Airdoc and Silibase Secure China's First Class III Certifications for AI-Powered Diabetic Retinopathy Diagnostic Software

VCBeat has learned that on August 10, the National Medical Products Administration (NMPA) issued an announcement on its official website stating that the “Fundus Image-Assisted Diagnostic Software for Diabetic Retinopathy” developed by Shenzhen SiBright Co., Ltd. and the “Fundus Image-Assisted Diagnostic Software for Diabetic Retinopathy” developed by Shanghai Airdoc Medical Technology Co., Ltd. (Airdoc) had received NMPA approval and obtained Class III medical device certificates. This marks another milestone victory in the commercialization of medical AI.

As early as the beginning of 2018, the National Institutes for Food and Drug Control (NIFDC) assumed responsibility for the quality evaluation and research of medical artificial intelligence products. At that time, leveraging the multi-party database construction experience of enterprises and hospitals, the initial standard databases were limited to two categories: color fundus images and pulmonary CT scans.

At that time, using the Guiding Principles for Technical Review of Medical Device Software Registration, the Guiding Principles for Technical Review of Mobile Medical Device Registration, and the Guiding Principles for Technical Review of Medical Device Cybersecurity Registration as the benchmark for database construction, multiple parties jointly established a standard fundus imaging database containing 6,327 cases and a standard lung imaging database containing 623 cases. Data indicates that the diabetic retinopathy database is the earliest established and largest AI evaluation database in China.

However, the first AI software to gain approval was not for diabetic retinopathy. Experts have noted that fields like diabetic retinopathy, characterized by numerous R&D enterprises and a broad patient population, require more refined standards to avoid patching loopholes after clinical application. Thus, seven months after the first AI software was approved, the regulatory outcome for diabetic retinopathy AI was finally settled. This time, the National Medical Products Administration approved two companies at once.

Airdoc (Shanghai Yingtong) and SiBright were fortunate to secure tickets to AI commercialization, but behind this “fortune” lies the years of accumulation and deep expertise of both companies.

For several years, the sensitivity and accuracy of AI products marketed by artificial intelligence companies have been widely criticized due to the lack of unified standard databases and evaluation criteria.

As collaborations with medical institutions continued to expand, algorithms underwent continuous iteration, and a deeper understanding of the approval process was achieved, Airdoc and SiBright ultimately outperformed other AI companies in the diabetic retinopathy market, becoming the first movers.

Therefore, the path to developing AI for medical imaging is not as smooth as it may appear. The companies that endure until the end and obtain regulatory approval have navigated countless detours along the way.

Certainly, this process also relies on the efforts of regulatory authorities. On July 17, 2019, the AI Medical Device Innovation Cooperation Platform was established. This platform aims to build an open, collaborative, and shared innovation ecosystem for AI medical devices, bringing together the Center for Medical Device Evaluation (CMDE), the China Academy of Information and Communications Technology (CAICT), several leading universities, and multiple top-tier medical institutions to jointly advance the approval of artificial intelligence technologies. In this review and approval process, CAICT was responsible for conducting cybersecurity testing of AI medical devices, while CMDE led the review of key clinical evaluation criteria for AI medical devices, taking into account product indications and usage scenarios.

Furthermore, in 2019, two AI-based ophthalmic products successfully passed the review and approval process for innovative medical devices, with Airdoc and SiBright both making the list. The accelerated approval was made possible by the significant efforts of institutions such as the Center for Medical Device Evaluation.

Hu Zhigang, General Manager of SiBright, previously highlighted key considerations for AI approval in his speeches. He stated, “The National Medical Products Administration (NMPA) primarily focuses on four aspects: First, it is necessary to clearly demonstrate that the core algorithm patents are utilized in the company’s own products and to elaborate on how AI is specifically applied within these products. Second, regarding product finalization, the sources of standard data must be clearly specified, with explanations provided from the perspectives of quantification and traceability. Third, concerning the significant value of clinical application, real-world clinical data must be provided to prove that the product can effectively address clinical issues, and such data should be made available to regulatory authorities.”

Overall, the key points for reviewing deep learning-assisted decision-making medical device software are similar to the conventional approach for medical device product development, which can be divided into stages such as clinical needs analysis, data collection, algorithm design, and verification and validation, among which data is the most critical element.

Data collection must prioritize the compliance of data sources and ensure data diversity, taking into account factors such as disease composition, geography, population distribution, equipment, institutional hierarchy, and epidemiology. Furthermore, data quality assessment, the degree of data de-identification, and data transfer methods must all adhere to established regulatory standards.

Data preprocessing comprises two components: data cleaning and data processing. The preprocessing approach must consider its impact on the product and associated risks, clearly defining the state of the data before and after preprocessing. Furthermore, the software tools used for preprocessing must be validated to demonstrate the reliability of the data processing.

Finally, after rigorous annotation as described above, an annotated dataset is formed, which needs to be split into training, validation, and test sets. Factors to consider in dataset splitting include clinicians’ experience and the requirements of algorithm experts. Dataset splitting must also account for data sample distribution and clinical needs. The training set should have a balanced sample distribution to ensure adequate learning across various disease types. For the validation and test sets, consideration should be given to the real-world disease distribution in clinical scenarios and potential confounding factors from other diseases.

From a commercialization perspective, SiBright follows the conventional approach to AI implementation, deploying its products through hospitals and primary healthcare institutions. In contrast, Airdoc has pursued more diversified commercialization strategies.

Furthermore, an increasing number of AI companies are beginning to partner with pharmaceutical firms to assist patient communities in managing chronic diseases. Diabetic retinopathy corresponds to the tens of millions of patients with diabetes; if this segment is effectively addressed, the commercialization pathway for AI companies specializing in diabetic retinopathy will differ entirely from that of applications in radiology.

In retrospect, the medical AI sector has become a fiercely competitive “red ocean.” Compounded by the impact of the 2020 pandemic, many companies have reached a critical make-or-break juncture, with only a few leading AI firms securing financing. At this point, is it already too late to obtain regulatory approval?

Regarding this question, Xie Guotong, Chief Medical Scientist at Ping An Technology, provided a negative answer: “AI will inevitably become a standard feature in hospitals in the future; however, the trajectory of this adoption may differ from what we currently observe, with the ecosystem being the key factor.”

Therefore, regulatory approval is no longer the most critical bottleneck constraining AI development. Instead, identifying sustainable business models that hospitals are willing to pay for and enhancing the clinical value of existing products will remain core challenges for AI enterprises for a considerable period.

Perhaps this year, many products such as those for pulmonary nodules and CTA will successively obtain Class III medical device certifications; however, AI still has a long way to go in terms of commercial implementation.