AI/ML-Based Medical Devices — Emerging Regulatory Questions

Traditional Medical Device regulation assumes a fixed product: what’s approved is what ships, indefinitely, until a formal change is filed and reviewed. Adaptive AI and machine learning algorithms break that assumption by design — a model that continues learning after deployment is, in a meaningful sense, a different product than the one initially validated, potentially every time it updates. Regulators worldwide are still working out how to govern that shift, and India is no exception, currently regulating AI/ML devices through frameworks that were not built with continuous learning in mind.

The Global Reference Point: Predetermined Change Control 

The FDA’s predetermined change control plan concept has emerged as an influential model for resolving this tension: rather than requiring a new approval for every algorithm update, manufacturers pre-specify the types of modifications a model may undergo — the scope of retraining, the performance boundaries within which the model may drift, the triggers for re validation — and the validation protocol governing those changes, allowing regulated evolution within defined, pre-approved bounds. 

This reframes the regulatory question in a fundamentally useful way: instead of asking “is this the same product” after every update, regulators and manufacturers agree in advance on “does this change fall within a pre-approved, monitored envelope.” It shifts oversight from a series of discrete approval events to something closer to continuous, structured governance — a model better suited to how AI products actually evolve.

Validation Challenges Unique to Adaptive Models

Continuously learning models raise validation questions that static devices don’t face. Performance drift over time needs active monitoring, since a model’s real-world accuracy can degrade gradually as the population it encounters shifts away from its original training distribution — a change that may not be obvious without dedicated tracking. Training data shifts need to be detected and controlled for, particularly when a model is retrained on newly collected field data that may differ systematically from its original dataset. And bias -particularly relevant given India’s population diversity across ethnicity, geography, healthcare access, and disease prevalence patterns — needs to be assessed across demographic and clinical subgroups on an ongoing basis, rather than validated once against a single, potentially unrepresentative dataset at the time of initial approval.

 Data provenance compounds all of this. Where and how training data was sourced, and how representative it is of the Indian patient population specifically, materially affects the credibility of an AI-based device’s performance claims. A model trained predominantly on data from other geographies and demographic profiles may perform very differently once deployed against India’s actual patient population — a gap that pre-market validation alone may not fully reveal. 

Where India Stands 

CDSCO has now issued its Guidance Document on Medical Device Software (CDSCO/MD/GD/MDSW/01/2026), providing India’s first comprehensive regulatory guidance for Medical Device software, including software incorporating AI/ML functionality. While the guidance does not establish a separate approval pathway for adaptive AI systems, it introduces expectations around software lifecycle management, risk management, validation, cybersecurity, post-market surveillance, continuous performance monitoring, and documentation of software modifications and algorithm changes that are directly relevant to AI-enabled devices. The guidance also highlights the importance of real-world performance monitoring and appropriate evaluation of AI-based software, including consideration of risks associated with data quality, model performance, and intended use.

Although the MDSW Guidance introduces important expectations relating to AI-enabled software, India has not yet established a fully developed regulatory framework for continuously learning or autonomous adaptive algorithms comparable to the FDA’s Predetermined Change Control Plan approach. Questions remain regarding the extent of pre-approved algorithm modifications, acceptable boundaries for model retraining, management of performance drift, and regulatory oversight of AI systems that evolve after deployment. 

CliniExperts is tracking the implementation of CDSCO’s Medical Device Software Guidance (CDSCO/MD/GD/MDSW/01/2026) and emerging regulatory approaches to AI/ML-enabled Medical Devices. We assist manufacturers in developing validation, change-management, clinical evaluation, cybersecurity, and post-market surveillance strategies that satisfy current CDSCO expectations while remaining adaptable to future AI-specific regulatory requirements.