Session Item

To use data from an international prospective cohort study of breast cancer patients (pts) to predict the risk of skin induration (SI) after radiotherapy (RT) using a machine learning algorithm that includes dosimetric/clinical/genetic factors.

Pts were treated after breast conserving surgery with conventional/moderate or ultra hypo-fractionated RT with or without a tumour bed boost based on clinical and pathological factors. Pts were enrolled in 7 countries in Europe/US; each centre followed local clinical practice, but the collection of data and genotyping was standardised and centralised. Our endpoint was late grade 1+ (G1+) SI 2 years after RT completion. Inclusion criteria were: no SI at baseline and availability of  complete dosimetric and genetic data.

For every pt, skin was defined as a 5-mm inner isotropic expansion from the outer body. To select a relevant portion of the skin DVH, we extracted the higher dose tail using different volume cutoffs (i.e. 25/50/100/150/200 cc volumes corresponding to 5x5-20x20cm2 areas). We corrected sub-DHVs for fractionation using two possible a/b values from the literature (1.8 Gy, Bentzen 1988 & Raza 2012; 3.6 Gy, Jones 2006 & Budach 2015). We calculated Equivalent Uniform Doses (EUDs) from corrected sub-DVHs, with n spanning from 1 to 0.05. We also considered the minimum dose of the selected DVH tail as an additional dose parameter (Dmin). Toxicity models were built using feed-forward neural networks (FNNs, 10 neurons and 1 hidden layer) following a wrapper method for feature selection. We used separate datasets for input: clinical/treatment/genetic features were constant, while the dosimetric factors (EUDs and Dmin) coming from sub-DVHs varied with volume cutoff and a/b (Fig 1).

The 647 pts included in the analysis had a G1+ SI rate at 2 years of 29.4%. 281 variables were considered: 127 published SNPs (GWAS literature), 40 clinical factors, 93 treatment factors and 21 dosimetric variables (for each volume and a/b). For volume thresholds <200cc, no dosimetric feature was selected by the wrapper method. Therefore, we derived a predictive model (16 features, no dosimetric variable) for use before RT planning (Model 1). At sub-DVH_200cc, for a/b=3.6Gy only Dmin was selected (Model 2) as dosimetric variable, while for a/b=1.8Gy, EUD (n=0.5) and Dmin entered the FNN (Model 3). 

Fig 2 reports the selected features and performance of the 3 models. 

A pre-planning SI model was derived that included information on genetics (6 SNPs), treatment (6 RT, 1 oncology) and clinical factors. Largest volume (200cc) sub-DVH allowed selection of dosimetric features, particularly with a/b=1.8Gy and EUD with n=0.5. Following validation, the model could be used to personalise use of new RT schedules, such as ultrahigh-hypofractionation, to minimise risk of SI.

This work was funded by REQUITE (EU 7th Framework Programm grant 601826) and RADprecise (ERA-NET ERA PerMed grant ERAPERMED2018-244).