AI/ML in Medical Devices: Planning Safe Model Updates with PCCPs and Iterative Development

This article continues a series on how agile-inspired practices, adapted for hardware and regulated environments, help medical device teams deliver better outcomes. As artificial intelligence and machine learning (AI/ML) become more prevalent in medical devices, teams face a practical challenge: how to improve models over time without introducing risk or triggering repeated regulatory delays.

Predetermined change control plans (PCCPs) provide a path forward. They allow pre-authorized updates when boundaries, validation methods, and evidence are clearly defined upfront. The challenge is operationalizing PCCPs in a way that keeps pace with how models actually evolve.

An iterative learning and execution approach—such as the MAHD Framework—offers a practical solution. It provides structure to define guardrails early, monitor performance continuously, and generate the evidence needed to support safe, compliant updates.

PCCPs Only Work When Boundaries Are Clear

At the core of any PCCP is a simple idea: certain changes can be anticipated, evaluated, and approved in advance, as long as they stay within defined limits.

For AI/ML systems, this may include updating model weights, adjusting thresholds, expanding datasets within a defined population, or improving performance within known conditions. Without clear boundaries, PCCPs are difficult to defend. With overly narrow boundaries, they lose value.

An iterative approach helps teams strike the right balance. Rather than attempting to define every possible change upfront, teams refine boundaries as they learn more about model behavior, variability, and real-world use.

A Simple Example to Visualize an Iterative Approach

Consider a connected health monitoring platform that analyzes patient data to detect early signs of cardiac irregularities. The team wants to improve the model over time by incorporating new data, reducing false positives, and improving sensitivity.

Each change can improve performance but also introduces risk. A small shift in behavior can affect clinical interpretation or user trust. PCCPs help define what is acceptable, but the development approach determines whether updates can be managed safely.

To make this more concrete, consider how a PCCP might define acceptable boundaries for a model update in this system. For example, the team may pre-authorize updates that retrain the arrhythmia detection model using additional patient data from the same intended population, provided that:

• Sensitivity for detecting clinically significant arrhythmias does not drop below an established threshold (e.g., ≥95%)
• False positive rates remain within a defined range (e.g., no more than a 10% increase relative to baseline)
• Performance remains stable across predefined patient subgroups (e.g., age ranges or signal quality bands)
• No new input signals or data sources are introduced
• Model architecture remains unchanged, with updates limited to parameter tuning

Changes such as introducing new sensor inputs, expanding to new
populations, or altering how outputs drive clinical decisions fall outside the PCCP and require additional review. Defined this way, the PCCP becomes practical. It enables improvement while maintaining predictable behavior and regulatory confidence.

How Traditional Development Handles Model Updates

In a traditional model, the model is trained, validated, and documented early. Updates later often require re-validation, re-documentation, and renewed regulatory review. As a result, teams delay updates or bundle them into larger releases.

This slows improvement, increases risk, and creates a reactive process where evidence is gathered late instead of continuously.

How an Iterative Approach Changes the Dynamic

An iterative approach treats model development as ongoing. PCCPs become part of a living system rather than a static document. Each cycle connects planning, execution, and evaluation. Teams ask:

• What change are we introducing?
• Does it fall within PCCP boundaries?
• How will we measure its impact?
• What evidence do we need?

This enables smaller, lower-risk updates and builds confidence over time.

Tracking PCCP Activities Within Iterative Development

This approach is reflected in how work is structured using an iterative approach, such as the MAHD Framework. As shown in Figure 1, with each iterative cycle, teams align design, risk, regulatory, and validation activities with a shared, flexible cadence. PCCP-related work—defining boundaries, monitoring behavior, and validating updates—is tracked alongside design artifacts and system verification.

Within this structure:

• Early iterations define initial boundaries and metrics
• Mid-stage iterations refine them based on observed behavior
• Later iterations generate consistent validation evidence

This alignment ensures model updates, risk controls, and regulatory evidence progress together rather than being reconciled late.

Monitoring Data and Model Health

AI/ML risk is not just about the initial model but how it behaves over time. Data shifts, user behavior changes, and performance can degrade.

In an iterative model, monitoring and improvement become routine. Teams track data characteristics, performance metrics, and emerging error patterns, identifying issues before they become systemic.

Validating Updates Against Defined Metrics

PCCPs depend on clear metrics such as sensitivity, specificity, false positive rates, and stability across populations. In the MAHD approach, these are defined early and reused consistently. Each update is evaluated against the same criteria, creating a growing body of evidence.

Separating Safety-Critical Logic from Model Tuning

Well-designed systems separate safety-critical logic from model outputs. Deterministic controls handle safety, while the model supports detection or prioritization. This allows updates within defined boundaries without compromising core safety functions.

In a monitoring system, the model may flag potential issues, while the system enforces thresholds or clinician review before action.

Conclusion

A PCCP is only as strong as its evidence. In traditional models, evidence is assembled late. In an iterative model, it builds naturally. Each cycle produces a consistent set of artifacts: change description, data used, performance results, and assessment against PCCP boundaries. AI/ML devices must improve over time while maintaining safety and regulatory confidence. PCCPs provide the mechanism, but they require disciplined execution from early concept through launch and field updates.

An iterative learning approach—supported by frameworks like MAHD—makes this practical. Teams define boundaries, monitor performance, validate updates, separate safety-critical functions, and build evidence as part of normal work.

The result is a system that improves in a controlled, transparent way, where each update is understood, measured, and aligned with the intent of the product.

MORE FROM THIS AUTHOR: Cybersecurity in Medical Devices: Enhancing SBOM & Threat Controls with Iterative Development

Dorian Simpson is an innovation, product management, and agile consultant, trainer, and speaker. He is the author of The Savvy Corporate Innovator and founder of the Modified Agile for Hardware Development (MAHD) Framework. He helps startups to Fortune 500 technology leaders build skills to improve their ability to identify, evaluate, plan, and develop innovative products. The MAHD Framework is a purpose-built agile approach for physical product innovation. MAHD combines agile principles with hardware-ready methods: On-Ramps to set strategic intent, IPAC Iterations to integrate and learn quickly, Aligned Backlogs to connect work across disciplines, and hardware-aligned roles that empower technical leaders. Organizations adopting MAHD report faster time-to-market, improved compliance confidence, and higher ROI—without sacrificing safety or quality. Learn more at www.MAHDFramework.com.