Preconditioned Random Forest Regression: Application to Genome-Wide Study for Radiotherapy Toxicity Prediction

Abstract

Urinary toxicity after radiotherapy (RT) limits the quality of life of prostate cancer patients, and clinically actionable prediction has yet to be achieved. We aim to exploit genome-wide variants to accurately identify patients at higher congenital toxicity risk. We applied preconditioned random forest regression (PRFR) to predict four urinary symptoms. For a weak stream endpoint, the PRFR model achieved an area under the curve (AUC) of 0.7 on holdout validation. Preconditioning enhanced the performance of random forest. Gene ontology (GO) analysis showed that neurogenic biological processes are associated with the toxicity. Upon further validation, the predictive model can be used to potentially benefit the health of prostate cancer patients treated with radiotherapy.