NMPA published Technical Guideline on AI-Aided Devices on June 28, 2019. Grace Fu Palma, CEO of China Med Device, and Zhiyong Rao, Testing Director of China Med Device, were interviewed by Elise Mak from BioWorld MedTech on the Guideline. Please see BioWorld’s article below.
For a series of NMPA regulatory establishments on AI products, please click HERE.
For English version of Guideline on AI-Aided Devices, please email info@ChinaMedDevice.com. We charge nominal fees for the translation.
Chinese Government Provides AI-aided Software Guidelines for Health Care Market
By Elise Mak, Staff Writer
HONG KONG – To respond to the emerging trend of adopting artificial intelligence (AI) in health care, China now tells its regulators what to look for when evaluating and approving AI-aided software, with a focus on the data and algorithms used.
In late June, China’s medical device regulators released a guideline for reviewing software using machine learning and data sourced from medical devices that can influence doctors’ decisions in diagnosis and treatment. AI applications can range from medical image recognition to diagnostic products for chronic disease management.
“AI-aided software as a medical device is completely new to the China industry. It was after April 2018 when the U.S. FDA approved the first AI-based device to detect certain diabetic retinopathies that AI software, especially diagnostics based on image recognition, was drawn to the attention of the Chinese public,” Grace Fu, CEO of China Med Device LLC, a Boston-based consulting firm focused on China’s med-tech regulations, told BioWorld MedTech.
“Due to its novelty, developers around the world do not know how to ensure the safety and effectiveness of the software. It would also be inappropriate for Chinese regulators to review such software based on existing guidelines. This guideline aims to guide manufacturers on how to develop and register such software,” she added.
During the review, the regulators will look at how the software is intended to be used, in what scenarios it is used and what core functions it has to ensure its safety and effectiveness.
They will place emphasis on how the software controls data quality and enables generalizations that allow an algorithm to be effective across a range of inputs and applications, as well as what clinical risks may arise. The risks arising from the software, according to the regulators, can lead to late or unnecessary diagnosis and treatment.
Companies should submit clinical data that supports their software. Regarding data, clinical institutions must meet the requirements for collecting, pre-processing and labeling data, and building datasets. For example, data masking and cleansing are required.
In addition, data used by the software for trial runs must be sourced from legitimate third-party databases to ensure the data is accurate, standardized and diverse.
For imported software, developers must provide the ethnic and epidemiological variations in the data. If they cannot ensure the software’s safety and effectiveness if used in China, they must conduct clinical trials in the country to prove its reliability within China.
As for algorithm design, regulators will look at algorithm selection, training and performance assessment as well as cybersecurity to ensure its interpretability.
The guideline also clarified that any algorithm-driven software updates would be considered major ones, which will require the regulators to re-assess the algorithm performance and clinical evaluation of the software to ensure its safety and effectiveness.
Zhiyong Rao, testing director of China Med Device, said overseas manufacturers may face a greater workload and higher costs to meet these requirements.
“AI-aided software’s safety and effectiveness lie in the quality and quantity of training, validation and test datasets,” Rao told BioWorld MedTech.
He noted that when developing products, manufacturers need to consider expanding the three datasets to include enough data from Chinese people, especially Chinese living in mainland China.
Manufacturers also need to enhance the generalization ability of algorithms by diversifying datasets, strictly controlling the collection, pre-processing, and labeling process of the datasets, while also conducting risk analysis, risk control and documentation for the entire process.
The guideline comes as China is quickening the pace of adopting AI in its medical sector to make health care more accessible and efficient for its huge population. Following the many other AI-specific policies, med-tech analysts believe this guideline is one more step in this direction.
“The Chinese government has supported the use of AI in health care through its Internet Plus health care policy,” Karen Simpkins, Medical Devices Analyst at Fitch Solutions, told BioWorld MedTech.
China’s Internet Plus health care policy allows medical institutions to provide online diagnostic services for some common and chronic diseases. Last year, it issued rules on e-Hospital, e-Diagnosis and telemedicine, as well as a more detailed plan to manage the standard, safety and service of health care big data. (See BioWorld MedTech, Oct. 24, 2018, and Oct. 29, 2018.)
The AI health care market in China has been estimated at around $3 billion. The vast market and clearer policies have encouraged several hundred Chinese companies, including internet giants and numerous startups, to develop AI-aided products for medical use.
“Both Alibaba and Tencent are active in developing artificial intelligence to assist doctors in making diagnoses – Alibaba through its Doctor You medical imaging software for the early detection of lung cancer and diabetes, and Tencent through its Mihying AI medical imaging application which also targets early cancer detection,” Simpkins explained.
“Ping An Health Technology, an affiliate of Ping An Insurance is also partnering with Varian to use artificial intelligence, cloud computing and big data technologies to expand access to cancer care in China,” she added.