Scientific Advisory Committee Digital Health Technologies, May 25 and 30, 2022, summary of proceedings

On this page

• Attendance
• Welcome
• Chair's remarks
• Summary and general considerations
• Presentations
• Discussion
• Closing remarks

Attendance

Committee members

Core members: May 25 - Joseph Cafazzo (Chair), Aviv Gladman, Trevor Jamieson, Doug Manuel, Kendall Ho

May 30 - Joseph Cafazzo (Chair), Trevor Jamieson, Doug Manuel, Kendall Ho, Kim Hanson, Kumanan Wilson

Ad hoc members: May 25 - Jared Perry, Chris Simotas, Emil Sosnowski, Adnan Sheikh

May 30 - Emil Sosnowski, Frank Rybicki, Christo Simotas

Regrets: May 25 - Kim Hanson, Kumanan Wilson

May 30 - Aviv Gladman

Presenters

Other: May 30 - Jaron Chong

Health Canada: May 25 - Andrew Smith, Marc Lamoureux

May 30 - Janet Hendry

Observers

Health Canada: May 25 - David Boudreau, Patrick Assouad, Ian Glasgow, Justin Peterson, Martina Buljan, Renate Kandler, Daniel Martire, Gregory Jackson, Julie Polisena, Andrea Katynski, Tudor Fodor, Janet Hendry, Emanuela Fedele, Kinga Michno, Colin McCurdy, Brian Dowling, Julien Proulx, Weimin Zhao

May 30 - David Boudreau, Ian Glasgow, Justin Peterson, Martina Buljan, Renate Kandler, Daniel Martire, Gregory Jackson, Julie Polisena, Emanuela Fedele, Kinga Michno, Colin McCurdy, Brian Dowling, Julien Proulx, Daniel Yoon, Marc Lamoureux, Panyada Phandanouvong, Tyler Dumouchel, Chad Sheehy

Other: May 25 - Joanne Kim (Canadian Agency for Drugs and Technologies in Health, or CADTH), Yannick Auclair (Institut national d'excellence en santé et en services sociaux, or INESSS), Glenne Grossman and Angela McAllister (Canadian Centre for Cybersecurity)

May 30 - Joanne Kim (CADTH), Yannick Auclair (INESSS)

Welcome

David Boudreau, Director General, Medical Devices Directorate (MDD), welcomed committee members. He then introduced the meeting topics:

• 3D printing for medical use
• medical device cybersecurity
• artificial intelligence (AI) transparency

He explained the process for the meeting, thanked the committee members for their time and for providing advice to Health Canada, and gave an overview of recent work being done by MDD.

Chair's remarks

Dr. Joseph Cafazzo, Chair, thanked members for participating in the meeting. He introduced the panel members and gave an update on declarations of affiliations and interests from those initially declared. There were none that restricted SAC-DHT members from participating.

Summary and general considerations

Marc Lamoureux, Manager, Digital Health Division, MDD, gave an overview of previous committee meetings and advice used.

Presentations

There were 4 presentations:

1. Andrew Smith, Medical Devices and Clinical Compliance Directorate, Regulatory Operations and Enforcement Branch (May 25):

• exploring oversight of 3D printing medical devices at point of care

2. Marc Lamoureux, Manager, Digital Health Division, MDD (May 25):

• introduction to cybersecurity

3. Janet Hendry, Senior Evaluator, Digital Health Division, MDD (May 30):

• overview on transparency

4. Dr. Jaron Chong, diagnostic radiologist in body imaging, Western University (May 30):

• clinician perspective on transparency

Presentation 1: Exploring oversight of 3D printing medical devices at point of care

Andrew Smith gave an overview of the traditional medical device manufacturing oversight model and described how 3D printing at point of care presents a new manufacturing paradigm. This emerging technology has the potential to offer innovative and personalized treatment. However, point-of-care manufacturing of medical devices in hospitals is not covered by the traditional regulatory framework that regulates sales between 2 parties. As such, there are some challenges in interpreting the regulations.

Mr. Smith acknowledged that health care facilities engaging in 3D printing at point of care may not be as experienced or familiar with Health Canada's regulatory framework. Any medical devices 3D printed at facilities should be of the same quality as those produced through traditional manufacturing processes. He also discussed the importance of increasing safety through federal oversight, without creating unnecessary burden or redundancy where provincial oversight may already exist.

