Abstract PS8-37: A breast cancer risk self-assessment model for Japanese women

Abstract

Abstract Introduction No fully validated breast cancer risk prediction model exists for Japanese women. We produce a collection of BC risk models and a means of selecting the most clinically interpretable and thus to a model that women can use directly for primary prevention. Methods A dataset of 2494 Japanese women were collected in 2014 - 2015 and was divided into six groups by parity (Parity and Nulliparity) and thereafter segmented into age as an estimate of menopausal status: premenopausal (PRE; 20 ≤ age < 45), perimenopausal (PERI; 45 ≤ age ≤ 55) and postmenopausal (POST; 55 < age ≤ 80). Some “history” status included multiple variables; breastfeeding history: breastfeeding experience (yes or no), number of children breastfed, breastfeeding duration (months); family history: BC cases within the 1st, 2nd, 3rd and 4th degree relatives (yes or no), number of BC cases in within the 1st, 2nd, 3rd and 4th degree relatives; smoking history: smoking experience (yes or no), Brinkman index dichotomized at 50 (less or more) and optimal value respectively; alcohol consumption history: regular consumption (yes or no), amount of monthly consumption (g), monthly consumption at optimal value (less or more). Thereafter, a logistic regression model was tested against all variables including age, BMI, parity, breastfeeding history, family history, smoking history, and alcohol history. Optimal fits were derived from adjusting thresholds and best variable representing the ‘history’ to optimize the ROC curves. All resulting optimally fit models were evaluated with AIC, and the top 10% of models (as calculated by AIC) were selected and further evaluated by AUC within the six groups. Internal validation was conducted by stratified five-fold cross validation (CV) (1000 times). Results A total of 2494 patient records were grouped by; Parity: PRE (149 cases and 184 controls): PERI (326, 415): and POST (439, 465); Nulliparity: PRE (73 cases and 148 controls): PERI (78, 122) and POST (28, 67). Based on criteria described in Methods, 480 (Parity) and 96 (Nulliparity) models were fit for use in the optimization. The mean AUC of the top 10% models after CV for Pre, PERI, and POST were (.635, .647, and .635) for Parity and (.625, .658, and .653) for Nulliparity respectively. The range of AUC for the “10% models” in each group was less than 0.02 thus there was no significant evidence that any given group 10% model was optimal. Therefore, we deferred to clinical domain knowledge in the form of which model contained the most clinically interpretable variables to select the model with the most translatable interpretation. The resulting “translatable” models include age and BMI; for Parity the additional variables included were number of childbirths, breastfeeding duration, number of BC cases within the 4th degree relatives, smoking experience and dichotomized monthly alcohol consumption for PRE and PERI (mean AUC after CV was .645 and .656 respectively) and breastfeeding experience, BC case within the 4th degree relatives, smoking experience, regular alcohol consumption for POST (.638); for Nulliparity, BC cases within the 2nd degree relatives, Brinkman index dichotomized at the optimal value, and dichotomized monthly alcohol consumption for PRE and POST (.626 and .648 respectively); BC case within the 2th degree relatives, smoking experience, and dichotomized monthly alcohol consumption at the optimal value for PERI (.657) Conclusion In this study, optimal clinically interpretable risk prediction models were derived for Japanese women. Our approach produced equally accurate models and the most clinically translatable was selected based on deep clinical knowledge. Future study will expand and re-execute the approach to produce higher accuracy while maintaining clinical interpretation and therefore use in clinical practice. Citation Format: Michiyo Yamada, Takashi Ishikawa, Sadatoshi Sugae, Kazutaka Narui, Peter Tonellato, Itaru Endo, Takashi Chishima. A breast cancer risk self-assessment model for Japanese women [abstract]. In: Proceedings of the 2020 San Antonio Breast Cancer Virtual Symposium; 2020 Dec 8-11; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2021;81(4 Suppl):Abstract nr PS8-37.