Regulatory science for adaptive AI decoders
The methods and evidentiary standards for evaluating BCI decoders that are AI-based and continually learning. Classical medical-device regulation assumes a 'locked' function that behaves identically after approval; an adaptive decoder that recalibrates to drift or personalizes to a user changes its behavior in the field, breaking the assumptions behind validation, traceability, and post-market surveillance.
Regulators treat these as Software as a Medical Device (SaMD) and are building frameworks for lifecycle oversight: defining what change is permitted, how it is validated, how ongoing performance is monitored, and how to detect degradation or bias after deployment. Real-world performance monitoring and meaningful human oversight become central rather than optional.
The core tension: the same adaptivity that makes a decoder clinically valuable — staying accurate for years without daily recalibration — is what makes it hard to regulate as a fixed product. Overly rigid rules freeze out useful updates; overly permissive ones ship an uncontrolled, drifting device.