Program Information
Machine Learning-Based Modeling to Predict Radiotherapy-Induced Genitourinary Toxicity in Prostate Cancer Using Genome-Wide Association Study
S Lee1*, J Oh1 , S Kerns2 , H Ostrer3 , B Rosenstein4 , J Deasy1 , (1) Memorial Sloan Kettering Cancer Center, New York, NY, (2) University of Rochester Medical Center, Rochester, New York, (3) Albert Einstein College of Medicine, Bronx, New York, (4) Icahn School of Medicine at Mount Sinai, New York, New York
Presentations
SU-I-GPD-T-100 (Sunday, July 30, 2017) 3:00 PM - 6:00 PM Room: Exhibit Hall
Purpose: Associations between radiotherapy toxicity and genetic factors such as single nucleotide polymorphisms (SNPs) are largely unknown. This study investigates whether a machine learning-based modeling can predict individual risk of radiation-induced urinary toxicity using genome-wide association study.
Methods: We studied 195 prostate cancer patients who met the following criteria: 1) received brachytherapy with or without external beam radiotherapy (EBRT), 2) minimum 3 years of follow-up, and 3) mild baseline urinary condition according to the International Prostate Symptom Score (IPSS) guidelines. A late toxicity event was defined as worsening of urinary condition to moderate or severe IPSS symptom levels (event rate 65%) between 1 and 3 years after radiotherapy. The patients were genotyped for 606895 SNPs. The dataset was split into training (2/3 of samples) and validation datasets (1/3 of samples). The pre-conditioned random forest regression (PRFR) model (Oh et al., 2017, Scientific Reports) was trained in a 5-fold cross validation (CV) setting and also using the entire training set (each repeated 100 times).
Results: Within the training set, 495 SNPs were highly associated with the toxicity (chi-square p value <0.001). When the training was repeated through CV, the PRFR model using these features recorded the validation area under the curve (AUC) of 0.59 (standard deviation (σ) = 0.01). When the entire training set was used for training, the validation performance was AUC = 0.62 (σ = 0.01) and Spearman’s ρ = 0.19 (p=0.13). The most strongly associated non-genetic variable was use of alpha-blockers before or during radiotherapy (p=0.03), which did not improve the performance of PRFR when included (p=0.49).
Conclusion: The validation results of our SNP-based prediction model show that the variation of individual radiosensitivity could be partially explained by genetic differences. We will refine the PRFR approach by incorporating more relevant clinical or dosimetric features to improve prediction accuracy.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by the NIH P30 grant.
Contact Email: