UCLA

Recent years have seen a tremendous amount of growth in the performance and adoption of artificial intelligence (AI) and machine learning (ML) systems. These systems now permeate our lives, underpinning everything from web search to credit card fraud detection and photography. In principle, these advancements could also be applied to the domain of healthcare, where they could improve patient outcomes.However, despite the almost fifty years that have elapsed since the first National Institutes of Health AI in Medicine (AIM) workshop in 1973 and the ubiquity of AI systems in our daily lives, AIM has not yet lived up to its lofty promises. AIM systems have seen limited deployment due to challenges including data missingness, data heterogeneity, explainability, and generalizability across variances in patient populations. The recent increase in the availability of electronic health record information, the variety and cost-effectiveness of mobile sensors, and the capabilities of machine learning algorithms promise to help improve healthcare delivery if challenges can be overcome. Through techniques such as interpretable analysis of heterogeneous information networks and missingness-aware modeling, we demonstrate that the challenges of AI in Medicine can be overcome in order to improve healthcare access, aid physicians, and generate new insights into disease.