Dosimetric Predictive Models for Urinary and Rectal Toxicities after Prostate Stereotactic Body Radiotherapy Assessed from the Expanded Prostate Cancer Index Composite Patient-reported Quality of Life

Abstract

The underlying correlations between treatment planning parameters and treatment toxicities are greatly needed for prostate SBRT. We aimed to assess dosimetric quantities and their associations with patient-reported quality of life (QOL) using advanced machine learning techniques. One hundred one prostate SBRT patients were pooled and analyzed. Patients received 40 Gy in 5 fractions on a LINAC using an IMRT arc therapy delivery technique. Patient-specific differential dose-volume histograms (dDVHs, in 1 Gy dose bins from 0 to maximum doses) for rectum and bladder were extracted from the treatment planning system. Other pre-selected quantities, such as planning target volume (PTV) /rectal/bladder volumes and maximum doses, conformality index, etc. were also considered as potential predictors. Patient-reported QOL scores from the Expanded Prostate Cancer Index Composite (EPIC-26) were collected at consultation (baseline) and at 3- and 12- months post-SBRT. Patients were grouped by their score changes (decrement/no change/increment) according to minimally important difference score thresholds of 4, 5, and 6 in bowel, urinary irritative and urinary incontinence domains, respectively. Ensemble machine learning methods including Gradient Boosting Decision Tree (GBDT) were applied to interrogate each dose bin of the dDVHs and the other potential predictors to evaluate their relationships with QOL changes. The most plausible dosimetric features and their importance were derived from the reduction of square loss after tree node splitting using the GBDT. The Pearson correlation tests were performed on final selected dosimetric features. Multivariable linear logistic regression models were then constructed using the selected top features. The performance of predictive models was evaluated using area under the curve (AUC) using 5-fold cross validation. An independent cohort of 15 patients were used to validate the predictive models. At 12 months, bladder dosimetry demonstrated predictive ability for urinary irritation and incontinence symptoms with AUCs of 0.72 and 0.81, respectively. AUC = 0.65 was observed for rectal toxicities. Predictive models, constructed using top 5, 10 and 20 most plausible dosimetric parameters, yielded AUCs of 0.72, 0.76, 0.75; 0.62, 0.8 and 0.8; and 0.73, 0.76, 0.73 for urinary irritation, incontinence and bowel toxicity, respectively. We demonstrated the correlations between dosimetric quantities and patient-reported outcomes for bladder toxicities, dosimetric metrics had less predictive ability for rectal toxicity at 12 months after prostate SBRT. The linear models built on top 10 dosimetric parameters achieved great performances to predict toxicities. The identified dosimetric metrics and the multivariable linear models can be used to refine planning guidelines towards personalized treatment.