Construction of a predictive early warning model based on machine learning neural network for prognosis of patients with traumatic brain injury

Jun Li^1,2Haoyang Wang²Xiaoli Cao²Lei Sun²Can Zhu²He Li^1*

• ¹Anhui Medical University, Hefei, China
• ²The 901st Hospital of the Joint Logistics Support Force of the Chinese People's Liberation Army, Hefei, China

Objectives: The analysis of prognostic regression of patients in the regional treatment programme for severe trauma can improve the survival rate and quality of life of patients. The aim of this study is to construct an accurate and effective prognostic prediction model for the optimization and development of the regional trauma care network. Methods: We firstly extracted the clinical data of patients admitted to the regional treatment programme for severe trauma in our hospital during the period from January 2020 to December 2022. The criterion weighting method was adopted to comprehensively evaluated the AIS scores of the cumulative patients in different parts of the body. Based on the regression, the patients were divided into cured group, improved group and poor prognosis group. Based on the dependent variables, the included influencing factors were subjected to univariate analysis, multivariate analysis, and prediction model construction and comparison study. Genetic algorithm was used to solve the planning model; combined with the results of unifactorial analysis and Xgboost, RF was used to screen the features, and the interpretable model (SHAP) and column charts were used to verify the effectiveness of the screened features. Results: After feature screening and interpretable model validation, 11 indicators such as the main diagnostic score, AIS score and albumin were ultimately included as the important influencing factors of outcome variables, among which albumin was the more important protective factor, and the diagnostic score and AIS score were the more important risk factors. In the comparative study of categorical prediction models, the RF-Transformer-LSTM model achieved the most excellent prediction effect, the accuracy rate of the model test set was 0.9556, the precision rate was 0.9615, the TPR was 0.9474, the TNR was 0.9619, F1 value of 0.9544 as well as AUC value of 0.9271, and in the construction of the three-classification model, the accuracy of the model test set reached 0.9310. Conclusion: We constructed RF-Transformer-LSTM prediction model has high prediction accuracy and good interpretability in practical applications, which can provide strong support for the optimisation of regional trauma treatment strategies.