In Depth: Chinese AI Developers Struggle to Win Over Health Care

The use of artificial intelligence (AI) in the Chinese healthcare sector is growing as various companies recognize its potential; however, challenges such as skepticism from hospitals and regulators about AI products' accuracy and cost-effectiveness remain prevalent[para. 1]. For example, diagnostic and gene technology company Sansure Biotech Inc. has integrated its intelligent sequencing system with an AI model from DeepSeek, while Shenzhen Mindray and Tencent Holdings have introduced a large language model (LLM) for processing clinical data[para. 2]. Despite DeepSeek's successful integration of AI into medical products, Chinese firms have generally been integrating AI and big data in healthcare over the past few years, with companies like Tencent, Baidu, and Alibaba being at the forefront[para. 3].

Most medical AI startups focus on diagnostics, particularly disease screening, medical image generation, and clinical decision support, as they can efficiently use large datasets and operate under asset-light models, according to Li Sheng of Roland Berger consultancy[para. 4]. Since the release of OpenAI's ChatGPT in late 2022, LLMs have been swiftly developed for various industries, including healthcare, where they enhance physician productivity, patient experience, and treatment periods, which is crucial considering the demand on healthcare due to an aging population[para. 5]. Regulatory challenges and prioritization of patient safety and data protection have inhibited the medical AI sector[para. 6]. Financing issues compound these obstacles; post-pandemic shifts from COVID-related products have led to a decrease in funding for medical AI, which dropped from 11.5 billion yuan in 2021 to 2.96 billion yuan in 2023[para. 7]. Commercialization remains challenging, as limited insurance coverage and safety concerns prevent widespread adoption by hospitals[para. 8][para. 9].

Regulatory hurdles are highlighted by concerns of AI tools generating inaccurate information, endangering patient safety[para. 10]. This concern is supported by research from Mass General Brigham revealing that AI-generated responses, such as those from OpenAI's GPT-4, could pose risks to patients, including death in some cases[para. 11]. LLMs struggle to gain trust among medical professionals due to difficulties in understanding and predicting errors, and in some regions, regulators limit AI's role to assessments rather than clinical decision-making[para. 12][para. 13]. Chinese regulations mandate that AI-assisted treatment can only occur with qualified personnel present[para. 14]. Data security efforts also result in limited data availability for AI development, a problem exacerbated by inadequate disease-specific databases in Chinese medical institutions[para. 16][para. 17].

Success in the consumer market for medical AI products depends largely on hospital endorsement[para. 18]. However, financial constraints and stricter control over medical expenses limit hospitals' willingness to invest in and endorse such technologies[para. 20][para. 21]. Evaluating the worth of medical AI products hinges on their ability to address genuine clinical needs without adding burdens, something which is not consistently achievable as seen in real-world scenarios, such as patient pre-diagnoses backed by AI[para. 23][para. 26].

Industry leaders advocate for more substantial policy support for medical AI, particularly concerning data accessibility and fee policies[para. 27]. Currently, regulatory bodies prioritize sectors like manufacturing for AI application over health care due to stronger economic benefits[para. 28]. Though the Ministry of Science and Technology and other bodies support AI growth, healthcare regulators focus on associated risks and insurance sustainability, impeding progress[para. 30]. Unlike the U.S, which includes some AI products in its Medicare program, China has yet to incorporate such products into its national insurance, emphasizing AI's supplementary role in clinical settings rather than as standalone diagnostic tools[para. 31][para. 32].