ID #860 AI-based Precision Prognostics and Therapy Personalization for Childhood Brain Tumors
Sabina Stefan, Amelie Amann, Kyle Smith, Fabio Boniolo, Estelle Pfitzer, Qian Li, Giles Robinson, Patrick Harter, Bill Lotter, Paul Northcott, Volker HovestadtAbstract
Background
Accurate prediction of long-term patient outcomes remains a critical challenge in oncology, in particular for childhood brain tumors such as medulloblastoma and ependymoma where therapeutic strategies need to carefully balance the risk of tumor recurrence and death against severe treatment-induced sequelae.
Methods
To address this challenge, we integrated genome-wide DNA methylation and copy-number profiles with time-to-event survival data from 2,540 patients to develop MANTIS (Methylation profiling and Artificial Neural networks for Time-resolved Individualized Survival predictions), a precision AI-based framework for medulloblastoma and ependymoma outcome prediction. Our approach utilizes sparse neural network models to generate individualized survival probabilities over a 10-year period. To move from prognosis to treatment recommendations, we leveraged large-scale pretraining to learn compact tumor representations from thousands of medulloblastoma methylation profiles. These representations enabled counterfactual survival prediction under varying craniospinal irradiation doses in the randomized ACNS0331 trial.
Results
In independent validation cohorts from clinical trials, MANTIS achieved concordance indices up to 0.80 ± 0.07 for medulloblastoma and 0.75 ± 0.09 for ependymoma, exceeding the predictive accuracy of current clinical indicators. In ACNS0331, MANTIS identified a large patient subset predicted to derive little incremental survival benefit from standard-dose versus low-dose craniospinal irradiation (∼60% of patients). In this subset, low-dose assignment was not associated with inferior survival (P = 0.683), while neurocognitive outcomes were significantly improved, with higher IQ scores than in standard-dose assignment (P = 0.037).
Conclusions
MANTIS delivers patient-specific, time-resolved survival estimates from routine molecular profiling and supports treatment-intensity stratification by identifying individuals likely to be appropriate candidates for radiation de-escalation, aiming to preserve survival while improving neurocognitive outcomes. This work provides a scalable framework for precision prognostics and therapy personalization in pediatric brain tumors and establishes a template for similar approaches across oncology.