The FDA, along with Health Canada, and the United Kingdom’s (UK) Medicines and Healthcare products Regulatory Agency (MHRA), released new guidance that will bring some major changes to the development Artificial Intelligence and Machine Learning (AI/ML) healthcare products.
This joint effort lays out a uniform approach to global standards that will serve to maximize efficiency and innovation throughout the industry, much like GMP did for manufacturing, GLP did for pre-clinical testing, and GCP did for clinical trials.
The Good Machine Learning Principles are:
Multi-disciplinary expertise is leveraged throughout the total product life cycle
Good software engineering and security practices are implemented
Clinical study participants and data sets are representative of the intended patient population
Training data sets are independent of test sets
Selected reference datasets are based upon best available methods
Model design is tailored to the available data and reflects the intended use of the device
Focus is placed on the performance of the human-AI team
Testing demonstrates device performance during clinically relevant conditions
Users are provided clear, essential information
Deployed models are monitored for performance and re-training risks are managed
Some of the major concerns regulatory authorities are hoping to address with the GMLPs?
Bias remains a major concern for AI/ML products, as inequity in the patient population can result in a product that passes performance criteria but is only truly effective for a subset of the intended patient population. Ensuring training and test data sets are representative of the whole intended patient population is imperative in developing non-biased AI software. Necessitating representative datasets as part of GMLP will mean that regulatory authorities should be looking to confirm that the data sets include appropriately diverse demographics, meaning appropriate numbers of patients of the intended races, sexes, and ages. So, if you’re developing an AI/ML product, be mindful of this during training and validation.
Establishing how the AI/ML product fits into the clinical workflow will be key for communicating with FDA, Health Canada, and the UK MHRA, as well as patients and providers. Clear articulation of the limitations of the product will be necessary to ensure that physicians and patients are able to interpret the outputs of the software accurately and appropriately. A balance of instruction and detail in the labeling may also aid in the users' understanding of the outputs of the device, allowing them to make informed decisions.
Separating training and test data sets is also imperative when testing the skill of the model. If these two sets are enmeshed, then the test set becomes biased and unable to provide an accurate representation of how the model will perform in the real world.
Impact on Patient Population?
The focus to provide superior AI/ML products, while also closing the door on loopholes or shortcuts that lead to inferior products, is great news for patients. If developed and deployed well, AI/ML products can provide access to care in underserved patient populations due to the products’ efficiencies and lower costs. The initiative to create a standard set of GMLP principles and expectations also gives guidance to companies developing artificial intelligence and machine learning applications for healthcare so that they know what to expect when engaging with regulatory bodies worldwide. This common understanding will make the product development and regulatory review processes smoother, resulting in the products more rapidly getting into the healthcare workflow.
What Does this Mean for Future AI/ML Developments?
Overall, this is good news for the AI development industry and the public. Standardized practices will only aid innovation, producing higher quality products and providing more efficient and equitable treatments.