Presentation 2: Introduction to cybersecurity

Marc Lamoureux gave an overview of Health Canada's activities in medical device cybersecurity as a member of the International Medical Device Regulators Forum (IMDRF). IMDRF provides an opportunity for international members to agree on regulatory topics, including medical device cybersecurity. Health Canada's work on legacy medical devices and software bill of materials at the international level is ongoing.

A legacy medical device cannot be reasonably protected against current cybersecurity threats. Manufacturers of software-enabled medical devices should implement planned end-of-support dates and publicly communicate those dates as early as possible. Before this date, manufacturers should give users as much information as possible to maintain cybersecurity, should the user wish to do so.

End-of-support dates would make it possible for health care providers to implement plans to:

• upgrade or decommission the medical device or
• ensure that they can control the risks of operating the device without support

A software bill of materials is a list that identifies each software component by name, origin, version and build. It includes any commercial, open source or off-the-shelf software components that are part of the medical device (similar to an ingredients list on food packaging). One vulnerable component will affect the final device, and potentially even the facility.

The IMDRF working group is looking at requiring manufacturers to provide an exhaustive software bill of materials to the regulator and health care facility for each medical device that contains software. Facilities should consider creating an internal software bill of materials repository for all their medical devices.

Presentation 3: Transparency overview

Janet Hendry gave an overview of relevant terminology and regulatory concepts. Transparency is the degree to which appropriate information about a device is clearly communicated to stakeholders and is important to consider when assessing a machine learning-enabled medical device. Transparency means that end users can make informed decisions and use the device properly. It also helps regulators and manufacturers evaluate and monitor performance, and fosters trust and confidence in the technology.

Health Canada gave committee members an advance draft of the upcoming guidance for machine learning-enabled medical devices to get input on:

• the transparency considerations in the draft guidance
• labelling expectations for machine learning-enabled medical devices
• how Health Canada can support and promote transparency beyond labelling

Presentation 4: Clinician perspective on transparency

Dr. Chong gave an overview of transparency of machine learning-enabled medical devices from the clinician perspective. He reviewed the notion of "explainability," but noted that health care providers may not always need or have access to an explainable model. In many cases, clinicians may need to act without fully understanding the model.

He also discussed including checklists for the elements that constitute the minimum for a "transparent" device. There are many different checklists available. However, due to the variety of applications being created, there is no agreement on their composition and the minimum information that may be acceptable.

Dr. Chong agreed with many of the transparency concepts set out in the draft guidance. He also reviewed several important concepts, including model cards, which are similar to a product monograph for drugs in Canada.

Discussion

3D printing

Science question 1:
"In your opinion, are there gaps in oversight related to the 3D printing of devices for a medical purpose, printed at the point-of-care?"

The committee identified gaps in oversight related to the 3D printing of medical devices for a medical purpose, when printed at point of care:

• quality assurance
  • 3D printing of medical devices has increased significantly in recent years, especially following pandemic-related stresses on the supply chain
  • hospitals may now be able to 3D print a similar device at a lower cost or that is unavailable but without the safety and quality assurance provided by federal oversight
• health care facilities printing many low-risk medical devices, unaware of medical device establishment licensing (MDEL) requirements
• health care facilities not able to track 3D-printed medical devices that are recalled
• 3D printing labs at hospitals not accredited, unlike diagnostic imaging labs
• lack of teaching expertise and gap in curricula
• inconsistent post-processing procedures among health care facilities
• lack of regulation in Canada on the use of 3D-printed jigs and guides as well as metal 3D printing
• environmental cost
• the materials used in 3D printing of medical devices at point of care do not undergo the same oversight as they would for typical devices
• gap in software used for 3D printing
  • generally, the software programs used for 3D printing of medical devices have not been approved

Science question 2a:
"Do you have suggestions for how Health Canada could achieve this in a least burdensome way? For example, through registrations or notification (similar to how we regulate other therapeutic products at point of care, such as blood, tissues and organs)."

The committee generally supported registration or notification as oversight approaches for 3D printing activities. For example, Health Canada could outline recommendations for post-processing and sterilization, and health care facilities could be required to attest to meeting those requirements. Provinces and territories could require that 3D printing labs be accredited.

