From Linear-quadratic to Machine Learning in Radiobiological Modeling for Radiotherapy Planning: Resolving the Complexity-Utility Paradox with Hybrid Physics-informed Neural Networks.

Radiobiological modeling for radiotherapy planning has evolved across three overlapping paradigms: postplanning biological evaluation using empirical dose-response relationships (c. 1960), biological optimization integrated into treatment planning system objective functions (c. 2000), and artificial intelligence (AI)-driven adaptive prediction (c. 2012-present). Despite this progression, increasing model complexity has not produced proportional clinical utility - a tension we describe as the complexity-utility paradox. A structured search of PubMed, Scopus and Web of Science (January 1960 to March 2026) yielded 71 included studies; evidence is synthesized narratively with critical appraisal of landmark mechanistic models, consensus guidelines and recent meta-analytic data. We operationalize the paradox across three measurable gaps: A performance generalizability gap, in which external validation of machine-learning normal tissue complication probability (NTCP) models is approximately 9% lower in relative terms (absolute difference 0.07; pooled internal area under a receiver operating characteristic [AUROC]: 0.76, 95% confidence interval [CI]: 0.73-0.78, versus pooled external AUROC: 0.69, 95% CI: 0.64-0.73), noting that 121 of the 225 pooled models were rated at high risk of bias on PROBAST and that the internal and external estimates derive from different model sets; an adoption gap, in which Quantitative Analyses of Normal Tissue Effects in the Clinic-derived Lyman-Kutcher-Burman constraints retain primary plan-evaluation authority despite demonstrated AI improvements; and an interpretability gap, in which regulatory and clinical-trust barriers penalize opaque model architectures. We argue that principled integration, rather than paradigm replacement, offers the most tractable resolution. Hybrid physics-informed neural networks (PINNs), anchoring data-driven learning to validated mechanistic priors within a federated multi-institutional training framework and uncertainty propagation compliant with the Guide to the Expression of Uncertainty in Measurement, may narrow all three gaps simultaneously. This position is explicitly caveated: PINNs inherit the bias of the embedded prior, and current radiotherapy-specific evidence remains preliminary. To our knowledge, no externally validated PINN-based NTCP model has yet been published, so the proposal is offered as a research direction rather than as an evidenced clinical solution. A tiered validation framework is proposed for structured clinical translation.