Correlating Dose Variables with Local Tumor Control in Stereotactic Body Radiation Therapy for Early-Stage Non-Small Cell Lung Cancer: A Modeling Study on 1500 Individual Treatments.

Abstract

CONCLUSIONS

BED_ave was generally better correlated with tumor control probability than either BED_max or BED_min. Because the average between near-minimum and near-maximum doses was highly correlated to the mean gross tumor volume dose, the latter may be used as a prescription target. More emphasis could be placed on achieving sufficiently high mean doses within the gross tumor volume rather than the PTV covering dose, a concept needing further validation.

BACKGROUND

Large variation regarding prescription and dose inhomogeneity exists in stereotactic body radiation therapy (SBRT) for early-stage non-small cell lung cancer. The aim of this modeling study was to identify which dose metric correlates best with local tumor control probability to make recommendations regarding SBRT prescription.

METHODS AND MATERIALS

We combined 2 retrospective databases of patients with non-small cell lung cancer, yielding 1500 SBRT treatments for analysis. Three dose parameters were converted to biologically effective doses (BEDs): (1) the (near-minimum) dose prescribed to the planning target volume (PTV) periphery (yielding BED_min); (2) the (near-maximum) dose absorbed by 1% of the PTV (yielding BED_max); and (3) the average between near-minimum and near-maximum doses (yielding BED_ave). These BED parameters were then correlated to the risk of local recurrence through Cox regression. Furthermore, BED-based prediction of local recurrence was attempted by logistic regression and fast and frugal trees. Models were compared using the Akaike information criterion.

RESULTS

There were 1500 treatments in 1434 patients; 117 tumors recurred locally. Actuarial local control rates at 12 and 36 months were 96.8% (95% confidence interval, 95.8%-97.8%) and 89.0% (87.0%-91.1%), respectively. In univariable Cox regression, BED_ave was the best predictor of risk of local recurrence, and a model based on BED_min had substantially less evidential support. In univariable logistic regression, the model based on BED_ave also performed best. Multivariable classification using fast and frugal trees revealed BED_max to be the most important predictor, followed by BED_ave.