e14118 Background: One of the emerging medical applications of artificial intelligence (AI) is the interpretation of mammograms for breast cancer screening. It is uncertain what attributes would result in acceptance of AI for breast cancer screening (AI BCS) among ordering clinicians. Methods: We performed qualitative interviews to identify the most important attributes of AI BCS for ordering clinicians. We then invited US-based primary care providers (PCPs) to participate in a discrete choice experiment (DCE). The experiment featured 15 choices between radiologist alone and two AI BCS alternatives where respondents traded better metrics on some attributes for worse metrics on others. Responses were analyzed using a mixed logit model adjusting for preference heterogeneity to determine the probability of recommending AI BCS. Results: In qualitative interviews, the six most important attributes to PCPs were AI sensitivity, specificity, radiologist involvement, understandability of AI decision-making, supporting evidence, and diversity of training data. Forty PCPs completed the DCE. Sensitivity was the most important attribute: a 4 percentage point improvement in sensitivity over the average radiologist increased the probability of recommending AI by 0.41 (95% confidence interval (CI), 0.38-0.42). Specificity was approximately half as important. Respondents were indifferent to whether radiologists confirmed all or only screens likely to be abnormal. However, no radiologist involvement reduced the probability of recommendation by 0.31 (95% CI, 0.29-0.31). An AI developed using data from diverse populations increased the probability of recommendation by 0.38 (95% CI, 0.36-0.39). Lastly, an AI that is transparent in the rationale for its decisions increased the probability of recommendation by 0.41 (95% CI, 0.39-0.41). Conclusions: PCPs prefer AI BCS that improves sensitivity versus specificity, and involves radiologists in the confirmation of abnormal screens. Improving sensitivity alone, however, will likely not be sufficient to support widespread PCP acceptance – algorithms will need to be developed with diverse data and more transparent explanations of their decisions.