Locked Algorithm

The core characteristic of a locked algorithm is the freezing of all learned parameters post-training. Once the model is validated and cleared, the mathematical weights that define its inferences are fixed. This prevents concept drift and ensures that every patient scan is analyzed with the exact same logic, guaranteeing diagnostic reproducibility across different sites and time periods.

In the FDA's regulatory framework, the locked state defines the boundary of the cleared medical device. Any change to the model's parameters or architecture crosses this boundary, transforming the software into a new device that requires a fresh marketing submission. This is distinct from an adaptive algorithm, which can change its behavior under a Predetermined Change Control Plan (PCCP) .

Locked algorithms demand rigorous version control as part of the Design History File (DHF) . Each released model must be tagged with a unique version identifier that maps to the exact training data, hyperparameters, and evaluation metrics submitted to regulators. This traceability is critical for post-market surveillance and auditing, allowing manufacturers to instantly identify which algorithm version produced a specific clinical result.

The locked nature simplifies Verification and Validation (V&V) activities under IEC 62304. Because the algorithm does not change, the analytical and clinical validation studies represent a static snapshot of performance. The V&V documentation must prove that the frozen model meets its Intended Use Statement and Indications for Use without the risk of post-deployment performance degradation due to self-modification.

A locked algorithm presents a stable target for cybersecurity risk assessment. The software's integrity can be verified through cryptographic hashing of the model file, ensuring it has not been tampered with or corrupted. This immutability is documented in the Software Bill of Materials (SBOM) , providing hospitals and IT departments with a fixed, auditable asset that does not introduce unexpected behavioral changes into the clinical network.

To improve a locked algorithm, a manufacturer must initiate a new regulatory submission, typically a 510(k) Premarket Notification if a predicate exists, or a De Novo Classification Request for novel devices. The submission must demonstrate Substantial Equivalence (SE) to the previous version or provide new evidence of safety and effectiveness. This contrasts sharply with a PCCP, which allows pre-authorized updates without a new submission.

An FDA-authorized plan detailing the specific types of modifications a manufacturer intends to make to a machine learning-enabled device without requiring a new marketing submission. A PCCP is the primary regulatory mechanism that allows a locked algorithm to be safely updated and unlocked in a controlled, pre-reviewed manner. It must define the scope of planned changes, the methodology for implementing them, and the verification and validation protocols to ensure continued safety and effectiveness.

A machine learning model whose behavior changes over time based on new data inputs, standing in direct contrast to a locked algorithm. An adaptive model continuously learns from real-world use, which introduces a dynamic performance profile. Without a Predetermined Change Control Plan, any modification to an adaptive algorithm would technically require a new regulatory submission, making the PCCP essential for the legal deployment of continuously learning diagnostic systems.

The most common FDA submission pathway for a locked algorithm classified as a Class II medical device. The manufacturer must demonstrate Substantial Equivalence (SE) to a legally marketed predicate device. For a locked AI model, this involves proving that the frozen model's performance, intended use, and technological characteristics are as safe and effective as the predicate. Any unlocking of the algorithm post-clearance would typically require a new 510(k) unless governed by an approved PCCP.

The overarching regulatory category for a locked algorithm intended for diagnostic purposes. SaMD is defined by the International Medical Device Regulators Forum (IMDRF) as software intended to be used for one or more medical purposes that performs those purposes without being part of a hardware medical device. A locked diagnostic AI model fits this definition perfectly, and its development lifecycle is governed by standards like IEC 62304 for software development and ISO 14971 for risk management.

A compilation of records that describes the design history of a finished medical device. For a locked algorithm, the DHF is the definitive proof that the model was developed according to an approved design plan. It must contain the frozen model architecture, the final training dataset specifications, the results of Verification and Validation (V&V) testing, and the formal risk analysis. The DHF is the auditable artifact that demonstrates the algorithm was intentionally locked at a specific, validated performance point.

The combined processes that provide objective evidence that a locked algorithm meets its specified requirements. Verification confirms that the final, frozen model's design outputs match its design inputs (e.g., the model architecture matches the specification). Validation confirms that the locked model meets user needs and intended uses in a clinical context (e.g., diagnostic accuracy meets the required sensitivity and specificity). The V&V report is a cornerstone of a regulatory submission for a locked SaMD.