The committee suggested that regulatory oversight should be in proportion to the risk level of a medical device. Health Canada should at least inform health care facilities that are engaging in these activities of the MDEL requirements and application process.

The committee also noted that health care facilities in Canada have different levels of resources and personnel. Larger organizations may be better equipped to have proper quality management programs, enforce standards and achieve accreditation. But they cannot assume the level of liability that regulated manufacturers can.

Science question 2b:
"Would these recommendations change if a health care facility manufactures medical devices via 3D printing for use within the same health care facility versus if they were being distributed to a separate health care facility?"

The committee discussed the possibility of health care facilities manufacturing a 3D-printed device for another facility. If this situation were to arise, Health Canada could consider having a third party evaluate the facility to ensure that quality management systems are in place. The facility would be required to meet the same standards as a medical device manufacturer to ensure the safety, effectiveness and quality of the device. This may not be feasible.

Cybersecurity

Legacy medical devices

Science question 3:
"Do you support manufacturers providing clear end-of-support dates for medical devices?"

The committee was generally in favour of the requirement for manufacturers to provide clear end-of-support dates for medical devices. Manufacturers should plan for this at the beginning of the medical device lifecycle.

Science question 4:
"What should manufacturers be requested to provide to users of devices nearing the end-of-support date?"

The committee made the following suggestions:

• Minimize legacy infrastructure, such as outdated operating systems.
  • Manufacturers should allow for underlying third-party operating systems or platforms to be updated without compromising the medical device, to ensure the security of the health care facility's overall network.
• Encourage manufacturers to provide configurations to operate the device with limited or no network access.
• Encourage manufacturers to integrate mechanisms so they can back up and restore configurations, to make it possible to rebuild safely.
• Encourage manufacturers to ensure their complex medical devices can support security monitoring software.
• Ensure that manufacturers are responsible for reporting, disclosing and patching any critical vulnerabilities they become aware of after the end-of-support date within a certain timeframe.

Software bill of materials

Science question 5:
"Do you support manufacturers providing a comprehensive software bill of materials for each medical device that contains software?"

The committee advised that requiring manufacturers to provide an exhaustive software bill of materials could allow regulators to trace any vulnerabilities through the supply chain and over the product's lifecycle. Software bill of materials are being used more frequently by industry and can be automatically generated as software is being built, making it less burdensome for manufacturers.

The committee advised against publishing a comprehensive software bill of materials, as this information could be used for criminal reasons. As medical devices contain regulated software, there are more restrictions on updating them. Health Canada should consider whether manufacturers would be able to update their devices in a timely manner in the event of a vulnerability or if publishing this information would make them a target.

Also, industry stakeholders may have concerns about publishing confidential business information.

Science question 6:
"Discuss the advantages and disadvantages of an exhaustive software bill of materials (compared to an abbreviated list of components that is not exhaustive). Can you comment on what Canadian health care institutions would prefer?"

The committee advised that an exhaustive list of software bill of materials could lead to health care facilities receiving these bills for hundreds of devices. Facilities would probably not be in a position to review and assess all of them. The committee said that active monitoring of the cybersecurity of health care facility networks is more important for identifying risks. Also, smaller companies may not have the resources to assemble an exhaustive software bill of materials.

Future work

The committee recommended that Health Canada explore the issue of electronic health records. In recent years, larger electronic systems have begun to incorporate machine learning-enabled medical devices, which can be high-risk and require frequent patching.

Transparency

Science question 7:
"Discuss the general labelling considerations included in the draft guidance document on machine learning-enabled medical devices."

The committee felt that the section on transparency and labelling expectations was comprehensive and aligned well with existing AI/machine learning checklists.

The committee made the following points:

• Labelling should be explicit about the precise scope of the software to allow clinical reproduction of performance claims. Items such as warnings, contraindications and population (genders or ethnicities, for example) are important and should be at the front of the label. End users (patients or clinicians) should be able to easily identify the criteria for use.
• Data, such as population, sample, ground truth, labels and splits, are as important as a model's technical characteristics.
• A description of how the machine learning system works is helpful in many cases, but may not always be available or relevant, especially when there's strong evidence that a product is effective. In contrast, when evidence is weak, clinicians are more dependent on the learning mechanism as this may impact clinical decision-making.
• The information provided via the software interface help users interpret each output. The description of the machine learning system output and how it should be used clinically will increase confidence in clinical decisions.
• Subgroup analysis and disaggregated data are important. External validation in different data sets is helpful for confirming the data in a real-world setting, but may not always be feasible.
• Inclusion and exclusion criteria are important, as manipulation of these criteria can affect the apparent performance of an AI/machine learning system. Clinicians should be able to determine if the effects are being driven by the data set or the model.

