ESTRO 2024 - Abstract Book

S5082

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Causal inference is the process of determining the causal effect of a particular phenomenon that is a component of a larger system. In this work, we aimed, for the first time, to combine a form of causal inference used for natural experiments with a proportional hazards analysis to allow direct estimation of a dose-effect relationship. As an example, we will estimate the total causal dose-effect relationship between dose to the base of the heart and hazard of death in a large cohort of lung cancer patients.

Material/Methods:

We utilised an instrumental variable analysis. A valid instrumental variable induces changes in the exposure but has no independent effect on the outcome, allowing the causal effect of the exposure on the outcome to be estimated without distortions. Here, we used the previously discovered relationship between residual setup errors and overall survival in stage III lung cancer [2]. These residual setup errors were caused by the use of an action threshold of 5 mm at patient setup during each treatment fraction. These residual errors act as a pure instrumental variable, by randomly modifying the OAR dose independent of all other variables including the outcome of interest - overall survival (Fig. 1). In previous work, setup error data was collected for 603 lung cancer patients treated with 55 Gy in 20 fractions [2]. For each patient, the average residual setup error was calculated by averaging the daily residuals, using the nearest neighbour method to interpolate missing data. As set-up aims to align the target to its planning position, such residuals shift the dose distribution relative to the OARs. Setup errors were projected in the direction of line between the clinical target volume and the heart. The region of the heart where local dose was most associated with survival was found [3] and its mean (planned) dose calculated for all patients. We arbitrarily assumed our dose-effect relationship to follow an integrated Gaussian form, with unknown D 0.5 (dose for half of the maximum effect) and steepness (γ) between region dose and the hazard for mortality. As stated, the planned dose depends on tumour size and therefore affects survival through both tumour control and toxicity (Fig. 1). Instead, we used an instrumental variable that modifies the dose, independent of these variables, to estimate the derivative of the dose-effect relationship and find the parameters for the derivative dose-effect relationship that best fit the data. We used the Cox regression with the lowest Akaike information criterion (AIC) to select the best fitting model. Cox regression included δNTCP as the product of the residual setup error (in mm), the observed dose gradient