Technical Feasibility of Implementing and Commercializing a Machine Learning Model for Rare Disease Prediction

Wild-type transthyretin amyloid cardiomyopathy is an under-recognized cause of heart failure. Pfizer previously developed a machine learning model that performed well in identifying wild-type transthyretin amyloid cardiomyopathy vs. nonamyloid heart failure. However, challenges exist when introducing machine learning applications into healthcare, mainly due to restrictions on sharing patient data. This requires the triggering and execution of the model outside the developer’s infrastructure using hosts with diverse information technology capabilities. With these barriers in mind, we investigated architectural designs to facilitate the delivery of the model to the customer. We considered manageability and scalability of the model, the host’s information technology maturity, maintenance of patient data privacy, and protection of Pfizer’s intellectual property. A container-based design, wherein the application is shipped as a container image to a third-party platform, fulfilled these criteria and was piloted on a platform hosted by Philips. Objectives of this pilot included defining and testing the architectural design and technical parameters for sharing the container image, creating a scalable and modular framework to manage multiple applications on different third-party platforms, and exploring a communication pattern based on clinical decision support Hooks/Cards and representational state transfer calls within the Philips platform. Implementing the model may enable earlier identification and treatment of wild-type transthyretin amyloid cardiomyopathy, and learnings from this pilot may lead to improved delivery of other machine learning models to healthcare providers, thereby increasing utilization. This article also presents an overview of architectural designs that may help others adopt new methodologies and ideas for machine learning model commercialization.