SURG-86. Predicting Postoperative Functional Status in Glioblastoma Using Diffusion Tensor Imaging, Disconnectomics, and Deep Learning

INTRODUCTION Surgical resection is a cornerstone of glioblastoma treatment but often poses risks of new or worsened functional deficits. Accurate prediction of postoperative functional status remains an unmet clinical need, limiting informed surgical planning and patient counseling. While prior studies have incorporated machine learning, structural MRI, and functional MRI with moderate success, none have explored the utility of white matter measures and brain network disconnectivity, which are factors critical to understanding functional outcomes. This study sought to develop a machine learning model that utilizes diffusion tensor imaging (DTI) and disconnectomics to predict postoperative functional outcomes in glioblastoma patients. METHODS We retrospectively analyzed 106 patients with glioblastoma who underwent surgical resection from 2008 to 2022 with preoperative DTI and structural MRI available. Preoperative clinical variables collected included age, sex, Karnosfky performance status (KPS), presenting symptoms, and neurological deficits. Postoperative variables included KPS. Tumor segmentation was performed automatically using HD-GLIO tool, a neural network based automatic tumor segmenter. The imaging platform 3D slicer was used for DTI estimation and tractography measurement extraction. BCBToolKit was used for disconnectomics. Clinical variables, DTI, and disconnectomics were used to train two random forest classifiers; one (numeric model) to predict postoperative KPS (≥70, <70), and another (deterioration model) KPS worsening versus non-worsening (deterioration model). RESULTS The numeric model random forest classifier demonstrated 75% accuracy in predicting postoperative KPS, while the deterioration model had an accuracy of 89%. The most important features for both models were pre-operative KPS score, presenting with motor deficit, tumor location, and tractography measurements. CONCLUSION Integrating DTI and disconnectomic features into a deep learning framework enables accurate prediction of postoperative functional deterioration in glioblastoma. This approach offers a promising tool for surgical risk stratification and individualized patient counseling.