A Machine Learning-based Approach To Building Predictive Models For The Field Of Traumatic Brain Injury

Objective Currently,the treatment and care of adult and pediatric traumatic brain injury(TBI)is an intractable global health problem.Predicting the prognosis and length of stay of traumatic brain injury patients can lead to improved clinical outcomes and achieve precision medicine,thus significantly reducing the health care burden on society.With the flourishing of computer technology in modern society,it is possible to apply new machine learning methods to the field of traumatic brain injury(TBI),and studies have shown that computer methods can be used to predict patient-related clinical outcomes with high accuracy,which may be valuable for determining the prognosis and cost-effectiveness of clinical treatment.value.In this study,we collected and retrospectively analyzed the clinical data of 1112 patients with a clinical diagnosis of TBI,and combined various machine-learning(ML)methods and the results of domestic and international studies to develop a hybrid model for predicting the prognosis and hospitalization time of adults and children with TBI.Methods We collected clinical information related to patients diagnosed with TBI at the Second Affiliated Hospital of Anhui Medical University Neurosurgery Center between May 2017 and May 2022 by screening and random segmentation,of which80%were used to train the model and 20%were used to test the model.For the two prediction results we compared and constructed four machine learning models that are commonly used and have high accuracy in current domestic and international studies,and the models were repeatedly trained and tested using a five-times cross-validation method to avoid over-fitting.In the machine learning models,11 independent variables were used as input variables,and the GOS scores used to assess patient prognosis and the length of stay of patients were used as output variables.After the models were trained,we obtained the errors of each machine learning model using a five-round cross-validation method and compared them to select the best prediction model.Finally,to further validate our experimental results,we collected screened and externally tested all models using clinical data from patients with TBI treated at the First Affiliated Hospital of Anhui Medical University Neurosurgery Center between June 2021 and February 2022.Results The accuracy(ACC)of the final convolutional neural network-support vector machine(CNN-SVM)hybrid model to predict the GOS classification outcome as prognosis of TBI patients was 93%and 93.69%in the test set and external validation set,respectively.93%and 93.69%,and the area under curve(AUC)was 94.68%and94.32%in the test set and external validation set,respectively.The mean absolute percentage error(MAPE)of the final convolutional neural network-support vector regression(CNN-SVR)hybrid model for predicting the length of stay of TBI patients in the test set and the external validation set was 94.68%and 94.32%,respectively.The coefficient of determination(R~2)of this hybrid prediction model is 0.93 and 0.92 in the test and external validation sets,respectively.The results are optimal and have a high confidence level.Conclusions This study demonstrates the clinical utility of two hybrid models built by combining multiple novel machine learning methods that can be used to accurately predict prognosis and hospital length of stay for adult and pediatric traumatic craniocerebral injuries.The application of these models can reduce the burden on physicians when evaluating patients with traumatic brain injury,assist clinicians in medical decision making,and provide some reference for related researchers.