Transforming AI Solutions in Healthcare—The Medical Information Tokens

Abstract

Comparatively to other fields, the use of artificial intelligence (AI)-based medical decisions remains limited. In general, AI-based medical decisions necessitate modeling the conditions being treated, including potential treatment alternatives and trajectories. Due to the heterogeneity of most medical conditions, developing these models requires a vast quantity of data. Although current healthcare systems generate abundant digital data, this data is frequently fragmented, stored in multiple locations, organized according to various structures, and lacks context, posing a challenge for integrating and utilizing these data to model the relevant conditions. Effective AI medical solutions, we argue, necessitate the development of an AI-specific medical vocabulary, which we call information tokens. The proposed vocabulary would enable AI methods to access diverse medical records and provide the context for developing treatment models, instead of only conditions-specific models. Then, we introduce a reinforcement learning-based agent, the human digital twin, which models medical conditions and provides patient-specific treatment recommendations based on the condition vocabulary. Finally, we define a framework for managing and coordinating the utilization and update of the vocabulary, and the treatment models throughout a healthcare system.