade 0 and 1 inhalation injury were classified into mild group and grade 2, 3, and 4 inhalation injury into severe group, gradient boosting algorithm-based machine learning provided a predicting model with accuracy (0.87), kappa score (0.72), sensitivity (0.8) and specificity (0.91). Using this model, we further identified five risk factors in distinguishing patients from mild and severe inhalation injury, including inhalation reminiscence, soot in upper airway, decreased oxyhemoglobin level, burn occurring in enclosed space and carbonaceous sputum, accounted for a total of 77% of prediction accuracy of the afore-mentioned model for such dichotomous grade distinction.

Conclusions:
We create the first machine learning approach for distinguishing mild and severe inhalation injury in burn patients, which is helpful when bronchoscopy examination is not quickly available. We also provide five important risk factors in assisting clinicians’ fast decision in emergency department.

Download