Research on predicting hematoma expansion in spontaneous intracerebral hemorrhage based on deep features of the VGG-19 network

Purpose To construct a clinical noncontrastive computed tomography (NCCT) deep learning joint model for predicting early hematoma expansion (HE) after cerebral hemorrhage (sICH) and evaluate its predictive performance.Methods All 254 patients with primary cerebral hemorrhage from January 2017 to December 2022 in the General Hospital of the Western Theater Command were included. According to the criteria of hematoma enlargement exceeding 33% or the volume exceeding 6 ml, the patients were divided into the HE group and the hematoma non-enlargement (NHE) group. Multiple models and the 10-fold cross-validation method were used to screen the most valuable features and model the probability of predicting HE. The area under the curve (AUC) was used to analyze the prediction efficiency of each model for HE.Results They were randomly divided into a training set of 204 cases in an 8:2 ratio and 50 cases of the test set. The clinical imaging deep feature joint model (22 features) predicted the area under the curve of HE as follows: clinical Navie Bayes model AUC 0.779, traditional radiology logistic regression (LR) model AUC 0.818, deep learning LR model AUC 0.873, and clinical NCCT deep learning multilayer perceptron model AUC 0.921.Conclusion The combined clinical imaging deep learning model has a high predictive effect for early HE in sICH patients, which is helpful for clinical individualized assessment of the risk of early HE in sICH patients. Key message What is already known on this topic: Prior to this study, existing research has highlighted various factors associated with hematoma expansion (HE) in intracerebral hemorrhage. These factors include age, Glasgow Coma Scale (GCS) score, hematoma characteristics, and time-related parameters. However, the predictive accuracy of individual factors has been limited, necessitating the development of more comprehensive prediction models. Deep learning features methods have shown promise in improving risk prediction but have not been applied to HE prediction. What this study adds: This study introduces the clinical noncontrastive computed tomography (NCCT) deep learning combined model, a novel approach that effectively predicts HE and enhances predictive accuracy. By incorporating a combination of clinical, imaging, and deep learning features, this model offers improved predictive capabilities compared to existing methods. Independent risk factors identified in this study for early hematoma enlargement include age, age group, GCS score, hematoma shape and location, midline shift, blend sign, pre-H-slicer, and time from onset to examination (H). How this study might affect research, practice or policy: The development of the clinical NCCT deep learning combined model presents a significant advancement in predicting HE, potentially aiding in the identification of high-risk patients for targeted interventions and enrollment in clinical trials. By demonstrating the value of combining clinical, imaging, and deep learning features in predicting hematoma enlargement, this study paves the way for future research to explore innovative approaches in risk assessment and management strategies for patients with intracerebral hemorrhage.