| Objective:This study aims to establish a model based on CT radiomics to predict the progression of Acute Intracerebral Hemorrhage(ICH)and verify the model.Methods:The imaging and clinical data of 289 patients diagnosed with ICH were obtained retrospectively from Guizhou Medical University Affiliated Hospital from October,2019 to October,2021.The clinical functional outcome of ICH patients was assessed using the modified Rankin Scale(m RS)at discharge after onset through Neurological examination.Referring to previous studies,a m RS score ≤ 3 was considered a favorable outcome,while a m RS score of > 3 was considered a poor outcome.Patients were randomly divided into a training cohort(n = 202)and an internal validation cohort(n = 87)at a 7:3 ratio by computer software generated random numbers,which were used for model construction and internal verification,respectively.The validation set cases are used to verify the effectiveness of the model,and 163 ICH patients in the same period from Medical University Affiliated Hospital are collected for external validation set.Collect the first head CT when the patient is admitted to the hospital,outline the image and obtain the Region of interest(ROI),use the obtained ROI for feature extraction,and use the Lasso(Lesat absolute shrinkage and seletion operator,LASSO)regression to reduce features.The optimal parameters were debugged by 5 times cross validation method and combine them with Logistic Regression Naive Bayesian Classification(NB)XGBoost algorithm Random Forest(RF)Ada algorithm 5 machine algorithms to establish the prediction model,compare the accuracy of all models.Radscore is calculated according to the selected features and their coefficients.Then,Radscore combined with the independent predictors of ICH patients’ short-term prognosis,a comprehensive imaging and clinical prediction model was constructed by multivariate Logistic regression analysis of the training cohort and visualized by drawing a nomogram.We can analyze the accuracy,Positive predictive value(PPV),Negative predictive value(NPV),sensitivity,specificity and F1 score of the model,and finally establish a nomogram to describe the model.Results:A total of 452 patients were included in the model development dataset and external validation set,of which 237(52.4%)had poor short-term outcomes.3 radiomics features were screened by LASSO regression method for dimension reduction.The 3 radiomics features were input into five machine algorithms to build the prediction model and draw the ROC curve.The AUC of them was 0.73 0.80 0.80 0.83 0.81 respectively.The nomogram were composed of six independent predictors:GCS,gender,breaking into ventricle,midline shift,gastrointestinal hemorrhage and Radscore.The nomogram display AUC values of training set and internal validation set were: 0.88,0.86,PPV,NPV,sensitivity,specificity and F1 score were 0.75,0.89,0.88,0.77,0.82 and 0.63,0.67,0.60,0.70,0.68,respectively.Through external verification,the AUC value was 0.86,and the sensitivity,specificity and F1 score were 0.85,0.68,0.62,0.88 and 0.77,respectively.Calibration curves showed that clinical radiographs showed satisfactory calibration in both the training team and the external test cohort,with less consistency in the internal test cohort.The clinical decision curve showed that the clinical benefit rate for predicting prognosis was higher in clinical radiographs than in clinical models when the degree of risk was 0.4 to 0.6.Conclusion:1.The nomogram is an effective computer-aided tool to provide personalized risk assessment of short-term prognosis for ICH patients.2.Combine image omics and machine algorithm to construct prediction model,among which RF algorithm model has the best performance.3.Our study showed that when the early clinical risk of ICH patients was 0.4-0.6,clinical radiological rograms could better assist clinical assessment of patients’ short-term prognosis. |
PDF Full Text Request