PATH-03. AI-based Precision Prognostics and Therapy Personalization for Pediatric Brain Tumors

Abstract

Abstract Accurately predicting long-term outcomes in pediatric oncology is critical for balancing risk of tumor recurrence with minimizing treatment-induced developmental impacts. To address this challenge, we integrated DNA methylation profiling with artificial intelligence to develop precision prognostic models for medulloblastoma (MB) and ependymoma (EP). Leveraging genome-wide DNA methylation array datasets from over 3,000 MB and 1,000 EP patients, we developed a deep learning-based framework to predict individualized survival probabilities and treatment responses over a 10-year period post-diagnosis. Our approach employs sparse neural network models which are trained on DNA methylation and copy number variation data. Predictions of our MB model were validated using independent clinical trial data from SJMB03 (n = 305) and ACNS0331 (n = 276), achieving concordance indices of 0.79 and 0.77, respectively. These results significantly exceed the predictive accuracy of established clinical indicators as well as current risk stratification schemes. Our EP model performed similarly, obtaining a c-index of 0.74 in an independent cohort (n = 100). To advance personalized therapy, we extended our framework to predict causal outcomes of varying radiation doses for average-risk MB patients (18 Gy vs. 24 Gy). Fine-tuned with data from the randomized clinical trial ACNS0331 (n = 170), our model estimated individual treatment effects (ITEs) that describe how a given tumor profile and clinical characteristics influence treatment sensitivity. We found that 61.2% of patients had low predicted ITE (n = 104) and showed no survival disadvantage with the less intensive treatment (P = 0.683), indicating that the model effectively identified individuals who safely received reduced therapy without compromised outcomes. Among these patients, those assigned to the 18 Gy regimen maintained stable IQ during follow-up, underscoring the benefit of our approach by enabling improved outcomes with fewer long-term side effects and better quality of life.

Ähnliche Arbeiten