Machine Learning Data Pipeline for the Democratization of AI

Abstract

The use of artificial intelligence continues to increase. In healthcare, there has been a recent increase in AI applications to real-time individual patient clinical care, as opposed to population-based research or quality improvement efforts. However, the expertise to evaluate and implement these solutions is limited and often congregates in academic medical centers, creating barriers to adoption for smaller community and rural centers. Lowering the barrier to entry for innovative tools can help address disparities in patient outcomes due to access and other urban/rural contributors. We describe a strategy for evaluating commercially available machine learning models to disseminate lessons learned from developing, validating, and implementing machine learning-based models in clinical care in radiation therapy. In addition, we share an end-to-end data pipeline as open-source code with the tools necessary to identify, extract, organize, and process the data for use in machine-learning applications. We illustrate the application of this data pipeline to the use of brachytherapy to treat female cervical cancer patients. The example will show how we used the proposed pipeline to extract 708 potential participants and applied the developed methods and visualizations to clean the data providing 144 study participants for inclusion in our study. Finally, we discuss the anticipated challenges in implementing machine learning models in commercially available FDA-approved devices and suggest solutions using discrete tools built in different programming languages.