Program Information
Computational Modeling of Risk of Radiation Pneumonitis
J.V. Logan, D.M. Trifiletti, J.M. Larner, J.V. Siebers, W.T. Watkins*, University of Virginia, Charlottesville, VA
Presentations
TH-AB-304-8 (Thursday, July 16, 2015) 7:30 AM - 9:30 AM Room: 304
Purpose: Current normal tissue complication probability models estimate incidence of radiation pneumonitis (RP) in the treatment of lung cancer based on single parameters, e.g. mean lung dose or V20, and therefore do not capture specific variations contained in dose volume histograms (DVHs). We therefore developed an RP-risk model which utilizes the entire DVH reasoning that such a model would be more robust for predicating radiation-induced lung disease (RILD).
Methods: A logistic regression model was developed to predict RP-risk on each point of the cumulative DVH utilizing >1200 patient outcomes summarized in the Quantitative Analysis of Normal Tissue Effects in the Clinic (QUANTEC) lung study. Stepwise regression including linear, quadratic, and mixed terms in dose and volume was applied to the single dose-volume parameter-versus-outcome data. The model computes a relative importance at each dose-volume point and these values were linearly combined to estimate an overall RP-risk. A set of 80 DVHs were used to validate the algorithm prediction by comparison with an independent predictor of RP-risk using mean lung dose.
Results: The logistic regression coefficients were all statistically significant (p<0.005). The proposed model predicted RP-risk consistently in agreement with the QUANTEC-RP estimates computed utilizing only mean dose (average difference 0.5%±1.8%) for the 80 DVH set. Most (>70%) of the model variability occurs at doses <30 Gy for DVH levels ranging from 30%-100%. The model is most sensitive to DVH variations in the V5-V30 range.
Conclusion: The model’s predictive power is comparable to estimates from QUANTEC which use mean dose as input. Since the model conveys the relative importance of each dose-volume level it has the potential to better predict RILD than currently available models. In order to determine if our hypothesis is correct we will compare the model to established methods including complete DVH data and clinical outcomes.
Funding Support, Disclosures, and Conflict of Interest: This abstract was supported by the George Amorino Pilot Grants in Radiation Oncology from the Department of Radiation Oncology, Univeristy of Virginia.
Contact Email: