Serverless on FHIR: Deploying machine learning models for healthcare on the cloud
This addresses the problem of efficient and safe model deployment in healthcare systems for clinicians and patients, but it is incremental as it builds on existing cloud and FHIR technologies.
The paper tackles the challenge of deploying and updating machine learning models for clinical decision support by introducing a four-tier cloud-based architecture called Serverless on FHIR, which uses containerized microservices, serverless computing, function as a service, and FHIR schema to improve maintainability, scalability, portability, and discoverability.
Machine Learning (ML) plays a vital role in implementing digital health. The advances in hardware and the democratization of software tools have revolutionized machine learning. However, the deployment of ML models -- the mathematical representation of the task to be performed -- for effective and efficient clinical decision support at the point of care is still a challenge. ML models undergo constant improvement of their accuracy and predictive power with a high turnover rate. Updating models consumed by downstream health information systems is essential for patient safety. We introduce a functional taxonomy and a four-tier architecture for cloud-based model deployment for digital health. The four tiers are containerized microservices for maintainability, serverless architecture for scalability, function as a service for portability and FHIR schema for discoverability. We call this architecture Serverless on FHIR and propose this as a standard to deploy digital health applications that can be consumed by downstream systems such as EMRs and visualization tools.