Abstract P2-06-05: Development and validation of a new non-parametric breast cancer risk assessment model on US and European screening populations

Abstract

Abstract Background: Stratified breast cancer (BC) prevention is a major option for the future but requires clinically meaningful internationally validated risk models. Non parametric models may be alternate methods for modeling in very large cohorts. We have previously shown that a non-parametric similarity-based k-nearest neighbors' (kNN) model performs better than the BCRAT/Gail model to on 65 000 women of the E3N French national cohort (Dartois et al 2015). We used this method to develop and validate a mammographic density-based model in larger general screening populations (pops). Methods: A modified version of a data-mining based algorithm, the kNN method, was implemented and adapted as previously described [ref Dartois]. Core concept of kNN algorithm is to gather similar profiles using a distance computation. We developed a BC risk prediction model on 629 229 women (wn) from the US Breast Cancer Research Consortium (BCSC), with 5 times random selection of learning and validation sets (75/25 %) within the cohort. 5 parameters were included: age, family history, previous breast biopsy, mammographic density and race. The model's performances (discrimination using c-stat (AUC) and calibration using E/0 ratio) were evaluated and compared to the parametric model developed on the same pop (param/BCSC)(Tice et al 2008). This kNN/BCSC model was then tested on two French screening pops after adjustment on French BC incidence: an urban area (Paris suburbs, N=316 775) and a rural area (Côte d'Or, N=32 930). Its performances were compared to those of a model directly developed (same methods) on the Paris cohort (kNN/Paris). Levels of individual risks assessed by the models were assigned into 4 risk categories. The sensitivity of the models was defined as the number of wn who had BC whose 5 yrs-risk category was intermediate (median risk at 50-yrs - 1.66%), high (> 1.66%) or very high (> 20% lifetime) divided by the total number of wn who had BC. Results: The performances of the different models are shown in Table 1. The kNN model developed on BCSC performed well (c-stat 0.653 and E/0 1.001). It had equivalent performances as the parametric model developed previously on the same pop. This kNN/BCSC had a good discrimination on French pops, although slightly lower than that on US pops. This is expected since French screening starts at 50 (vs 40 in BCSC) and French parameters do not include race. The calibration of such model was excellent on Paris, while it overestimated the risk on Côte d'Or, in which BC incidence is lower. It performed as well as the kNN/Paris model directly developed on Paris' pop. The sensitivity of the kNN/BCSC on US and French pops was good. Conclusions: A new non parametric kNN breast cancer risk model developed on an American screening cohort (BCSC) was successfully validated on two French screening cohorts. This new international model could allow stratified prevention. Performances of the models Param/BCSC on BCSCParam/BCSC on PariskNN/BCSC on BCSCkNN/BCSC on PariskNN/BCSC on Cote d'OrkNN/Paris on ParisN629 229313 817629 229313 81732 930313 817c-statistic0.6580.6020.6530.6020.5930.605E/0 global1.031.071.0011.061.521.00Sensitivity--76.18%72.87%72.58%- Citation Format: Ragusa S, Gauthier E, Dartois L, Tice J, Dancourt V, Arveux P, Brixi Z, Bernoux A, Soyer P, Delattre H, Brechenade S, Catajar N, Kaufmanis A, Hélin VM, Clavel F, Kerlikowske K, Miglioretti D, Delaloge S. Development and validation of a new non-parametric breast cancer risk assessment model on US and European screening populations [abstract]. In: Proceedings of the 2016 San Antonio Breast Cancer Symposium; 2016 Dec 6-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2017;77(4 Suppl):Abstract nr P2-06-05.