"Trust through evidence"

Summary of key messages:

​

• AI solutions require both model evidence (for the underlying algorithm) and solution evidence (for the product in which the algorithm is embedded). Models require validation internal and external datasets (internal and external validation).​

​

• Given their likely use to support decision making in healthcare, it is critical for AI developers to show that their solution does not reinforce or exacerbate existing biases and inequities, including cognitive biases present in the augmented decision-making process.

​

• Evidence is required throughout the product life cycle (product development; regulatory approval; market access/payment; post-market surveillance). Different types of evidence are required for different aims and different stakeholders.

​

• The evidence required to enter markets (e.g., FDA/CE/UKCA marking) differs depending on the class of the tool and may require additional evidence.

​

• AI developers will face an evidence limbo between the early-stage evidence that is sufficient for product development and the evidence needed at later stages of the product life cycle to demonstrate specific outcomes. Innovative evidence generation methodologies like clinical simulation can help bridge that gap.

​

• Once the solution is in use, there are significant opportunities to generate real-world evidence to demonstrate its value. This evidence needs translation to determine whether it can be considered generalisable to other populations/workflows or should be considered only applicable to a single context.

​

• Ultimately, commercial success will depend on whether a product can deliver value for money for a client. Economic modelling and analysis are therefore essential.