CHINA Law and Practice Contributed by: Gil Zhang, Diana Li, Muran Sun and Yongqi Tao, Fangda Partners
6. Data Governance in Healthcare AI 6.1 Training Data Requirements For healthcare AI systems that qualify as medical devices, the Guiding Principles for AMD Registration Review provide key compliance requirements for the training data. • Data quality: It is mandated that the data training process consider compliance and quality control requirements during the key processes of data collection, collation and annotation, as well as the construction of a data set, particularly in relation to (i) data collection devices and personnel man - agement; (ii) data desensitisation; and (ii) process management, the establishment of data collec - tion-, cleansing- and annotation-related operational standards, quality assessment processes, etc. • Fair representation: The training dataset should, in principle, ensure that the sample distribution is bal - anced, scientific and rational (taking into account the epidemiological characteristics of the target disease). Data should be collected as extensively as possible based on the intended use and appli - cation scenarios of the product. This includes data from representative clinical institutions across multiple hospitals, regions and levels, as well as data from representative collection devices from multiple manufacturers, of multiple types and with multiple parameters. • Documentation: The source of training data and quality control processes (including data collec - tion quality assessment results, annotation quality assessment results, etc) shall be traceable, well- documented and structurally managed. For other generic healthcare-related Gen AI services, the Gen AI Measures will regulate the data training process, which primarily requires service providers to use training data and models from legal sources, ensure there is no infringement, take measures to improve data quality and prevent bias, etc. • Data quality: establish data screening and data quality assessment mechanisms, including man - aging illegal and harmful content (less than 5%), abandoning data sources with third-party infringe - ment risks, identifying and removing misleading,
fake or false content in vertical fields like health - care, etc, and assessing data quality through annotation. • Data representation: ensure diverse sources of data of the same format (eg, code, images, audio, video, text in different languages, etc), use both overseas and local training data, etc. • Data documentation: sources of training data shall be traceable and documented, including through the provision of relevant authorisation documents (eg, open-source licence agreements, commer - cial contracts, authorisation records of users, etc) and data collection (from the Internet) records that comply with the limitations of robot protocols and technical restrictions. For bias-mitigation measures, please refer to 5.3 Bias and Fairness . 6.2 Secondary Use of Health Data If healthcare data used for training contains personal information, the PIPL and Measures for the Ethical Review of Life Science and Medical Research Involv - ing Humans requires – as a general principle – that data processing activities, including secondary use of healthcare data for AI training and development, shall be disclosed in the privacy policy/informed consent form to patients, and consent shall be obtained. Although it might not be feasible to obtain consent for secondary use, current laws do not provide con - sent exemptions for secondary use of healthcare data, nor do they specifically provide that secondary use is compatible use. Having said that, the recommended national standard GB/T 39375-2020 Health and Medi - cal Data Security Guidelines provides a mechanism to request secondary use of healthcare data from medical institutions, albeit that this is limited to non- identifiable data and can only be used for non-profit purposes. 6.3 Data Sharing and Access Current legislation governing data sharing and access remains centred around: • the PIPL, if personal information is involved in the training; and
48
CHAMBERS.COM
Powered by FlippingBook