Medulloblastoma, the most common malignant brain tumour of childhood, is curable in roughly three-quarters of cases, but cure comes at a cost. Craniospinal irradiation and intensive chemotherapy leave survivors with lifelong cognitive, endocrine, and audiological deficits. A central goal of contemporary trials is therefore to reduce treatment intensity, particularly radiotherapy dose, in patients whose biology suggests they can safely tolerate less, while preserving the headline survival figures that have been so hard-won.

Current stratification schemes pursue this by stacking binary biomarkers to assign patients into discrete risk groups. In a disease with ~70 UK cases/year, this approach atomises an already-rare cohort into subgroups too small to power and too overlapping to disentangle, making them increasingly difficult to translate to the clinic.

I will argue that risk in medulloblastoma is better expressed as a continuous, probabilistic quantity. Using cross-validated feature selection on DNA methylation profiles from ~1,100 patients, we derived continuous molecular signatures predicting overall survival, externally validated in an independent ~770-patient clinical trials cohort. Treating predicted survival as a continuum improves discrimination, and data-driven risk binning identifies substantially larger patient populations eligible for safe de-escalation.

I will then talk about generative AI models: tabular variational autoencoders trained on our cohorts produce synthetic patients statistically indistinguishable from real data, enabling explicitly counterfactual in silico trials of radiotherapy de-escalation strategies. I will close with what this suggests for paediatric clinical trials design, where slow accrual, small numbers, and ethical constraints make conventional randomised comparison increasingly difficult to justify or complete.

Local Event Organiser: Maryam Garba
Local Group Secretary: Louis Aslett