Questions

Employ version control for source code, data, and ML artifacts to ensure traceability and reliability of production ML deployments. Version control and traceability may be required by applicable regulatory frameworks, if for example models are deployed in support of medical devices.

Data science in healthcare often depends on open-source libraries for data processing, model development, training, and hosting. Establish a process to review the privacy and license agreements for all software and ML libraries needed throughout the ML lifecycle. Verify that these agreements comply with your organization’s legal, privacy, and security requirements.

Many ML workflows do not require identified health data. Applying ML to deidentified data is one way to develop AI-powered applications without compromising privacy or data security. Cloud services like the Amazon Comprehend Medical DetectPHI API can streamline generating deidentified datasets.

Statistical models trained on real-world health data are susceptible to biases. Health data may inadvertently be collected from populations of individuals with similar characteristics, such as median household income, social determinants of health, and access to care. Care setting and health insurance coverage may also impart biases. For example, treatment cohorts may exhibit higher household income because such individuals may have greater access to advanced care.

Trained models may be misleading if biases are not quantified and mitigated. Also, models may be inaccurate when trained on biased data and applied to settings with different distributions. Examine data distributions and perform analyses to quantify and mitigate biases before training models.

Health data is often complex, and subject to temporal variations in quality and concept expression. Model performance may degrade over time due to data quality, model quality, and concept drift. Create a baseline for data quality, and automate monitoring performance in production. Automate alerts for changes in data quality or distributions, such as age deciles and prevalence of relevant chronic diseases. SageMaker AI Model Monitor provides an end-to-end framework model monitoring and lifecycle management.