The Consumer Technology Association (CTA) has released a new artificial intelligence standard that requires model developers to meet specific accuracy and explainability requirements for pre-market solutions.
The standard for predictive AI focuses on accuracy and reliability of models and emphasizes complete data collection and the mitigation of bias in data. Using the specifications will help standardize the AI industry and build trust in the technology, the CTA writes.
The CTA releases many health technology standards, like performance requirements for integrated continuous glucose monitors and sleep tracking consumer devices. The fifth of its AI standards, called Performance Verification and Validation for Predictive Health AI Solutions, creates a set of requirements for accuracy, data verification, explainability and real-world testing.
The CTA standard is only applicable for non-generative AI technologies. The association clarifies that the standard is not relevant for use cases like transforming unstructured electronic medical record data into structured data, or AI scribes, though it notes that generative AI will be addressed in a later version of the standard.
“This standard emphasizes a holistic approach that considers data quality, model accuracy, utility, and explainability,” the standard document says. “This focus strives to ensure high quality healthcare applications spanning the health journey which may be used for diagnosis, treatment selection, patient monitoring, improved patient experience, and even help with administrative tasks for the caregiver.”
Data verification is one of the pillars of the CTA standard. The data verification part of the standard requires transparency of input and output data elements such as how the data were obtained.
The standard gives the example of a predictive model for assessing breast cancer risk where the input is the presence or absence of the breast cancer gene BRCA1 or BRCA2 and the dependent variable is the high risk of breast cancer as assessed by biopsy. The model result, or prediction, is based on the statistical relationship between the independent and dependent variables.
Model developers are required to report the results of at least one accuracy measure like an F1 score or Mean Absolute Error as part of data verification.
Model developers also are required to disclose the number, age and gender demographics of the people included in the final model. Disclosure of the race and ethnicity split of the test and validation population is suggested to be disclosed but not required by the standard.
The standard seeks to ensure that the model is well explained enough for local personnel to be able to implement and understand the solution. The CTA will require the model developer to describe the purpose of the AI solution, explain how to install and use it, provide user manuals and provide contact information for technical support.
It also offers additional recommendations and resources for explaining the model.
Model developers should have a plan to address model degradation and drift, the CTA writes, such as quality controls to track variances in the model and pre-identified benchmarks that signal a need to recalibrate the model.
The final components of the predictive AI standard are basic deployment testing and full operational validation. The two stages of testing happen after initial internal tests for accuracy and involve testing the model at other sites and in real-world settings. Model developers are required to report the accuracy results of the external validation compared to the initial tests.
The standard promotes compliance with the Health Insurance Portability and Accountability Act and the EU’s data privacy act.