In short
- ISO 14971:2019 defines the risk management process for medical devices, including software as a medical device, and requires manufacturers to set objective risk acceptability criteria without prescribing acceptable risk levels.
- For device software, FDA expects a risk management file with a plan, a risk assessment and a report, and it suggests focusing on the failures that lead to hazardous situations, setting the probability of a software failure to 1 when it cannot be estimated realistically.
- Risk controls should follow FDA's order of design, protective measures and information for safety, be traced to requirements and tests, and the file stays open after release because production and post-production information has to feed back into it.
ISO 14971 is the international standard for risk management of medical devices, and it explicitly covers software as a medical device. For a software team, applying it means keeping a risk management file: a plan with acceptability criteria, a hazard analysis tied to the intended use, risk controls traced into requirements and tests, an evaluation of residual risk against benefit, and a report. FDA asks for that file in premarket submissions for device software and refers sponsors to an FDA-recognized version of ISO 14971 for how to build it.
What does ISO 14971:2019 require?
The IEC webstore summarizes ISO 14971:2019 as specifying terminology, principles and a process for risk management of medical devices, including software as a medical device and in vitro diagnostic devices. The process helps manufacturers identify the hazards associated with the device, estimate and evaluate the associated risks, control those risks and monitor the effectiveness of the controls. It applies to all phases of the device life cycle and to risks such as biocompatibility, data and systems security, electricity, moving parts, radiation and usability.
Three scope points matter for software companies. The standard requires objective criteria for risk acceptability but does not specify acceptable risk levels, so you set them. It does not cover business risk management. And it does not require a quality management system, although risk management can be part of one. The current edition is the third, published on December 10, 2019.
What changed in the 2019 edition?
According to the Johner Institute's review of the third edition, the main changes for software teams are these:
- A definition of benefit, meaning a positive impact or desirable outcome of the device's use on the health of an individual, not an economic benefit.
- An explicit requirement to analyze reasonably foreseeable misuse, intentional or unintentional.
- Risks from inadequate data and system security are explicitly in scope, without specific security requirements.
- Separate acceptance criteria are allowed for individual risks and for the overall residual risk.
- More detailed production and post-production requirements, so manufacturers actively collect information and check whether new hazards have appeared and whether risk estimates still hold.
- Many informative annexes moved to the companion document ISO/TR 24971, whose second edition was published in 2020.
How does ISO 14971 fit with IEC 62304?
IEC 62304 defines the life cycle processes for medical device software, whether the software is itself a device or an embedded or integral part of one. ISO 14971 supplies the risk process those life cycle activities depend on. The 62304 safety class, for example, depends on the risk or severity of possible harm, and the Johner Institute recommends deriving the hazards that come from the software system from risk management first, then setting the class. FDA names both, along with ANSI/AAMI SW91 on classifying defects in health software, among the consensus standards it harmonized its software guidance with. In practice, one team keeps one risk file and the software process reads from it and writes back to it.
What goes in the risk management plan?
FDA says that when it reviews a risk management plan, it focuses mainly on two things: the individual risk acceptability criteria, including when risk reduction is needed, and the method for evaluating the overall residual risk after all controls are implemented and verified. The criteria should be written into the plan before the first risk evaluation of the software under review, and the plan should make clear how the overall residual risk will be weighed against the benefits of the intended use.
A practical plan for a software product also names the scope, meaning the device, its intended use and the releases covered, the people responsible and the reviewers, the severity and probability scales and the matrix that combines them, the verification activities for controls, and how production and post-production information will be collected.
How do you analyze hazards in software?
FDA's software guidance lists what a risk assessment should document, often in a table. It starts with the known or foreseeable hazards and their causes, based on the intended use, reasonably foreseeable misuse and the impact on safety, treatment or diagnosis. For each hazard, you trace the sequences or combinations of events that lead to a hazardous situation, estimate the risk and record the severity of the harm. FDA uses three terms here: a hazard is a potential source of harm, a hazardous situation is a circumstance in which people, property or the environment are exposed to a hazard, and harm is injury or damage to the health of people, or damage to property or the environment.
Estimating probability is harder for software than for hardware. FDA notes that it is often difficult to estimate the probability of software failures, and that unrealistically low estimates lead to inappropriate risk controls. It suggests focusing on identifying the software functionality and failures that could result in hazardous situations, and, where probability is used, considering the worst case and setting the probability of the software failure occurring to 1. The Johner Institute adds a practical method: go through the system requirements and ask whether there is a risk if each one is not met, or not met to the specified extent.
