Using ML techniques to predict extubation outcomes for patients with central nervous system injuries in the Yun-Gui Plateau.

No predictive models have been reported for tracheostomy extubation success in plateau region rehabilitation departments. Hence, the primary objective of this retrospective study was to evaluate the predictive capabilities of different models for extubation outcomes in CNS injury patients in plateau rehabilitation departments, as well as investigate the influence of clinical features on these outcomes. Data were collected from 501 adult tracheostomy patients in the Department of Rehabilitation Medicine, including 196 successful extubations. Logistic regression was employed to identify the significant features linked to extubation outcomes from a pool of 31 variables. A total of eight independent models and a weighted posterior voting ensemble model were developed. Hyperparameter optimization and tenfold cross-validation were used to assist in choosing model parameters. Random forest (ACC = 84.15, AUC = 0.85), extra trees (83.17%, 0.87), K-NN (82.18%, 0.85), and gradient boosting (81.19%, 0.85) performed well. An ensemble model (85.15%, 0.87) combining random forest, Gaussian naive Bayes, and K-NN via the WPV method was developed. Dysphagia and low GCS scores have been linked to increased difficulty in extubation, as indicated by SHAP values and previous studies. Moreover, there could be a relationship between chronic inflammation and albumin levels in patients, which may collectively impact extubation success. This study evaluated the effectiveness of conventional models for predicting extubation outcomes and analyzed the factors influencing extubation results at high altitudes, laying the groundwork for clinical use and future research. Nevertheless, further research will see advantages in using multicentric approaches and broadening clinical indicators.