Room: AAPM ePoster Library
Purpose: Better understanding of the relationships between treatment planning parameters and treatment-related toxicities may further inform planning objectives. We assessed dosimetric quantities and their associations with patient-reported urinary and rectal toxicities using advanced machine learning and multivariable modeling.
Methods: 116 prostate SBRT patients treated on a Brainlab Novalis Tx linear accelerator using RapidArc delivery technique (40 Gy in 5 fractions) were pooled and analyzed. Patient-specific differential dose-volume histograms (dDVHs, in 1 Gy dose bins from 0 to maximum doses) for the rectum and bladder were extracted from Eclipse. Other dosimetric quantities including planning target volume (PTV) /rectal/bladder volumes and maximum doses, conformality index, etc were also considered as potential predictors. Patient-reported QOL (EPIC-26) scores were collected and fed into eML algorithms to identify most plausible dosimetric predictors. Performance of the classifiers was evaluated via area under the curve (AUC) using 5-fold cross-validation for the subset of 86 patients. The Pearson’s R tests were performed on the top selected dosimetric features, which were used to construct multivariable linear logistic regression predictive models. An independent cohort of 30 patients was used to validate the models.
Results: Dosimetric features demonstrated predictive ability for urinary irritation, urinary incontinence and rectal symptoms with AUCs of 0.72, 0.81, and 0.65, respectively. Multivariable linear regression models, constructed using top 20, 10 and 5 most plausible dosimetric parameters, achieved AUCs of 0.75, 0.76, 0.72; 0.8, 0.8 and 0.62; and 0.73, 0.76, 0.73 for urinary irritation, incontinence and bowel toxicity, respectively.
Conclusion: Dosimetric correlations were found for patient-reported bladder toxicities at 12 months after prostate SBRT. The linear models utilizing the top 10 dosimetric parameters achieved great predictive ability. The identified dosimetric features and the multivariable predictive models can be used to refine planning guidelines and potentially reduce toxicity.
TH- Dataset Analysis/Biomathematics: Machine learning techniques