Towards Minimally Domain-Dependent and Privacy-Preserving Architecture and Algorithms for Digital Me Services: EdNet and MIMIC-III Experiments

Abstract

Digital Me can be defined as a service that reflects the user's goals, measures, predicts, and evaluates the individual's status, and recommends actions to improve the status. This paper explores the architecture and algorithms for predicting user status and recommending actions, using only the user's data. It aims to minimize the reliance on domain-dependent models while ensuring the sustainable and privacy-preserving accumulation of data. To validate the architecture and the algorithms, we used EdNet data in education and MIMIC-III data in healthcare. We developed algorithms to recommend activities that optimize the achievement of users' goals. These algorithms operate on a simple principle applicable to general Digital Me services: recommend actions most likely to improve the user's next state, based on their likelihood of success and the expected score. We also demonstrate the feasibility of creating effective health prediction algorithms using personal federated learning, which avoids centralizing sensitive health data by storing it directly on individual devices or in private clouds.