Data-driven evaluation of quantitative metrics and models for normal-tissue sparing in FLASH radiotherapy

Objective: Preclinical studies demonstrated that ultrahigh dose rate (UHDR) irradiation can spare normal tissue compared to conventional dose rate (CDR) irradiation while maintaining tumor control. However, the conditions and magnitude of that FLASH sparing remain uncertain, limiting clinical translation. Most proposed metrics and models for predicting FLASH sparing lack rigorous experimental validation. This study systematically benchmarks them against extensive in vivo data.Approach: Normal tissue toxicity electron and proton data comparing CDR and UHDR irradiations were compiled from murine skin, gut, and zebrafish studies. The dataset encompassed single-fraction, fractionated, dose-rate scan, and split-dose experiments.Isoeffective dose ratios were derived, referred to as FLASH-modifying factors (FMF).Metrics exhibiting the strongest correlations with experimental data were identified and used to model FMF via empirical transition functions. Model performance was compared for two metric-based models established in this work and proposed models from literature, including the FLASH effectiveness and equilibrium state dynamics (ESD) models.Main results: FMF correlated stronger with fractional dose (Spearman's ρ ≈ -0.64) than with total dose (ρ ≈ -0.11), following a piecewise-linear dependence in terms of FMF-weighted dose versus dose. Among the investigated temporal dose delivery metrics, mean and pencil beam scanning (PBS) dose rates showed the strongest correlations (ρ ≈ -0.75 to -0.80) and the associated FMF values were well captured by exponential relationships. Metrics such as dose-averaged dose rate and UHDR-contribution did not account for loss of FLASH sparing due to delivery pauses. The ESD model provided a more accurate description of FMF than metric-based models or the FLASH-effectiveness model.Significance: Among the assessed metrics, fractional dose combined with mean or PBS dose rate best captured FLASH-induced normal-tissue sparing, while ESD models provided the highest descriptive accuracy across all evaluated datasets. Validating FLASH metrics and models is essential to standardize FLASH treatment planning, inform delivery system design, and enable optimized clinical translation.