Document resource
Background Spontaneous basal ganglia hemorrhage is associated with high morbidity and mortality. The MISTIE-III trial evaluated minimally invasive clot evacuation with thrombolysis as a strategy to improve functional recovery in patients with intracerebral hemorrhage, though treatment response varied substantially among individuals. Radiomics extract quantitative imaging features, including subtle structural and textural characteristics of hemorrhage. We sought to identify whether radiomic variables can be used to improve identification of patients with basal ganglia hemorrhage who would respond well to surgical treatment.Methods Of 305 total patients in the MISTIE-III dataset, 161 were treated surgically. Radiomics variables for surgical patients were extracted from DICOM image analyses and evaluated for clinical association and statistical significance (p<0.05). Least Absolute Shrinkage and Selection Operator (LASSO) regression was then used to identify the most informative clinical predictors. Selected radiomics features were combined with baseline clinical variables to construct predictive models of functional outcomes. Good outcome was defined as a modified Rankin Scale score of 0-3 or a Glasgow Outcome Scale - Extended score of 4-8. Model performance was assessed using confusion matrices and area under the receiver operating characteristic curve (AUC).Results A linear regression model combining baseline clinical variables - age at consent, intracerebral hemorrhage volume at diagnosis, and GCS score - predicted GOSE (AUC 0.797, 95% CI 0.741-0.843) and modified Rankin Scale scores (AUC 0.789, 95% CI 0.732 - 0.861). Features with strong correlation with surgical outcomes included distribution and density of low-intensity voxels, while features correlated with poor outcomes included hyperdense, uniform pixels at hemorrhage rims. Sequential incorporation of additional radiomics variables reached a peak AUC for favorable outcomes of 0.841 (95% CI 0.791 - 0.888) and 0.862 (95% CI 0.796 - 0.921). The model classified patients with a sensitivity of 0.682 (95% CI 0.489 - 0.747) and a specificity of 0.873 (95% CI 0.796-0.941).Conclusion When used alone, radiomics features were not particularly powerful in predicting outcomes for surgically treated patients. The use of baseline clinical variables alone produced a similar result, but the combination of clinical and radiomic variables led to a modest increase in AUC and predictive strength. Applications of this data in basal ganglia hemorrhage are limited given the small size of the MISTIE-III data set and bias toward negative outcomes in the baseline dataset. Further evaluation of these variables can be augmented with expanded use of DICOM images and larger, more diverse training and test sets of patients.Disclosures A. Katewa: None.Abstract E-312 Figure 1