Research on Machine Learning Models Based on Cranial CT Scan for Assessing Prognosis of Emergency Brain Injury

To evaluate the prognosis of patients with traumatic brain injury according to the computed tomography (CT) findings of skull fracture and cerebral parenchymal hemorrhage.

We retrospectively collected data from adult patients who received non-surgical or surgical treatment after the first CT scan with craniocerebral injuries from January 2020 to August 2021. The radiomics features were extracted by PyRadiomics. Dimensionality reduction was then performed using the max relevance and min-redundancy algorithm (mRMR) and the least absolute shrinkage and selection operator (LASSO), with 10-fold cross-validation to select the best radiomics features. Three parsimonious machine learning classifiers, multinomial logistic regression, a support vector machine (SVM), and a naive Bayes (Gaussian distribution) classifier, were used to construct radiomics models. A personalized emergency prognostic nomogram for cranial injuries was created using a logistic regression model based on selected radiomic labels and patients' baseline information at emergency admission.

The mRMR algorithm and the LASSO regression model extracted 22 top-ranked radiologic features and based on these image histologic features, the emergency brain injury prediction model was built with SVM, logistic regression, and naive Bayes classifiers, respectively. The SVM model showed the largest area under the receiver operating characteristic curve in training cohort for the 3 classifications, indicating that the SVM model is more stable and accurate. Moreover, a nomogram prediction model for Glasgow Outcome Scale prognostic score in patients was constructed.

We established a nomogram for predicting patients' prognosis through radiomic features and clinical characteristics, which provides some data support and guidance for clinical prediction of patients' brain injury prognosis and intervention.