ESTRO Course 2026 - Quantitative methods in Radiation Oncology: "Bayesian statistics – the doctors view"
Year: 2026
Type: article
Abstract: Delivered 2026-04-20, Lisbon, Portugal, Faculdade de Ciências da Universidade de Lisboa, Campo Grande. Precis: This invited presentation, delivered at the 2026 QMRO symposium, offers a clinician-oriented introduction to Bayesian statistical reasoning and its growing practical relevance in radiation oncology and medical physics. Aimed explicitly at non-statistician audiences, the talk translates foundational Bayesian concepts into accessible clinical frames of reference without sacrificing mathematical rigor.The lecture opens by grounding attendees in Bayes' theorem — the formal relationship between prior beliefs, observed data (likelihood), and updated posterior probabilities — and contrasts this paradigm with frequentist inference. Where frequentists treat model parameters as fixed and data as random (seeking the parameter value maximizing likelihood under repeated sampling), Bayesians treat parameters as random variables updated by data, permitting direct probability statements about hypotheses. This distinction has immediate practical consequences: Bayesian credible intervals yield the intuitive interpretation that clinicians often (incorrectly) apply to frequentist confidence intervals, and Bayes Factors provide direct evidence ratios between competing hypotheses rather than the indirect p-value statement of data probability under the null.A core focus is the taxonomy of prior distributions — informative, weakly informative, and non-informative (diffuse) priors — with worked examples illustrating how sequentially updating priors from pilot data or historical studies mirrors the actual knowledge accumulation process in clinical research. The talk then addresses one of the most practically powerful applications for radiation oncology: Bayesian data borrowing. Drawing on the FDA's January 2026 Bayesian Methodology Guidance, the presentation systematizes four borrowing archetypes — prior trials of the same intervention, external or non-concurrent controls, pediatric extrapolation via, and cross-disease/subsite pooling through Bayesian hierarchical models (BHM) — each paired with radiation oncology analogues. Adaptive Bayesian trial design architectures are reviewed comprehensively: sequential monitoring, sample-size re-estimation, response-adaptive randomization, adaptive enrichment, and seamless phase transitions. The lecture closes with practical guidance on when Bayesian approaches offer the greatest yield across trial phases, and with personal observations from leading a Bayesian adaptive trial in a rare radiation toxicity population. The central message is that Bayesian methods are not merely a statistician's tool: they are conceptually more intuitive for clinical decision-making, increasingly supported by regulatory guidance, and directly applicable to the adaptive, limited-sample, borrowing-rich research environment that characterizes modern radiation oncology.
Source: Figshare
Author: Fuller, Clifton D.
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7154995798
FWCI:
Cites:
Cited by:
Related to: