Current Situation of Approval of Artificial Intelligence Medical Device Registration Certificate in China
»Development status of AI medical device industry from 2019 to 2021
Market Scale
The scale of domestic AI medical device market has increased significantly from 125 million yuan in 2019 to 292 million yuan in 2020, with a year-on-year increase of 134%. It is expected to continue to increase significantly in 2022, reaching 1.156 billion yuan.
»Current situation of approval of AI medical device registration certificate in China
Analysis on the current situation of AI medical device registration certificate approval in 2021
According to the report of the State Drug Administration (NMPA) on medical device registration in 2021, as of December 31, 2021, drug regulatory departments across the country have handled a total of 31509 category II medical device registrations and 4596 category III medical device registrations.
Comparison of domestic AI medical device examination and approval rules
The evaluation requirements and partial description of domestic AI medical device products in 2022 are quite different from the 2019 Medical device Software Review Essentials of Deep Learning decision-making (hereinafter referred to as “key points”).
Here’s a comparison:
| Difference | Key Points of Deep Learning-assisted Decision Software Review for Medical Devices in 2019 | 2022 Ai Medical Device Registration Review Guidelines in 2022 |
| Policy orientation | It is suitable for the registration and application of medical device software (including independent software and software components) for deep learning decision making | Artificial intelligence general guidelines for medical devices to replace the Deep Learning Auxiliary Decision Review Key Point of Medical Equipment Software non-clinical part requirements |
| Software purpose | Assisted decision-making: Assisting medical staff in clinical decision-making by providing recommendations for diagnosis and treatment activities | Assistant decision-making (Assistant):
Assist users (such as medical staff and patients) in medical decision-making by providing suggestions on diagnosis and treatment activities, such as auxiliary triage, auxiliary detection, auxiliary diagnosis and auxiliary treatment through lesion feature identification, lesion nature determination, medication guidance and treatment plan formulation. It is divided into real-time and non real-time, and the risk of real-time auxiliary decision-making is higher than the latter. |
| The non-auxiliary decision: using deep learning technology for pretreatment, such as image quality improvement, imaging speed improvement and image reconstruction), process optimization, conventional post-processing (image segmentation, data measurement) | Non auxiliar decision-making（工具）：
Only provide medical reference information without making medical decisions Process optimization Diagnosis and treatment drive The former, such as simplified imaging process, simplified diagnosis and treatment process, and the latter, such as improved imaging quality, improved imaging speed, automatic measurement, automatic segmentation, three-dimensional reconstruction, etc It is divided into real-time and non real-time, and the risk of real-time non auxiliary decision-making is higher than the latter. |
|
| Core algorithm | Deep learning | Artificial intelligence (deep learning, integrated learning, transfer learning, reinforcement learning, federated learning, generation of confrontation network, continuous learning / adaptive learning) |
| Algorithm transparency | Black box | Black box, white box, gray box (combination of white box and black box) |
| Data sources | Medical device data (medical images and medical data generated by medical data) | Objective data generated from medical devices and used for medical purposes. Such as medical image data generated by medical imaging equipment, physiological parameter data generated by medical electronic equipment and in vitro diagnostic data generated by in vitro diagnostic equipment |
| Under special circumstances, the objective data generated by general-purpose equipment (non-regulatory objects) for medical purposes also belong to medical device data, such as skin photos taken by digital cameras for skin disease diagnosis, ECG data collected by health electronic products for early warning of heart disease, and so on. | ||
| The generation and use of medical device data, including the separate use of medical device data, or the combined use of non-medical device data based on medical device data (such as patient complaint information, conclusions of inspection reports, electronic medical records, medical literature, etc.). | ||
| Data collection | Ensure the diversity of data on the basis of compliance, in order to improve the generalization ability of the algorithm, such as from as many representative clinical institutions as possible, different regions and different levels of clinical institutions, as much as possible from a variety of collection equipment with different acquisition parameters. | Data collection should consider the compliance, sufficiency and diversity of data sources, the scientificity and rationality of data distribution, and the sufficiency, effectiveness and accuracy of data quality control. |
