Artificial Intelligence for Combining Evidence-based and Data-driven Medicine
Rapid advances in Artificial Intelligence (AI) and machine learning may help to unlock the value of healthcare data and contribute to solving some of the major problems of modern healthcare, prevent diseases and save lives. Two of the key problems for realizing these objectives are the challenges of accessing healthcare data needed for developing AI models, and the inherent sensitivity of the models to biases in healthcare datasets.
In particular, the standard assumption is that statistical regularities observed in data used for developing AI models are similar to those of the target populations where the models are intended to be used. However, in healthcare this assumption rarely holds. For example, hospitals may serve disparate populations with differing needs, demographics, socioeconomics, and clinical characteristics; and vary in the quality of procedures and levels of innovation. As a result, AI models may learn to rely on spurious or counter-intuitive local regularities and not generalize across hospital settings.
Pharmatics’ solution, MedAI, helps to address both problems by marrying evidence-based and data-driven medicine. MedAI operates by retrieving medical knowledge about clinical conditions and outcomes from millions of medical publications and formalizes the retrieved information using formats suitable for enriching the development of diagnostic models and preventative interventions. By identifying, formalizing, and aggregating a priori evidence from biomedical literature, MedAI indirectly employs hundreds of data sources that had been used for deriving the evidence summarized in the literature.
Using the extracted evidence as the starting point in the development of medical AI models, MedAI helps to achieve higher-quality predictions using less data, reduce the effects of local biases, and develop more explainable and reproducible artificial intelligence for applications in medicine. The approach has been used at the backend of clinical decision support and digital therapy developments for long-term conditions, including, for example, UK-wide developments in diabetes. It has been used to develop patient-facing guidance for self-management of chronic lung disease during COVID-19 (adopted by NHS in Scotland and helping thousands of high-risk patients). A limited instance of one of the knowledge bases retrieved by the MedAI platform will be used for educating a new generation of European engineers in the new AI for eHealth course prepared by the Erasmus+ consortium ATHIKA.
Check out a video describing MedAI: