Landmark Policy Unveiled for Medical AI Registration Review: Five Key Points Explained

Slow Progress in the Registration and Approval of Medical AI: The Absence of Relevant Regulations Is the Key CauseAs a non-traditional software-based medical device, artificial intelligence (AI) employs algorithms, utilizes data, and targets functions that fall outside the scope of traditional regulatory guidelines. Consequently, reviewers and approvers are forced to proceed with caution, navigating uncharted territory in advancing the registration and approval process for medical AI.

On March 7, the Center for Medical Device Evaluation of the National Medical Products Administration (NMPA) released a 41-page “Guiding Principles for the Registration and Review of Artificial Intelligence Medical Devices” (hereinafter referred to as the “Guiding Principles”), marking the end of the aforementioned situation. The Guiding Principles provide a detailed description of the concepts, basic registration principles, lifecycle processes, and technical considerations for artificial intelligence medical devices, offering clear regulations and standards for registering enterprises.

To understand the implications of the Guidelines and gain insight into the future trajectory of medical AI, VCBeat interviewed relevant experts to comprehensively review the policy content, aiming to help medical AI practitioners clarify the NMPA’s approval rationale and key considerations. This article will interpret the full text of the policy from three perspectives: the incremental information contained in the Guidelines, the strategic adjustments for medical AI enterprises, and the overall development of the medical AI industry.

From Standing at the Forefront to Steady Progress: Over the Past Seven Years, the Center for Medical Device Evaluation Has Been Remarkably Reticent in Issuing Documents Related to Medical AI. The Previous Comprehensive Guideline, “Key Points for the Review of Medical Device Software Assisted by Deep Learning for Decision-Making,” Was Released Two Years Ago. Under the New Standards, Artificial Intelligence Medical Devices Now Have a General Guidance Standard, and the Current Approval and Review Requirements Differ Significantly from Those in the Past.

Comparison of Key Incremental Information between the “Guiding Principles for Registration Review of AI Medical Devices” and the “Review Points for Deep Learning-Assisted Decision-Making Medical Device Software”

In terms of key information comparison, the two versions of the policy have undergone substantive changes in positioning, software usage, core algorithms, and algorithm transparency, while wording adjustments were made regarding data collection, data sources, and algorithm design.

Since Google’s top-secret lab, Google X, successfully trained a neural network to recognize images of cats by feeding it data in 2012, convolutional neural networks (CNNs) and deep learning—the algorithms behind this breakthrough—have become synonymous with artificial intelligence. The vast majority of the first wave of medical AI companies chose to leverage deep learning to design AI-assisted diagnostic systems for healthcare.

Therefore, the previous round of AI policy, titled “Key Points for the Review of Medical Device Software with Deep Learning-Assisted Decision-Making,” was essentially a set of norms and standards for mainstream algorithms. Judging from the approval outcomes, the first 12 medical AI devices approved by the Center for Medical Device Evaluation were all explicitly labeled with the term “deep learning” in their approval documents.

However, with the continuous advancement of artificial intelligence algorithms, companies such as Deepwise Medical, United Imaging Intelligence, and Tencent Healthcare have been actively exploring novel AI approaches, including few-shot learning and unsupervised learning. Although these emerging research outcomes also fall within the scope of artificial intelligence, they are not currently covered by regulatory review and approval processes. In particular, supervised learning-based AI has been prominently highlighted in official documents for its positive significance.

Content and Regulation of Various Artificial Intelligence Algorithms

Ding Jia, Co-founder and CTO of MedAI Intelligence, stated, “The approval process for AI-based medical devices has been one of continuous exploration. When companies propose innovative algorithms beyond deep learning, they often need to engage in multiple rounds of communication with the National Medical Products Administration (NMPA) before finalizing product functionalities and intended uses for registration purposes.”

From this perspective, one of the core significances of the release of the "Guiding Principles" is the redefinition of artificial intelligence. Under the new policy, various innovative AI algorithms are expected to accelerate through the review and approval process.

In the "Key Points for Review of Medical Device Software with Deep Learning-Assisted Decision-Making," the first sentence of the section "Key Focus Areas for Review" states that "deep learning is essentially a black-box algorithm based on massive datasets and high computational power." Consequently, subsequent review priorities focus on areas such as data quality control, algorithm generalizability, and clinical use risks.

In the Guidelines, a distinction is made between “white-box” and “black-box” algorithms used in artificial intelligence. It requires that white-box algorithms, whether employing supervised or unsupervised learning, adhere to the model/data quality control standards applicable to supervised deep learning. Specifically, they must clearly specify feature information, such as feature categories (e.g., demographic, biological, morphological), feature attributes (e.g., shape, texture, properties, size, boundary), and feature presentation methods (e.g., shape, size, boundary, color, quantity).

In this section, VCBeat raised the question of whether “black-box algorithms will gradually be phased out from the requirements for AI-assisted clinical decision support software.” In response, Ding Jia, Co-founder and CTO of Yizhun Smart Medical Technology, stated, “Although regulatory policies encourage companies to enhance algorithm interpretability through methods such as manual feature extraction, black-box algorithms currently demonstrate significantly higher accuracy than white-box algorithms in the industry. While their internal mechanisms remain unexplainable, the feasibility of black-box algorithms can be validated using performance metrics such as stability and robustness.”

Typically, we have been accustomed to viewing imaging AI as auxiliary diagnostic software. In past policy definitions regarding software intended use, a distinction was made only between “decision-support” and “non-decision-support” categories, with specific terminology appearing solely in the context of Class III medical device certifications. Now, the Guidelines provide a precise differentiation of the term “decision support.”

Under the new standards, AI-based medical devices are categorized into four types: auxiliary triage, auxiliary monitoring, auxiliary diagnosis, and auxiliary treatment. A review of regulatory approvals shows that products such as those for diabetic retinopathy by Airdoc, Siyuan Intelligence, and Zhiyuan Huitu, as well as MRI-based intracranial tumor detection by Ande Medical Intelligence, are classified under auxiliary diagnosis; CT-based pulmonary nodule detection by Deepwise and Infervision, and X-ray-based fracture detection by Huiyi Huiying are classified under auxiliary detection; while CT-based pneumonia screening by Infervision, United Imaging Intelligence, and Tencent Healthcare is categorized as auxiliary triage. This demonstrates that the classification of software is not related to the maturity level of AI, but rather to the specific clinical problems addressed by the AI.

The “Key Points for the Review of Deep Learning–Assisted Decision-Making Medical Device Software” does not provide a detailed analysis of the data sources for artificial intelligence medical devices, merely stating that they are “medical images and medical data generated by the medical device.”

In the Guidelines, medical device data is explicitly defined as objective data generated by medical devices for medical purposes, such as medical image data produced by medical imaging equipment (e.g., X-ray, CT, MRI, ultrasound, endoscopy, and optical images), physiological parameter data generated by medical electronic devices (e.g., waveform data including electrocardiogram [ECG], electroencephalogram [EEG], blood pressure, non-invasive blood glucose, and heart sounds), and in vitro diagnostic (IVD) data generated by IVD equipment (e.g., pathological images, microscopic images, and invasive blood glucose waveform data).

Notably, under specific circumstances, objective data generated by general-purpose equipment (which is not a regulated entity) for medical purposes also qualifies as medical device data. Examples include dermatological images captured by digital cameras for the diagnosis of skin diseases, and electrocardiogram (ECG) data collected by consumer health electronics for early warning of cardiac conditions. This implies that skin photographs taken by patients using their mobile phones, when utilized via skin-assisted decision-making mini-programs such as those developed by Voxel Tech and DXY, can likewise be regarded as medical data. This requirement provides greater room for advancement for AI enterprises specializing in dermatology.

Furthermore, the text mentions that “medical device data is based on aspects such as the generation and usage of such data, wherein ‘usage’ includes the standalone use of medical device data, or the combined use of medical device data alongside non-medical device data (such as patient chief complaints, laboratory and diagnostic test report conclusions, electronic medical records, and medical literature).” In other words, in future regulatory approvals of AI-based medical devices, text-based data is expected to serve as an important supplement, or even function as a standalone AI tool.

In July 2021, Senyi Intelligence’s VTE system obtained a Class II medical device registration certificate from the National Medical Products Administration (NMPA). It is foreseeable that a large number of companies will subsequently follow suit in applying for medical device registration approval for their VTE systems.

Two critical components—data collection and algorithm design—determine the actual performance of medical artificial intelligence. Both the “Review Points for Medical Device Software Using Deep Learning to Assist Decision-Making” and the “Guiding Principles” provide detailed descriptions of these two aspects. Indeed, in past regulatory reviews and approvals, many companies failed to meet regulatory requirements for algorithm accuracy, sensitivity, and robustness, primarily due to deficiencies in these two areas.

In the “Key Points for Review of Medical Device Software with Deep Learning-Assisted Decision-Making,” the Center for Medical Device Evaluation (CMDE) did not explicitly require data to originate from distinct sources, but rather encouraged manufacturers to ensure data diversity to the greatest extent possible. In contrast, in the “Guiding Principles,” the CMDE requires manufacturers to “consider the compliance, sufficiency, and diversity of data sources; the scientific rigor and rationality of data distribution; and the sufficiency, effectiveness, and accuracy of data quality control,” aiming to promote the robustness of AI-based medical devices through the approval process.

Algorithm design has undergone a similar shift. The “Review Points for Medical Device Software with Deep Learning-Assisted Decision-Making” only recommends that enterprises adopt a combination of data-driven and knowledge-driven approaches in algorithm design to enhance algorithm interpretability, whereas the “Guiding Principles” require that black-box algorithms undergo an analysis of factors influencing algorithm performance to improve interpretability.

For enterprises, official guidance serves as a lighthouse; under its illuminating direction, businesses no longer grope in the dark, and operations begin to proceed in an orderly manner.

“In the short term, the standardization of processes has introduced new requirements. Companies may conduct more multi-center trials and perform multi-dimensional dataset testing prior to application, which could increase R&D and registration costs for medical AI enterprises,” said Ding Jia. “However, taking a long-term view, a standardized system will inevitably offer sustained advantages by ensuring the quality of medical AI, driving the development of related industry chains such as CROs for AI-based medical devices and dataset validation, and making registration costs for R&D companies more stable.”

It is foreseeable that in the coming year, we will see a more diverse and richer array of AI-powered medical devices take root in hospitals. As the industry matures and stabilizes, emerging fields such as proton and heavy ion therapy and brain science are poised to unlock new waves of AI-driven opportunities.

In addition to the guidelines, the National Medical Products Administration (NMPA) successively released two major documents on March 7 and March 9: the “Announcement of the Center for Medical Device Evaluation of the NMPA on Issuing the Guidelines for Registration Review of Cybersecurity in Medical Devices (2022 Revised Edition)” (No. 7 of 2022) and the “Announcement of the Center for Medical Device Evaluation of the NMPA on Issuing the Guidelines for Registration Review of Medical Device Software (2022 Revised Edition)” (No. 9 of 2022). These initiatives aim to establish a robust and comprehensive regulatory review and approval system for digital healthcare.

Nowadays, emerging concepts such as digital therapeutics and clinical decision support are developing rapidly. Not only artificial intelligence, but also various innovations arising from the wave of digital healthcare require corresponding regulatory approvals to standardize them, thereby safeguarding the healthy development of the industry.

For artificial intelligence, the emergence of new policies may not once again thrust it into the spotlight. However, as a consensus on AI gradually takes shape and the integration of digital healthcare becomes normalized, medical AI companies that have lingered for years may finally accelerate their entry into an era of profitability.