FDA PCCP for AI Medical Devices: Strategic Framework for 2026
Predetermined Change Control Plans let AI/ML medical devices evolve post-market without requiring new 510(k) submissions. Here's how to build a PCCP that FDA accepts.
What Is a PCCP?
PCCP solves a fundamental challenge: traditional regulatory frameworks assume "locked" algorithms. AI/ML algorithms that learn and adapt require pre-approved boundaries for change.
Why PCCP Matters
Without PCCP:
- Every algorithm update requires new 510(k)
- 6–12 months per update cycle
- Frequent submissions delay iteration
- Innovation pace slows
With PCCP:
- Updates within scope: no submission needed
- Updates implemented in days/weeks
- Pace of iteration matches modern AI development
- FDA visibility maintained through Notification
The 3 Components of a PCCP
Component 1: Description of Modifications
What changes can the algorithm make?
Must be specific:
- Quantitative boundaries (e.g., "Sensitivity range: 0.85–0.95")
- Categorical limits (e.g., "Trained only on adult populations")
- Algorithmic constraint types (e.g., "Threshold tuning only; no model architecture changes")
Common mistake: Vague descriptions ("Algorithm may be updated to improve performance"). FDA rejects ambiguity.
Better approach: Specific modification types with quantitative limits.
Component 2: Modification Protocol
How will changes be tested, validated, and monitored?
Must address:
- Data management for algorithm updates
- Validation methodology for each modification type
- Performance acceptance criteria
- Quality assurance for changes
- Documentation of each change
Format: Typically 15–30 pages with detailed test methodology.
Component 3: Impact Assessment
How do changes affect safety and effectiveness?
Must address:
- Risk analysis for each modification type
- Benefit-risk justification
- Monitoring for performance drift
- Rollback procedures
- Post-market performance reporting
PCCP Structure Example
For an AI radiology software (CADx for chest X-rays):
PCCP COMPONENTS:
Modification 1: Sensitivity Threshold Adjustment
- Range: 0.80 to 0.95
- Validation: ROC analysis on 1000+ image validation set
- Monitoring: Monthly performance review
Modification 2: Training Data Expansion
- Scope: Additional images from FDA-cleared institutions
- Validation: Performance maintained within ±2% on locked test set
- Monitoring: Quarterly performance evaluation
Modification 3: Demographic Subgroup Performance Update
- Scope: Performance optimization for previously underrepresented demographics
- Validation: Subgroup analysis showing maintained or improved performance
- Monitoring: Continuous bias assessment
Out of Scope (Requires New Submission):
- Model architecture changes (CNN to transformer, etc.)
- New imaging modalities (chest X-ray → CT)
- New indications (lung nodule → cardiomegaly)
- New patient populations (adults → pediatrics)
How FDA Evaluates PCCP
FDA reviewers assess:
| Criterion | What FDA Looks For |
|---|---|
| Specificity | Quantitative boundaries, not vague descriptions |
| Validation rigor | Adequate test data, methodology, acceptance criteria |
| Safety preservation | Each modification type has documented safety analysis |
| Monitoring adequacy | Detectable performance drift, defined intervention triggers |
| Rollback capability | Documented procedures if modification fails post-market |
| Scope discipline | Clear inside-vs-outside scope criteria |
Building a PCCP — 6-Step Framework
Step 1: Algorithm Lifecycle Planning
- Identify likely modification types over device lifetime
- Categorize by risk and complexity
- Prioritize for PCCP inclusion
Step 2: Modification Type Scoping
- Define specific modification categories
- Set quantitative boundaries for each
- Identify explicit out-of-scope cases
Step 3: Validation Methodology
- Define validation approach per modification type
- Specify data requirements (training set, validation set, test set)
- Set performance acceptance criteria
Step 4: Safety Analysis
- Risk assessment per modification type
- Bias analysis methodology
- Rollback procedures
Step 5: Monitoring Plan
- Real-world performance monitoring
- Drift detection methodology
- Reporting thresholds and intervention criteria
Step 6: Documentation Package
- Compile components into PCCP document
- Cross-reference with 510(k) submission
- Submit as part of premarket submission
Common PCCP Failures
Failure 1: Vague Modification Boundaries
"Algorithm performance may be improved" — too vague.
Fix: "Sensitivity may range from 0.85 to 0.95 based on threshold tuning. Specificity must remain ≥ 0.80."
Failure 2: Insufficient Validation Methodology
"Performance will be validated" — doesn't specify how.
Fix: "Validation on 1,000+ image locked test set, ROC analysis required, AUC ≥ 0.85 acceptance criterion."
Failure 3: No Real-World Monitoring Plan
PCCP allows changes; FDA expects ongoing monitoring.
Fix: Monthly real-world performance review with drift detection thresholds.
Failure 4: Out-of-Scope Boundary Unclear
When does a change require new submission?
Fix: Explicit out-of-scope list (model architecture, new indications, new patient populations).
Failure 5: Bias Analysis Missing
AI bias concerns are top of FDA review.
Fix: Bias analysis methodology with demographic subgroups, fairness metrics, intervention triggers.
PCCP and Other Jurisdictions
Korea MFDS
MFDS announced PCCP-equivalent framework in 2025 guidance. Korean version aligns with FDA approach with Korea-specific monitoring requirements.
EU MDR
EU MDR allows similar concepts but with less prescriptive framework. Risk-based change control under existing MDR provisions.
China NMPA
NMPA AI medical device framework includes change management plans similar in concept to PCCP.
PCCP Cost-Benefit Analysis
| Item | Without PCCP | With PCCP |
|---|---|---|
| Initial submission cost | $50K–$150K | $70K–$200K (PCCP adds 30–40%) |
| Per-update cost | $30K–$80K per 510(k) | $0 (within scope) |
| Time per update | 6–12 months | Days to weeks |
| Annual update capacity | 1–2 updates | 5–10+ updates |
For AI/ML devices with frequent updates, PCCP investment recovers within 1–2 update cycles.
Frequently Asked Questions
Q: Can existing 510(k) cleared devices add PCCP later?
A: Yes, via supplementary submission. However, prospective PCCP inclusion in initial submission is more common.
Q: How specific must the modification boundaries be?
A: Specific enough that FDA reviewer can determine "in scope" vs "out of scope" for any proposed change. Quantitative boundaries are preferred over qualitative descriptions.
Q: Does PCCP cover bug fixes?
A: Typically yes, if bug fixes don't fundamentally change algorithm behavior. Define bug fix scope explicitly in PCCP.
Q: Can we update training data within PCCP?
A: Yes, if training data updates are defined in modification scope. Common boundary: "Additional training data from FDA-cleared institutions, maintained or improved performance."
Q: Does Leanabl help with PCCP development?
A: Yes. Leanabl's Regulatory Pathway Strategy and Korea SaMD Approval services include PCCP development for FDA and equivalent Korean frameworks.
How Leanabl Helps
- Korea SaMD Approval — AI/ML device pathway for Korea
- Regulatory Pathway Strategy — PCCP-aligned submission strategy
- Medical Device Cybersecurity — for AI/ML devices with cybersecurity complexity
Contact Leanabl for PCCP development.
Last updated: 2026-05-15.
Have a regulatory question?
Talk to a Korea regulatory specialist about your device, your timeline, or your next submission.
Talk to a specialist
