Objective To develop and identify the optimal machine learning model for predicting the short-term prognosis of comatose patients with severe traumatic brain injury （sTBI）.Methods A retrospective analysis was conducted on 262 comatose patients with sTBI admitted between January 1， 2022， and December 1， 2024. Based on the Glasgow Outcome Scale （GOS） score at the 6-month follow-up， patients were divided into a favorable prognosis group （GOS≥4， n=112） and an unfavorable prognosis group （GOS ≤3， n=150）. Prediction models were constructed using clinical data collected at admission with seven machine learning algorithms： logistic regression， support vector machine， multivariate adaptive regression splines， random forest， k-nearest neighbors， extreme gradient boosting， and artificial neural network （ANN）. Model performance was evaluated via 10-fold cross-validation， with the area under the receiver operating characteristic curve （AUC） serving as the primary evaluation metric. Accuracy， sensitivity， and specificity were also compared.Results Among the seven models， the ANN model demonstrated superior prediction performance， achieving an AUC of 0.905. The AUCs for the remaining models ranged from 0.581 to 0.760. The ANN model also exhibited the highest accuracy， sensitivity， and specificity. Further validation confirmed the stability of the ANN model， as indicated by the consistent trends of the loss function， mean squared error， and accuracy over 100 iterations on both the training and test sets. Analysis of key prognostic factors identified the Glasgow Coma Scale score， bispectral index， and partial pressure of oxygen as the most significant predictors in the model.Conclusions The prediction model based on the ANN algorithm demonstrates excellent performance in predicting the short-term prognosis of comatose patients with sTBI， outperforming other traditional machine learning models. This model may provide a valuable reference for clinical assessment of disease severity and prognosis.