Machine learning–based risk stratification for early tracheotomy in geriatric intracerebral hemorrhage: model development and web deployment

Background Respiratory complications are common among elderly patients with intracerebral hemorrhage (ICH), which can cause more hypoxemia and neuronal injury, further leading to poor prognosis. Tracheotomy (TT) can effectively improve airway protection and oxygenation, but the timing of tracheotomy is usually determined by the experience of the attending clinician and lacks objective and standardized decision criteria. A reliable prediction tool for the early identification of elderly ICH patients who need TT is urgently needed. This study aimed to establish and validate a machine learning (ML) model to predict the early application of tracheotomy in elderly ICH patients. Methods A total of 435 elderly patients (≥ 65 years) with ICH from two tertiary hospitals between 2015 and 2023 were retrospectively reviewed. A total of 40 clinical, laboratory, and imaging features were collected at admission. We applied LASSO regression for feature selection. Six machine learning algorithms (Random Forest, XGBoost, SVM, Gradient Boosting, Logistic Regression, and KNN) were trained and tested with a train–test split of 70/30. The imbalanced class distribution was handled by SMOTE and class-weight adjustments. Hyperparameter optimization was conducted with Bayesian optimization. The performance of each algorithm was evaluated based on ROC–AUC, PR curves, accuracy, precision, recall, and F1 score. SHAP values were calculated to interpret the models. The best-performing model was deployed as an online prediction tool. Results LASSO identified five predictors: Glasgow Coma Scale (GCS) score, intraventricular hemorrhage (IVH), midline shift, albumin level, and CT biphasic sign. Among the six algorithms, Random Forest performed the best (composite score 0.831; test AUC 0.723; accuracy 0.793; precision 0.784; recall 0.793; F1 = 0.788). SHAP analysis identified that GCS was the most important feature that contributed to tracheotomy risk, followed by IVH and midline shift. An online prediction tool was successfully deployed to estimate the early risk of tracheotomy in real time. Conclusions A ML–based model with five readily available clinical features accurately predicted the early need for TT in elderly ICH patients and may be helpful in timely airway management decisions.