Session Item

Radiation Pneumonitis (RP) is still a major complication for non-small cell lung cancer (NSCLC) patients. Traditionally, the mean lung dose (MLD) and the volume of the lung receiving at least 20 Gy (V_20Gy) are used to predict RP. However, Hoffmann and Nahum (2013) proposed to use the underlying actual dose-distribution of the lungs, and Ghobadi et al. (2012) found the dose to the heart to also contribute to pulmonary toxicities. Therefore, we investigated the use of the standard deviation of the dose distribution, the effective α/β parameter as well as the Biologically Effective Dose (BED) of the lungs and the heart as predictive parameters for RP in NSCLC patients.

Data on the occurrence of RP and dose-volume parameters of 96 patients diagnosed with NSCLC and treated with passively-scattered proton beam therapy was retrospectively retrieved from MD Anderson Cancer Center, Houston, TX, (Gjyshi et al., 2021) and the University Hospital Carl Gustav Carus Dresden, Germany, (Zschaek et al., 2016). Data was randomly split into a training-set (64 patients) and a validation-set (32 patients). The eff. α/β parameters were calculated using the equation derived by Hoffmann and Nahum. Statistical analyses were performed with R version 4.0.1 using binomial logistic regression models as well as bootstrapping for these models. For multivariate logistic regression parameters with a high mutual Pearson Correlation (>0.5) were represented by the most important parameter to avoid collinearities. Comparison of models was performed using a 2-sided test comparing the differences between the AUC-distributions of the bootstrapped models. To find out if the fraction size has an impact on our results we recalculated the 33x2Gy fractionation scheme of the patient-plans of the second patient group using a 24x2.75Gy fractionation scheme and compared to the original plans using the AUC-distributions of the bootstrapped models.

The BED of the lung significantly predicted RP Grade ≥ 2 in a univariate model in training (p=0.019, AUC_Train=0.72;Fig. 1A) and in a multivariate model in combination with the eff. α/β parameter of the heart (p_BED=0.003, p_{α/β_eff}=0.028, AUC_Train=0.78). These results did not hold in validation (AUC_Val=0.52 and AUC_Val=0.43, respectively). Also in training, these models were not significantly better in predicting RP Grade ≥ 2 (p=0.96 for univariate models, p=0.99 for multivariate models) than a model built with the MLD (p=0.015, AUC_Train=0.73, AUC_Val=0.51;Fig. 1B) and a multivariate model additionally including the V_20Gy of the heart (p_MLD=0.039, p_HeartV20Gy=0.58, AUC_Train=0.74, AUC_Val=0.53). The performance of the model was insensitive to the underlying fractionation scheme (33x2Gy vs. 24x2.75Gy: p=0.923 for BED, p=0.926 for MLD).

Structural parameters of the dose distribution of the lungs and heart did not outperform commonly used factors, such as MLD, in predicting RP Grade ≥ 2. Pulmonary and cardiac performance indicators may be added to future analyses.