NMPA issued the “Guideline on Artificial Intelligence Medical Devices (Draft)” on June 4 for feedback. The document is to direct manufacturers on how define, design, and register the AI medical device and conduct the life-cycle management.
The guideline applies to the AI medical devices and IVDs, including the Class II and III AI independent software and medical devices containing AI software components.
The AI medical devices refer to medical devices that use artificial intelligence technology to achieve their intended use based on “medical device data”, such as medical image data generated by:
- medical imaging equipment (X-ray, CT, MRI, ultrasound, etc.)
- medical electronic equipment (ECG, brain, etc.)
- waveform equipment (blood pressure, non-invasive blood glucose, etc.)
- In vitro diagnostic equipment (pathological images, microscopic images, invasive blood glucose waveform data, etc.)
- general equipment (skin photos taken by digital cameras for skin disease diagnosis, and ECG data collected by health electronic products for early warning of heart disease, etc.)
- combined use of non-medical device data (such as electronic medical records, the result of medical inspection reports, etc.)
The guideline pays attention to AI medical device life-cycle process, per the “Good Manufacturing Practice for Medical Device Independent Software and Guideline for On-site Inspection”. It lists the following items:
- Demand analysis
- Data collection
- Algorithm design
- Verification and confirmation
- Update control
For registration application, the document specifies that:
Product name should comply with the requirements of common naming conventions, and usually reflect the characteristics of input data (such as CT images, fundus photos), target disease (including lesions and disease attributes), and intended use (such as auxiliary triage, auxiliary detection, auxiliary diagnosis), etc. word.
Software components generally do not need to be reflected in the information contained in the registration certificate. Its software function names can refer to the requirements for independent software of artificial intelligence. If so, the scope of application reflects the functions of the auxiliary decision-making software.
In the software research materials, the algorithm research report of each artificial intelligence algorithm shall be submitted with the algorithm as the unit. For specific requirements, please refer to Chapter 6.
If the product technical requirements contain performance indicators based on the evaluation database, the basic information of the evaluation database (such as name, model specification, complete version, responsible party, etc.) should be clearly defined.
Performance indicators based on other types of third-party database tests do not need to be reflected in product technical requirements in principle.
For auxiliary decision-making products, the instruction manual should specify the algorithm performance evaluation summary of the artificial intelligence algorithm (basic test set information, evaluation indicators and results), clinical evaluation summary (basic clinical data information, evaluation indicators and results), and the definition of decision-making indicators (or provide References such as clinical guidelines, expert consensus, etc., on which the definition of decision indicators is based) and other information. At this time, if a data-based artificial intelligence algorithm is used, the instruction manual should also supplement the algorithm training summary information (basic training set information, training indicators and results).
If the product adopts the AI black box algorithm, according to the analysis report of the influencing factors of the algorithm, the product use restrictions and necessary warning information shall be specified in the manual.
For products whose software security level is critical, user training materials shall be provided.
Source: https://chinameddevice.com/guideline-on-artificial-intelligence-medical-devices/
