Abstract
A hallmark of glioblastoma (GBM) is disease recurrence, which occurs in all patients despite tumor resection, radiation, and chemotherapy. A critical challenge in glioblastoma treatment is the management of recurrent disease, for which there is no standard of care. Predicting the location of glioblastoma recurrence may improve the efficiency of advanced-stage therapies. Here, we present an artificial intelligence (AI)-based model to predict the risk of unprocessed surgical tissues at initial resection. AI-informed label-free optical microscopy was used to generate a normalized tumor infiltration value (AI-infiltration) for whole-slide optical images of samples taken from resection cavity margins. These values, in combination with clinical, radiographic, and molecular variables, were used to build a predictive model of focal recurrence. In a cohort of 80 patients, comprising 367 samples and 133,454 unique images, glioblastoma infiltration was significantly higher in margin samples from recurrent tumors (p = 0.026) compared with those from non-recurrent tumors. A random forest (RF) machine learning classifier was able to predict site recurrence with an average area under the receiver operating characteristic curve (AUC) of 86.6% ± 10.0 for the training cohort and 80.3% (95% CI: 0.641-0.965) for the validation cohort. AI-infiltration was the strongest contributor to recurrence prediction, outperforming tumor molecular features. Model performance remained high regardless of tumor location, resulting in random forest model predictions of recurrence at 5 and 10 millimeters of each sampled site. These findings represent the potential of AI to predict sites of tumor recurrence, thereby improving accessibility to targeted, precision, multimodal therapy for the highest-risk areas of disease. One Sentence Summary: Machine learning estimates of tumor infiltration predict focal glioblastoma recurrence.