Prognostic prediction of hypertensive intracerebral hemorrhage using CT radiomics and machine learning

Objectives Spontaneous intracerebral hemorrhage remains a major cause of death and disability throughout the world. We tried to establish accurate long‐term outcome prediction models for hypertensive intracerebral hemorrhage (HICH) using CT radiomics and machine learning. Methods In a retrospective study of 270 patients with HICH between June 2013 and June 2018, CT images and patients' 6‐month outcome based on the modified Rankin Scale were collected. Hematomas on CT images were selected as volumes of interests (VOIs), and 1,029 radiomics features of the VOIs were extracted. Based on correlations with patients' outcome, radiomics features underwent dimensionality reduction analyses. Then, the support vector machine (SVM), k‐nearest neighbor (KNN), logistic regression (LR), decision tree (DT), random forest (RF), and XGBoost algorithms were applied with the screened features to establish prognostic prediction models of HICH. Accuracies of all models were compared. Results Eighteen radiomics features were screened as prognosis‐associated radiomics signature of HICH based on the variance threshold, SelectKBest, and least absolute shrinkage and selection operator (LASSO) regression models. Patients were randomly allocated into training (n = 215) and validation (n = 55) sets. Accuracies of all 6 machine learning algorithms in the validation set exceeded 80%. The sensitivity, specificity, and accuracy in the validation set were 93.3%, 92.5%, and 92.7% for the RF model and 92.3%, 88.1%, and 89.1% for the XGBoost model, respectively, which were the best two among all models. Conclusions Taking advantage of radiomics and machine learning, we established accurate prognostic prediction models of HICH. The RF model and XGBoost model returned the best accuracies. We tried to establish accurate long‐term outcome predictions models for hypertensive ICH using CT radiomics and machine learning. Six machine learning algorithms were used at the same time, and the accuracies of different models were compared and selected. Finally, the random forest algorithm model demonstrated a sensitivity of 93.3%, a specificity of 92.5% and the area under the curve was 0.92, which were the best among all 6 models.