Machine learning reference architecture
This section depicts a typical machine learning lifecycle and data flow.
ML lifecycle with detailed phases and expanded components
The following steps detail the end-to-end data flow for machine learning:
-
Data is collected and pre-processed using a data lake, as described in the preceding healthcare analytics scenario.
-
Features representing clinically valid events, concepts, and processes of care are extracted from raw data and stored in feature stores for model training.
-
The ground truth of data labels is populated and reviewed by humans, which can be used to build supervised classification or regression models.
-
Standard ML training, tuning, and evaluation workflows are used to develop models.
-
Models are reviewed by cross-functional stakeholders, such as clinical leaders and regulatory reviewers. Models are evaluated based on performance and explainability requirements
-
Accepted models may be integrated with IT systems used for care delivery, such as EHRs and medical devices.
-
Model inferences are incorporated in clinical workflows. Providers may be trained on how to use the models as they deliver care.
-
Model inferencing pipelines are monitored, and the performance of deployed models is periodically checked.