| Algorithm design | Algorithm design should consider the quality control requirements of algorithm selection, algorithm training, network security protection, algorithm performance evaluation and other activities. It is recommended to combine data-driven and knowledge-driven algorithm design to improve algorithm interpretability | It mainly considers the requirements of algorithm selection, algorithm training, and algorithm performance evaluation. For black-box algorithms, the algorithm design should carry out analysis of factors affecting algorithm performance, and it is recommended to establish association with existing medical knowledge to improve the interpretability of the algorithm. |
| Algorithm research data | Including data source compliance statement, analysis data of algorithm performance influencing factors, and comparison analysis data of algorithm performance evaluation results under various test scenarios. | The algorithm research report is applicable to the initial release and re release of artificial intelligence algorithm or algorithm combination, including algorithm basic information, algorithm risk management, algorithm requirements specification, data quality control, algorithm training, algorithm verification and confirmation, algorithm traceability analysis, conclusion and other contents. The reasons for not applicable contents are detailed. |
Key point analysis：
I. Auxiliary decision-making / non auxiliary decision-making positioning and use segmentation
The Key Points in 2019 gives a brief explanation on the definition of auxiliary decision-making / non auxiliary decision-making of AI medical device software, and the Guiding Principles for Registration and Review of Artificial Intelligence Medical Devices in 2022 (hereinafter referred to as the principles) gives a more detailed supplementary explanation.
Auxiliary decision-making products are generally divided into four categories: auxiliary triage, auxiliary detection, auxiliary diagnosis and auxiliary treatment. Non auxiliary decision-making defines its positioning and only provides medical reference information, such as process optimization and diagnosis and treatment drive, without medical decision-making. It is helpful for the applicant to clarify the specific positioning and purpose of its AI medical device products.
II. Refine three types of data sources
For the data source of artificial intelligence medical devices, the 2019 Key Points describes “medical images and medical data generated by medical devices”. The 2022 principles specify three types of data sources. What these data sources have in common is objective data for medical purposes. It is worth mentioning that in special cases, patients use mobile phones to take their own symptom photos for auxiliary diagnosis of diseases. Such images also belong to medical data.
III. Association between black box algorithm design and medical knowledge
In the Key Points in 2019, the algorithm design suggests the combination of data-driven and knowledge driven. Specific areas of relevant knowledge are not mentioned in the Key Points. The 2022 Principles mentiones that for the black box algorithm, the algorithm design should carries out the analysis of the factors affecting the algorithm performance. At the same time, it is suggested to establish an association with the existing medical knowledge to improve the interpretability of the algorithm.
Approval comparison of class II / class III certificates of AI medical devices
In August 2018, the new version of Medical Device Classification Catalogue officially came into force, setting up approval channels for AI medical device products according to class II and class III medical devices for the first time.
In combination with the use descriptions of class II and class III AI medical device products in the Key Points in 2019 and the Principles in 2022, AI medical device products with low maturity and used for real-time decision-making are more suitable for applying for class III Registration Certificate.
| Product Performance |
Applicable Policy |
||
|
Is it used for medical purposes |
Maturity |
Function |
|
|
Yes |
Both high and low | Non auxiliary decision-making (providing clinical reference information for data processing and measurement) | Class II Certificate |
| Yes | low | Assist in decision-making (provide clinical diagnosis suggestions such as lesion feature identification, lesion nature determination, medication guidance, treatment plan formulation, etc.). | Class III Certificate |
| Yes |
High |
Assist in decision-making | Medical device classification catalogue > > and classification definition documents |
| No | / | / | Scope of non-medical devices |
Both class II and class III certificates need technical review, but the approval time is different. The technical review of class II Certificate is completed within 60 working days, while the technical review of class III certificate is completed within 90 working days.
Review process of AI medical device registration certificate (Category II and III):
Source: Physician Weekly
Edit and Translate：BradyKnows
BradyKnows provides turn-key solutions to help overseas manufacturers facing the regulatory challenges in China. Pls contact us for any questions via info@bradyknowsmedical.com
