Abstract
The continuous delivery of applied machine learning models in healthcare is often hampered by the existence of isolated product deployments with poorly developed architectures and limited or non-existent maintenance plans. For example, actuarial models in healthcare are often trained in total separation from the client-facing software that implements the models in real-world settings. In practice, such systems prove difficult to maintain, to calibrate on new populations, and to re-engineer to include newer design features and capabilities. Here, we briefly describe our product team's ongoing efforts at translating an existing research pipeline into an integrated, production-ready system for healthcare cost estimation, using an agile methodology. In doing so, we illustrate several nearly universal implementation challenges for machine learning models in healthcare, and provide concrete recommendations on how to proactively address these issues.