ESTRO 2024 - Abstract Book

S5083

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

(~1 Gy/mm) and the derivative of the dose-effect relationship to be estimated (in Gy -1 ). Results are summarized as a heat map of AIC values for varying D 0.5 and γ (Fig. 2). Bootstrapping with 100 resample steps was used to evaluate model stability.

Results:

Residual setup errors ranged between +5 and -5 mm and were not correlated with age, tumour size, performance status or comorbidities [2]. Median follow-up time was 15 months. The lowest AIC was obtained for a D 0.5 of 19Gy and γ of 2.1Gy. Cox models including δNTCP (beta = 0.09Gy -1 , p=0.02) outperformed those without (p=0.01). Bootstrap analysis showed the D 0.5 to be stable with 50% of bootstraps finding a D 0.5 between 17 and 21Gy. The estimate of γ was unstable, with 75% below 2Gy. This is likely due to low number of events contributing to each bootstrap, less than 50 patients contribute to the fit due to the steep dose-effect relationship (Fig. 2).

Conclusion:

A full OAR dose-response curve has been derived for the first time using causal inference with an instrumental variable approach. This allows complex outcomes to be modelled, here from random perturbations of the dose. Bootstrapping shows the estimate of D 0.5 for the base of the heart to be stable around 19Gy, but the steepness γ is uncertain, with a tendency to be quite steep (below 2Gy).

Keywords: Dose-response, causal inference, heart toxicity

References: