Radiomics Features on Computed Tomography Combined with Clinical Factors Predicting Hypoproteinemia in Patients with Traumatic Brain Injury

Abstract

Background Traumatic brain injury (TBI) is a major cause of death and disability in all age groups, placing a heavy burden on society and families, serum albumin levels have a significant impact on mortality and length of hospitalization patients. This study was made to develop a predictive model based on Computed Tomography (CT) and clinical parameters to explore the predictable power of the model in the development of hypoproteinemia with TBI patients. Methods A total of 72 TBI patients were prospectively recruited and confirmed as hypoproteinemia in 26 cases. A cranial CT and clinical parameters such as age, gender, admission Glasgow score were collected to establish the clinical model. The least absolute shrinkage and selection operator (LASSO) was applied to extract radiological features. Then a total of five different machine learning methods (RF, SVM, GNB, XGB, KNN) were used to establish the prediction model of radiomics. Finally, a combined model: clinical-radiological was constructed. The average area under the curve (AUC) were used to evaluate the performance of each model. Results Comparing these three different models, we found that the radiomic models combined with clinical parameters showed the best performance, which had an AUC with 0.8704 compared with clinical model only AUC = 0.8512 and radiomics model only AUC = 0.7040, respectively. Conclusions The model of radiomics features combined with clinical parameters is superior to the radiomics model and clinical model alone, and the model of radiomics combined with clinical parameters is a useful tool for predicting the occurrence of hypoproteinemia in patients with traumatic brain injury.