Machine learning for healthcare
Artificial intelligence/machine learning (AI/ML) is being applied to a growing set of problems across healthcare, such as prioritizing treatments, predicting health outcomes, guiding provider workflows, and streamlining revenue cycle operations. A key strength of AI/ML technology is the ability to continually learn from real-world data and improve performance over time. However, healthcare applications of AI/ML pose unique problems, including regulatory oversight, design control obligations, and interpretability requirements imposed by stakeholders. Machine learning development is often performed in concert with traditional analytics and can leverage elements of the infrastructure described in the Healthcare analytics scenario.
Characteristics of machine learning for healthcare architectures include:
-
Data for training models is commonly extracted from production systems within payer and provider organizations, such as EHRs, medical imaging systems, claims and revenue cycle solutions, scanned or faxed documents, biobanks, and genomics data stores. Healthcare tends to be high dimensional, multi-modal, and often suffers from missingness due to the episodic nature of care resource consumption.
-
Data lakes are well suited to landing, preparing, and storing health datasets for ML. The traditional data silos of healthcare data can hamper training models that perform well. Extracting data from multiple systems and data modalities can improve model performance and transferability between settings.
-
Healthcare analytics tools may be used for exploratory data analysis before machine learning development.
-
Data is commonly extracted in bulk, cleaned, and prepared for use training models. Data for inferencing may also be processed in batches, or in real-time using streaming data integrations such as HL7 v2, FHIR, or other interoperability standards. Data lineage should be tracked, providing a map of various data sources and transformation steps that the data flows through.
-
Organizations may use identifiable health data for ML development. In such cases, organizations must adhere to the principle of least access, protecting health data from inappropriate access while enabling data scientists to run ML workflows. De-identifying health data before making it accessible to ML teams can mitigate privacy and security concerns.
-
Federated Learning (FL) may be adopted to facilitate multi-centric, collaborative ML modeling when data privacy and anonymity is required. Federated paradigm has been proven to be domain-agnostic and framework independent for distributed modeling training. The heterogeneous datasets from diverse sources and patient populations are desired to build a robust and accurate ML model. To overcome the privacy concern and complex data sharing process, FL in healthcare can be used to rapidly and securely develop new ML models without data ownership hurdles.
-
Stakeholders may require explainability and repeatability of models employed in healthcare, especially in care delivery and revenue cycle use cases. Validating a model may require that structure and output of the model have face validity and align with medical knowledge.
-
Thresholds of acceptable model performance may follow from the risks and benefits of the outcome modeled. For example, if the outcome being modeled is high risk to patients or high cost, then stakeholders may demand that models display high precision
(confidence in a positive prediction). Alternatively, if the model is used for a low risk, low cost, yet high benefit, stakeholders may prioritize higher recall from models. -
Model deployments should be automated with pipelines to minimize human touchpoints, deploy data integration workloads consistently and repeatedly, and formalize how code is promoted from development to production.
-
Healthcare AI/ML models often need to be integrated in the workflows of healthcare providers, patients, and other actors to have utility. This may require integrating systems (such as an EHR), updating workflows, and retraining providers. It may also require integration with medical equipment and require regulatory oversight.
-
Model performance should be monitored after deployment in production. The complexity and variability of health data makes it especially important to monitor the health of inferencing pipelines.
Machine learning resources
Refer to the following resources to learn more about our best practices for machine learning.
Documentation and blogs
Whitepapers
