Integration of multi-modality treatment planning for early stage breast cancer (BC) into Watson for Oncology, a Decision Support System: Seeing the forest and the trees.

e12042 Background: IBM Watson for Oncology (WFO), trained by Memorial Sloan Kettering (MSK) is designed to assist medical oncologists in making expert treatment decisions driven by key disease and patient-specific attributes. As previously reported, medical logic instilled by MSK and IBM Watson's machine learning on actual and manufactured MSK cases allow WFO to consider nuances not addressed by existing guidelines. Decision making for early stage BC extends beyond adjuvant systemic therapy, to encompass breast surgery (BS), radiation (RT), clinical genetics (CG), and fertility preservation (FP) issues. Oncologists are imperfect in applying these referrals, with potential for both over- and under-referral. Contemporary digital decision support tools do not yet integrate capabilities that ensure optimal multidisciplinary patient care. Methods: We utilized published evidence, guidelines, and MSK expertise in training WFO to recommend clinically indicated referrals to multidisciplinary specialists. We measured the frequency with which WFO prompted consultation to ensure optimal breast/axillary surgery, adjuvant RT, and CG or FP management, and the degree of its agreement with two MSK breast medical oncology specialists. Results: Across 103 de-identified historical stage I-III BC cases entered into WFO after primary surgery, WFO output had 98% agreement with MSK recommendations on RT referral; 94% on CG counseling; and 91% on FP counseling as detailed in Table 1. WFO recommended BS consultation in 8/8 manufactured cases where MSK recommended completion axillary surgery, but also in 7/12 where MSK did not. Conclusions: An expert-guided machine learning approach informed by published evidence and guidelines succeeds in generating multidisciplinary care plans for early stage BC. We continue to optimize WFO with additional training data and by focusing on natural language processing and the incorporation of patient preference. Rate (%) of WFO recommendation by subgroup. Referral Subgroup Patient Age Rate of WFO Recommendation FP Counseling All <=40 100 41-45 6 >45 0 CG Counseling HR+ and/or HER2+ <=50 98 >50 25 TNBC <=60 92 >60 67