Big-RT: Big Data analysis to identify combinatorial predictors of radiotherapy toxicity for personalised treatment in prostate cancer patients.


Year:

Session type:

Navita Somaiah1,Sebastian Poelsterl1,Anna Wilkins1,Sarah Gulliford1,James Campbell1,Carmen Rodriguez-Gonzalvez1,Sheng Yu1,Veronica Garcia-Perez1,Claire Griffin1,Judith Bliss1,John Yarnold1,Uwe Oelfke1,David Dearnaley1,Emma Hall1,Bissan Al-lazikani1
1The Institute of Cancer Research

Abstract

Background

Nearly two thirds of cancer patients will receive radiotherapy (RT), with about 20% experiencing long-term RT-induced toxicity. Currently patients cannot be stratified by risk of side effects, which restricts treatment intensity for all patients. The pathogenesis of RT-induced toxicity is complex and multi-factorial, yet most predictive analyses to date are restricted to isolated data types and standard statistical techniques, with limited success. We developed and applied bespoke, state-of-the-art machine learning techniques to large-scale, high-dimensional multidisciplinary data from the CHHiP prostate RT-fractionation trial (CRUK/06/016) to identify multi-parametric predictors of RT toxicities.

Method

We performed a fully integrative analysis of clinician- and patient-reported outcomes, co-morbidities, dosimetry and genetic data (via RAPPER/PRACTICAL consortium) collected as part of the trial. Dosimetric data was presented as a functional fit to the DVH. Genetic data was reduced from ~15 X 106 measures per patient to 100 by Bayesian single-SNP tests of association and further refined in a hybrid functional-scalar multivariate model with elastic-net penalty. Our final model was selected using K-fold cross-validation.

Results

We applied our methodology to 721 patients (out of 3,212 recruited) with complete data profiles. Our final model was trained on a feature matrix consisting of 12 clinical, 100 germline and 21 dosimetric variables. We identified novel candidate predictive markers integrating dosimetric, clinical and genetic variables to predict rectal bleeding, which have increased predictive power over isolated measures (mean ROC AUC of 0.713 vs. 0.524 for dosimetry and 0.641 for genetic data).

Conclusion

We have demonstrated the predictive power of our data-rich, integrative machine learning analysis driven by a multidisciplinary team. The resulting novel combinatorial markers predicting RT-induced toxicity will need to be validated in an independent data set. The successful techniques developed in this project will allow similar approaches to be applied to other tumour types treated with RT.

Share this abstractTweet about this on TwitterPrint this pageShare on FacebookEmail this to someoneShare on LinkedIn