Application Of Ct-based Machine Learning And Radiomics In Prognosis Assessment Of Spontaneous Intracerebral Hemorrhage

Objective: To improve the BAT score using machine learning algorithms to predict early hematoma expansion(HE)after spontaneous intracerebral hemorrhage(s ICH).Methods: The clinical and imaging data of 232 s ICH patients who were admitted in our hospital from May 2015 to May 2020 were analyzed.Early HE was defined as an increase in hematoma volume >6 m L or 33%on the follow-up CT images within 24 hours compared to baseline CT images.The BAT score of each patient was calculated according to the algorithm of the original BAT score,and its performance in identifying patients with early HE was evaluated.Then,all s ICH patients were randomly divided into a training subset(n=162)and a validation subset(n=70)in a ratio of 7:3.In the training subset,based on the three variables of the original BAT score(blend sign,any hypodensity,and time from onset to CT examination),five common machine learning algorithms(random forest,gradient boosting,Naive Bayes [NB],logistic regression and k-nearest neighbors)were used to construct the modified BAT scores.Subsequently,the modified BAT scores were independently validated in the validation subset.Receiver operating characteristic(ROC)curves were plotted to evaluate the discriminative ability of all BAT scores.The De Long test was used to compare the performances of all BAT scores.Decision curve analysis(DCA)was performed to evaluate the clinical usefulness of all modified BAT scores.Results: The area under the ROC curve(AUC)of the original BAT score for predicting early HE was 0.57.Among the five modified BAT scores,the modified BAT score based on NB algorithm performed the best,with the AUC of 0.83 in the training subset and 0.77 in the validation subset respectively.The De Long test showed that the prediction performance of the modified BAT score based on NB algorithm was significantly better than the original BAT score(AUC=0.57)in both training and validation subsets(both P<0.001).The DCA curves showed that the modified BAT score based on NB algorithm was better than other four modified BAT scores in predicting early HE.Conclusion: Machine learning algorithms using the same variables could improve the ability of the original BAT score.For predicting early HE after s ICH,the modified BAT score based on NB algorithm could be used as an effective tool to identify patients at high risk of HE.Objective: To construct a model for predicting early HE after s ICH by combining clinical risk factors and CT-based radiomics features of cerebral hematoma,and to evaluate its predictive performance.Methods: The clinical and imaging data of 339 patients with s ICH who were admitted to our hospital from April 2014 to September 2020 were analyzed.Early HE was defined as an increase in hematoma volume >6 m L or 33% on the follow-up CT images within 24 hours compared to baseline CT images.The radiomics features of cerebral hematoma were extracted from the patients’ baseline plain CT images,and the optimal radiomics features were obtained through feature screening.Then a Rad-score model was constructed using the optimal radiomics features.Independent risk factors associated with HE were detected by using univariate and multivariate analyses,and were used to construct a clinical model.The clinical-radiomics model was constructed by combining clinical independent risk factors and Rad-score,and a nomogram was formulated based on the optimal prediction model.The ROC curve and calibration curve were plotted to evaluate the predictive performances of these models.Results: Univariate analysis showed that CT time from s ICH onset,diabetes history,Glasgow coma scale(GCS)score,platelet count and hematoma volume were significantly different between the non-HE group and the HE group(P<0.05).Multivariate analysis showed that CT time from s ICH onset(odds ratio [OR]=0.855;P=0.032),diabetes history(OR=0.522;P=0.014),GCS score(OR=0.914;P=0.039)and hematoma volume(OR=1.015;P=0.046)were independent risk factors for early HE.After feature screening,20 optimal radiomics features related to early HE were obtained.The AUC of the clinical-radiomics model(0.870)for predicting HE was higher than the clinical model(0.650)and Rad-score(0.860).The clinical-radiomics model predicted the probability of early HE was in good agreement with the actual probability of early HE.Conclusion: The clinical-radiomics model based on clinical data and CT radiomics features at admission can effectively predict early HE in patients with s ICH,and its predictive performance is better than the clinical model.The nomogram of clinical-radiomics model has great potential for clinical application in the risk assessment of early HE.Objective: To construct and validate location-specific Rad-score and clinical-radiomics models based on plain CT images for predicting poor functional outcome at 6 months after deep and lobar s ICH.Methods: Totally 494 patients with s ICH between January 2014 and January 2021 were retrospectively reviewed.Poor functional outcome of a patient with s ICH was defined as a modified Rankin scale score >2 at6 months after the onset of s ICH.Firstly,the radiomics features of cerebral hematoma were extracted from the region of interest of hematoma in baseline plain CT images,and then the optimal radiomics features related to poor functional outcome were selected through feature screening.Finally,the Rad-score was constructed using the optimal radiomics features.ROC curve was plotted to assess the predictive power of Rad-score for poor functional outcome at 6 months after deep and lobar s ICH in the derivation and validation cohorts,respectively.The optimal location-specific Rad-score cut-offs for predicting poor functional outcome after deep and lobar s ICH were determined.The Rad-scores were converted into the binary values according to cut-offs for deep s ICH and lobar s ICH,respectively.In the derivation cohort,univariable and multivariable analyses were applied to identify independent predictors associated with poor functional outcome.The location-specific clinical-radiological models and clinical-radiomics models for deep and lobar s ICH were constructed in the derivation cohort and validated in the validation cohort,respectively.The AUC,sensitivity and specificity were calculated for each predictive model.Results: Of the 494 s ICH patients,392(79%)patients had deep s ICH,and 373 patients(76%)had poor functional outcome.The GCS score,hematoma volume,hematoma location,hematoma expansion and Rad-score were determined as independent predictors of poor functional outcome(all P<0.05).A Rad-score of 82.90(AUC=0.794)for deep s ICH and a Rad-score of 80.77(AUC=0.823)for lobar s ICH were determined as cut-offs for predicting poor functional outcome.For deep s ICH,the AUCs of the clinical-radiomics model were 0.856 in the derivation cohort and0.831 in the validation cohort,respectively.For lobar s ICH,the AUCs of the clinical-radiomics model were 0.866 in the derivation cohort and 0.843 in the validation cohort,respectively.Conclusion: The location-specific Rad-scores and clinical-radiomics models based on plain CT images can identify patients at high risk of poor functional outcome after deep and lobar s ICH,and the location-specific Rad-scores could be considered as an inclusion criterion for screening target populations to help improve clinical trial design related to the evaluation of the therapeutic efficacy of s ICH.