Although there are many risk scores in the health field to predict disease risk, they are not as used as they could be to individualize and enhance prevention based on an estimated risk level. In order to facilitate the production of risk scores that are efficient in detecting high risk profiles and that fit to the context of use, we suggest a risk score building process. In order to conduct experiments, we build an information system architecture that supports the building and use process of risk scores. Thanks to the implementation of this architecture, we use our process to experiment the creation of breast cancer risk scores based on a publicly available american database and on the E3N French cohort study database. Using the breast cancer example, we show that it is possible to obtain comparable performances in terms of discrimination and better performances in calibration than available risk scores of the literature, using a readable k-nearest-neighbor algorithm and less attributes.
