Mdb-28. deep learning-based personalized survival prediction for medulloblastoma

Abstract Medulloblastoma is one of the most common malignant brain tumors in children and a leading cause of cancer-related death in this age group (5-year survival rate of ~70%). However, survival outcomes are heterogeneous, and currently associated with only a small number of clinical and molecular parameters. Instead, in this work we aim at exploiting the wealth of hidden information in genome-wide DNA methylation array data to predict accurate personalized survival probability curves for Group 3 and 4 medulloblastoma patients, the subgroups associated with the most heterogeneous outcomes. To accomplish this, we implemented a sparse neural network trained on DNA methylation and copy-number variation (CNV) profiles of over 900 medulloblastoma patients. Our network is designed specifically to leverage censored data for training, and makes no assumptions about the underlying distribution of the output survival probability curves. Furthermore, the architecture of our network is biologically interpretable, enabling us to probe the underlying biology driving survival. Our results demonstrate excellent discrimination and calibration, with a 5-year AUROC of 0.8. We compare our results to currently used stratification schemes on validation cohorts, including SJMB03 clinical trial data, and demonstrate significantly improved risk stratification (c-index of 0.66 vs 0.74, p < 0.0001). We find several cases in which patients assumed to be low-risk were predicted to progress rapidly, and vice versa, motivating possible refinement of risk categories in clinical practice. Our approach, which is based entirely on DNA methylation array data, may facilitate the de-escalation of therapy for low-risk patients to minimize treatment-related side effects and the intensification of treatment for high-risk individuals. The approach can be extended to other cancer types, and highlights the value of AI in healthcare towards realizing the goal of personalized medicine.