AI as SaMD – Understanding the concept
If considered as SaMD, Health Canada classifies it using a risk-based approach, as laid out in the MDR (Schedule 1) and elaborated in the SaMD guidance[2]. The classification depends on the severity of the condition, the role of the software in clinical decision-making, and whether it provides informative vs. directive outputs. More precisely Health Canada defines two axes:
- state of healthcare situation or condition (from critical to serious and non-serious) and
- significance of information provided by SaMD to inform healthcare decision (from treat or diagnose to drive clinical/patient management and inform clinical/patient management)
A critical healthcare situation would be a life-threatening state of health, a major therapeutic intervention or a fragile intended target population (e.g., pediatrics). A non-serious situation would be a slow and predictable disease state, minor (and normally non-invasive) therapeutic interventions or a non-patient target population.
This classification determines the level of regulatory scrutiny and evidence required for deployment. Class IV is the most demanding from a regulatory perspective and requires more comprehensive evaluation, documentation, and quality management, while Class I is the least demanding and can often be addressed within existing institutional research infrastructures.
Examples of SaMD classification in practice
While the principles of SaMD classification may seem abstract, seeing how they apply to real or hypothetical AI can help researchers determine the class of their tool. The following examples are taken from Health Canada’s guidance on SaMD[3], including non-medical device software as well as Class I to Class III examples. Note that, in principle, SaMD can also fall into Class IV, but typically not as a standalone software which is why they are not considered in the following examples.
Read through the examples and check your intuition:
Examples of SaMD Classification in Practice| Example | Class |
|---|
| Software that provides a diabetic patient with simple tools to organize and track their health information. The healthcare professional can input medical information for diabetes-related conditions such as kidney and eye function as well as drug dosage. The software also obtains data from a closed-loop blood glucose monitor. The software analyzes the information collected to provide early diagnosis of a diabetic emergency. When a patient experiences a diabetic emergency, the software will alert the healthcare professional and use the analyzed information to make treatment decisions based on the patient’s unique health profile. | III |
| Software intended for health care professionals that uses an algorithm to analyze patient information, such as blood pressure, heart rate, weight, and age to determine which treatment plan is likely to be most effective in treating the patient's condition. These patient-specific treatment recommendations would otherwise not be available to the health care professional. The healthcare professional cannot independently review the calculations made by the algorithm. | II |
| Software that provides patients with simple tools to organize and track their health information. The information is intended to be shared with a healthcare provider as part of a pre-diabetes management plan. | NA (not a medical device) |
| Software that uses artificial intelligence to tailor mental health treatment plans, including medication dosing, based on patient-derived signals. The software also calculates the probability of remission for a given treatment plan. The software is intended for use in the clinician's office during a routine appointment. | II |
| Software that is intended for use in rehabilitation and active range of motion assessment. This software analyzes images from cameras to generate objective measurements of joint mobility and posture. The data generated by the software is intended to function as a clinical aid in the assessment and rehabilitation of joint dysfunction and posture imbalance. | I |
| Software intended to be used as an annotation tool for the production of diagnostic and prognostic reports of clinical electroencephalogram (EEG) studies, based on a user's traditional visual interpretation of EEG. It is intended to standardize and structure reporting of cl inical EEG according to an international academic standard. The software is also intended to reduce the workload of the reporting doctor by automatically importing minimal pieces of information from an external EEG system. | NA (not a medical device) |
| Chat-based triage software intended to guide users to the most appropriate form of help based on their medical symptoms. The chat-based triage software outcomes are intended to indicate to a user the next best step to take when seeking further health care, and to provide safe and appropriate clinician-vetted advice where appropriate. Triage advice outcomes include self-care, pharmacy, primary care, dental, ophthalmic, sexual health and emergent care advice outcomes. | NA (not a medical device) |
| Radiation treatment planning software that accepts patient images, allows for a healthcare professional to identify and delineate tumours, healthy tissues, and critical organs, calculates a complex plan for how the therapy system will deliver radiation, and then serves as input for medical linear accelerators in the treatment of cancer. | III |
You can find details and further examples in the Health Canada’s guidance on SaMD[3]. These examples illustrate that SaMD classification is often highly nuanced, and that the outcome ultimately depends on the intended use. The intended use of SaMD is reflected across multiple sources, such as labelling, user manuals, and websites. Regulators expect the relevant details to be captured clearly in the intended use statement.
