Receiver operating characteristic curves to assess predictors of radiation-induced symptomatic lung injury

Purpose: To assess the utility of dosimetric/functional metrics as predictors of symptomatic radiation pneumonitis using receiver operating characteristic curves. Methods: Between 1991 and 1999, 277 patients were enrolled on a prospective clinical study to relate radiation therapy (RT) induced changes in lung function with dosimetric and functional metrics. Pre-RT whole and regional functional assessments included pulmonary function tests and single photon emission computed tomography lung perfusion scans. Patients had three-dimensional planning scans and dose calculations (reflecting tissue density heterogeneity) to provide a dose-volume histogram of the lung and associated dosimetric parameters (MLD = mean lung dose, V30 = % of lung receiving greater than or equal to30 Gy). Fusion of single photon emission computed tomography and computed tomography scans provides perfusion-weighted dose-function histograms and associated dosimetric parameters (mean perfusion-weighted lung dose). The incidence of clinically relevant radiation pneumonitis requiring steroids was related to the dosimetric and functional metrics. The predictive abilities of models (sensitivity and specificity) were calculated and compared based on the area beneath receiver operating characteristic (ROC) curves (Wilcoxon rank-sum and chi-square). Results: Twenty-seven of 162 evaluable patients with :6 months' follow-up developed pneumonitis requiring steroids. Single metrics were typically not good predictors for pneumonitis (area under ROC curve = 0.5-0.68). The two-dimensional models (e.g., MLD and pre-RT diffusion capacity for carbon monoxide) generally provided greater ROC areas (0.61-0.72). Overall, the models that considered a measure of pre-RT lung function (i.e., pulmonary function tests), the MLD, and mean perfusion-weighted lung dose were best correlated with outcome (ROC area: 0.7) (p < 0.05 compared to unidimensional models). However, because the area under the ROC curve for these models was < < 1, they too seemed not to be ideal. Conclusion: Predicting symptomatic radiation pneumonitis remains difficult. Multiparameter models that consider pre-RT pulmonary function and the three-dimensional dose distribution seem to be best able to predict outcome. Additional studies are needed to better understand the dosimetric/functional determinants of radiation pneumonitis. © 2002 Elsevier Science Inc.