For a typical software device, that review covers wrong outputs, late or missing outputs, outputs attached to the wrong patient, displays that users can misread, failures at interfaces with other systems, known anomalies in third-party components, and compromise through inadequate security.
Which risk controls count, and in what order?
FDA lists risk control measures in descending order of priority:
- Design: removing unnecessary features, changing the software architecture to prevent hazardous situations, or changing the user interface to prevent use errors.
- Protective measures: defensive programming checks that detect unexpected faults and automatically halt the delivery of results or therapy, or alarms that let the user intervene.
- Information for safety: written warnings, on-screen warnings and training.
Each control needs two verifications: that it was implemented, and that it actually reduces the risk. FDA describes doing this by tracing each hazard to its controls, for example a hazard HAZ-001 to a requirement in the software requirements specification, a design element in the software design specification and the unit, integration and system test cases that cover it. Because one hazard can trace to many requirements and tests, the guidance allows that traceability to live in a separate document. You also assess whether a control introduces new hazards or changes the initial risk evaluation.
How do you evaluate residual risk and benefit?
After controls are in place, each hazardous situation gets a residual risk evaluation against the acceptability criteria in the plan. If a residual risk is not acceptable and further control is not possible, FDA expects a documented benefit-risk analysis showing that the benefits of the intended use outweigh it. The overall residual risk of the device is then evaluated with the method the plan defined, which under the 2019 edition can use different criteria from those for individual risks.
The risk management report comes last. FDA says it should show that the plan was implemented, that appropriate personnel assessed the file and found the overall residual risk acceptable, and that methods are in place to collect and assess relevant production and post-production information.
What about AI models and security risks?
For AI-enabled devices, FDA's guidance on predetermined change control plans tells manufacturers to follow their risk management processes when they write the protocol for future model changes, and cites AAMI CR34971:2022, a consensus report on applying ISO 14971 to artificial intelligence and machine learning. The impact assessment in such a plan uses the ISO 14971 concept of harm, and asks for the risks of each planned change, including unintended bias.
For security, ISO 14971:2019 brings data and systems security risks into scope, and FDA's documentation level assessment counts the likelihood that device functions are compromised by inadequate cybersecurity. FDA covers the security content of a submission in its separate cybersecurity guidance, so plan for a separate security risk process whose results feed into the safety file.
Why does the risk file stay open after release?
The 2019 edition expects manufacturers to collect and review production and post-production information and act on it: new hazards, wrong estimates, risks that are no longer acceptable. For a software product that means connecting the risk file to the places where that information shows up: complaints and support tickets, field defects, anomaly reports from third-party components and monitoring of model performance. Every release should end with a review of the risk file as well as a changelog.
What do software teams get wrong?
- Writing acceptability criteria after the analysis. FDA expects them in the plan before the first evaluation.
- Assigning very low probabilities to software failures to make risks look acceptable.
- Relying on warnings and training first, when design changes and protective measures come earlier in the order.
- Listing controls without verifying that they reduce the risk, or without tracing them to tests.
- Treating the file as a submission document and closing it at release.
- Leaving reasonably foreseeable misuse and security out of the hazard analysis.
Building a risk file for your product?
We build medical device software under an IEC 62304 process and produce the documentation an FDA submission needs, including the requirements, architecture, traceability and test evidence a risk file depends on. The manufacturer owns the risk file and the submission. Onix Systems has built software since 2000 and has 370+ engineers. Talk to one of our engineers, or start with our free five-day technical audit, which ends with a written report on where your risk file and software records stand.
Sources
- ISO 14971:2019, Medical devices: Application of risk management to medical devices (IEC webstore)
- Johner Institute: The third edition of ISO 14971
- FDA: Content of Premarket Submissions for Device Software Functions (June 2023)
- FDA: Content of Premarket Submissions for Device Software Functions, final guidance PDF
- IEC 62304:2006+AMD1:2015, Medical device software: Software life cycle processes (IEC webstore)
- FDA: Marketing Submission Recommendations for a Predetermined Change Control Plan for AI-Enabled Device Software Functions
- Wikipedia: ISO 14971 (editions and ISO/TR 24971)
- Johner Institute: Safety class according to IEC 62304