Maturity Framework for Operationalizing Machine Learning Applications in Health Care: Scoping Review

Area of Science:

• Healthcare Informatics
• Machine Learning Engineering
• Clinical AI Implementation

Background:

• The application of machine learning (ML) in medicine is rapidly expanding, yet practical implementation in clinical settings faces significant barriers.
• Machine Learning Operations (MLOps) practices, common in IT, are under-researched in healthcare, limiting the deployment of ML models.
• Existing literature lacks comprehensive details on the feasibility and operationalization of MLOps within the unique context of healthcare.

Purpose of the Study:

• To investigate and detail the implementation of MLOps within healthcare environments.
• To propose a novel MLOps maturity framework tailored for healthcare applications.
• To identify key components and considerations for successful MLOps deployment in the medical field.

Main Methods:

• A scoping review was performed following the Joanna Briggs Institute Manual for Evidence Synthesis.
• Four major databases (MEDLINE, Embase, Web of Science, Scopus) were searched for studies on MLOps proof-of-concept or real-world implementation in healthcare.
• A 3-stage basic qualitative content analysis was used to synthesize findings from 19 included studies.

Main Results:

• The MLOps workflow in healthcare encompasses data extraction, preparation, model training, evaluation, validation, deployment, continuous monitoring, and continual learning.
• A 3-stage MLOps maturity framework (low, partial, full) was proposed, with 13 of 19 studies demonstrating full maturity.
• Eight studies addressed critical ethical, legislative, and stakeholder considerations for MLOps in healthcare.

Conclusions:

• Limited studies report on ML implementation in healthcare, underscoring the need for improved data infrastructure and collaborative development.
• Engaging patients, policymakers, and healthcare professionals is crucial for the successful creation and implementation of ML healthcare applications.
• Variations in study quality impacted the depth of analysis for each MLOps workflow step, highlighting a limitation in current research.