MedTech Terms
    The authoritative reference
    Back to AI/ML Devices Deep Dive
    Software & AIAI / ML in DevicesSoftware LifecycleClinical Evidence

    Foundation Model (Healthcare)

    Large pretrained model adaptable to many downstream clinical tasks via fine-tuning or prompting.

    Reviewed by Christian Espinosa, Founder, Blue Goat CyberLast reviewed September 19, 2026

    Definition

    Healthcare foundation models - for medical imaging, pathology, EHR text, multi-modal patient records - are trained on broad datasets and then specialized. They blur traditional boundaries between intended use and raise novel validation questions. Because the same base model can be adapted to many tasks, a clear intended use and task-specific validation remain essential for any device built on one. Manufacturers also need to manage dependencies on third-party models they do not control, including version changes. The EU AI Act includes separate obligations for providers of general-purpose AI models, which apply alongside high-risk requirements for a device built on top of them.
    What the regulation says
    Regulatory bodies like the FDA and EMA recognize that foundation models, due to their broad applicability and subsequent specialization, present unique challenges that don't always fit neatly into existing Software as a Medical Device (SaMD) frameworks. They may be considered a medical device (or part of one) if they have a medical intended use, as per documents like the FDA's 'Clinical Decision Support Software' guidance.

    What this means in practice

    FDA, MHRA, and Health Canada have flagged foundation models as a regulatory frontier - neither pure SaMD nor pure platform.

    Examples

    • A foundation model initially trained on general medical texts and then fine-tuned to assist with differential diagnosis in radiology would be subject to medical device regulations for its diagnostic aid intended use.
    • A biotech company develops a generic protein-folding foundation model, which is then adapted by a separate entity for predicting targets for a specific disease, thus becoming part of a medical device development.
    • A hospital system uses a foundation model pre-trained on anonymized patient EHRs to optimize resource allocation, which may not be a medical device itself, but if subsequently used for individual patient risk stratification, it could become one.
    Common pitfalls
    • •A common misconception is treating a specialized foundation model as a standalone SaMD without considering the regulatory implications of its underlying general-purpose capabilities.
    • •Failing to document the full lifecycle of a foundation model, including its initial pre-training, fine-tuning, and deployment, can lead to compliance issues.
    • •Underestimating the resources required for continuous monitoring and validation of a deployed foundation model's performance in clinical settings is a significant pitfall.

    Frequently asked questions

    Foundation models are pre-trained on vast datasets for broad applicability, then fine-tuned for specific tasks, unlike traditional AI/ML models often developed for a single intended use from inception. This broad initial training introduces unique considerations for bias, generalization, and continued validation.

    Related terms

    Grouped by theme
    Cited by

    Where this term appears across MedTech Terms.

    Learning paths (1)

    Latest in MedTech

    Sources

    3 sources

    Every citation below opens the original document. Each is graded against our source-tier hierarchy so you can see what rests on binding law versus commentary.

    Tier 2Regulator guidance and consensus· 3
    Link health: 3 verified· last checked 2026-06-20
    FDA·1IMDRF·1MDCG·1
    1. 1
      FDA Digital Health Center of Excellence
      Tier 2 Verified
      FDAfda.gov
    2. 2
      IMDRF - Software as a Medical Device
      Tier 2 Verified
      IMDRFimdrf.org
    3. 3
      MDCG Software Guidance
      Tier 2 Verified
      MDCGhealth.ec.europa.eu

    Inline markers like [1] jump to the matching reference above.