Published on May 6, 2019

Today, six of the top ten highest-grossing drugs in the US are effective in less than 10% of patients and even the most effective drugs from that list have positive responses in only 25% of patients. This “imprecision medicine” practice not only harms certain populations of patients, it also burdens the healthcare system financially. By finding meaningful and homogeneous subgroups prior to conducting costly clinical trials, researchers can further study focused populations and identify potential risk factors through slicing and dicing from complex phenotypic/genotypic information sources. Advancements in machine learning algorithms have shown promising results in many biomedical applications and may provide a potential solution for subgroup discoveries. Unfortunately, many of the high-performing approaches result in black boxes that are not explainable and often fail to close the known gap between computational innovation and clinical practice. In this talk, I will introduce a novel deep exploratory mining framework that is designed to answer the following two questions: “What hypotheses are likely to be novel and produce clinically relevant results with well thought-out study designs?” and “Which subgroups of patients might benefit from interventions that are likely to be effective for the selected populations?” Innovations in actionable and explainable AI (X2AI) and applications of Big Data technologies make it feasible to tackle this complex biomedical informatics problem. Although this framework is domain independent, I will use a case study in autism spectrum disorder (ASD) to explore a large cohort of ASD patients and find subgroups with similar behavioral and communication characteristics that share underlying genetic patterns for potential interventions.