The committee also suggested that Health Canada consider adding the following information in the guidance document:

• the minimum skillset that a user should have to be able to use the device safely, as AI/machine learning devices may be used by a wide range of users (for example, primary care doctor to specialist)
• estimates of the range within which performance is expected, as this range establishes thresholds for action and remediation
• failure mode analyses results that provide warning signs for the user to look out for
• transparency about whether synthetic data sets were used instead of real-world data
• real-world evidence if possible (may not be a feasible standard for all manufacturers to meet)
• saliency maps to help users interpret output, in cases where there is a user involved in the workflow to interpret or disagree with the output
• a figure that shows the user interface

The committee also suggested that Health Canada consider:

• being mindful of instructions about "clinical workflow" information, as a clinician may not always be involved in the use of these devices
• include instructions for use (such as calibration, monitoring) rather than describe how the technology works
• how the labelling should be structured to highlight the most important information for the purposes of safety, as the average user may feel overwhelmed by information around training and test data sets and may not read the most important aspects of the labelling
• the transparency of the decision support that the machine learning system is providing, as decision support tools may have very different performance when used at different sites

Science question 8:
"Discuss what could be pre-market expectations tailored for machine learning-enabled medical devices labelling, such as:

• minimum acceptable level of information that Health Canada should require in the labelling for all machine learning-enabled medical devices
• abbreviated approaches (such as model facts, model cards) versus detailed descriptions of model development and performance (for example, pros and cons, strengths and weaknesses)
• publication citations in the user manual or instructions for use if the relevant information is presented in a peer-reviewed article
• device-specific transparency expectations dependent on device risk, intended user or intended use"

There are general principles and elements that are common between labelling checklists. Creating a universal checklist can be difficult due to the variety of applications. Determining which components should be optional and which mandatory will likely evolve as we learn more about these devices and reviewers gain more experience.

The committee talked about the significant limitations associated with citing publications in user materials. Although publications could be included as references, they are not enough on their own. These publications are intended for a different audience and manufacturers should reframe this information for the intended reader. It's not likely that patients and clinicians will seek out publications.

Transparency expectations may vary depending on the intended use of the device. For example, flexibility in terms of labelling may be permitted for alerting tools and decision support tools. However, autonomous diagnostic tools (currently, none are authorized for sale in Canada) should have more stringent labelling requirements and undergo a higher level of scrutiny. Some committee members cautioned that even algorithms intended to support decisions can lead clinicians to make decisions that may not be appropriate, especially when updates happen automatically without the user being alerted.

It was suggested that labelling should take into account its audience and distinguish between patient-facing, primary care-facing and hospital-facing products. For example, patient-facing labelling may require a different reading level or terminology than clinician-facing labelling. Also, independent professionals may not have the same tools to vet a system as a large hospital would.

Science question 9:
"Discuss how Health Canada can support and promote machine learning-enabled medical devices transparency, beyond labelling expectations. Are there other potential methods of providing patients and users with access to transparency information, besides the traditional user manual or instructions for use?"

The committee suggested there may be an opportunity to go beyond traditional labelling (for example, user manual or instructions for use) with machine learning-enabled medical devices. Because the user interacts with the technology through a digital interface, risk information could be presented "just in time" through this interface. Examples are built-in advisories and safeguards to prevent the technology from being used in an inappropriate population. Manufacturers should be encouraged to enforce contraindications in this way whenever possible rather than relying on labelling and manuals.

From a patient perspective, the committee said there can be risks associated with having a lot of user manuals. This type of information may not always have the intended impact. Manufacturers should be encouraged to consult with people with lived experience to gain their support in using these devices.

Closing remarks

Dr. Joseph Cafazzo, Chair, thanked the members for their participation before closing the meeting.