AI in Healthcare in China: The 2026 State of Play

China has a healthcare resource problem, and it has had one for decades. Too many patients, not enough senior doctors, and the best of them concentrated in a few dozen tertiary hospitals in Beijing, Shanghai and Guangzhou. Artificial intelligence was never going to solve that on its own. But it did stop being a pilot project. In 2026 AI sits inside radiology reading rooms, inside hospital triage desks, inside the WeChat account patients use to book an appointment, and increasingly inside the drug discovery pipeline of Chinese pharma.

This article was first published in 2018, back when the domestic AI medical device market was worth about 13 billion yuan and everybody was quoting the same forecast. The forecast stopped being the interesting part a long time ago. What matters now is what has been approved, what has been deployed, and what the rules allow you to do with patient data.

I am Olivier Verot, founder of Gentlemen Marketing Agency, based in Shanghai since 2012. My team works with medical device makers, diagnostics companies and health brands entering China, so we spend our days on the unglamorous side of this topic: registration timelines, Baidu visibility in a regulated category, and what a hospital procurement officer actually reads before shortlisting a supplier. Find me on LinkedIn.

Where China’s medical AI market actually stands in 2026

China’s medical AI solutions market was worth around 16.4 billion yuan in 2024, up from 9.1 billion yuan in 2021, and was estimated at roughly 18.2 billion yuan for 2025 by Chinese industry analysts. Frost & Sullivan puts the longer curve at a 43.1% compound annual growth rate through 2033. Growth is real, but the base is smaller than the headlines suggest. Anyone quoting tens of billions of dollars is adding up hardware, hospital IT and consumer apps, or talking about the artificial intelligence industry in China as a whole.

The more useful number is regulatory. As of mid-2026 the National Medical Products Administration has issued 127 registration certificates for AI medical software, and 126 of them are Class III, the highest risk category. That is both the barrier to entry and the proof of adoption: a Class III certificate means a hospital can legally buy the product.

Read the last line twice if you are a foreign vendor. Seven imported certificates out of 127. That is not a rounding error, it is the shape of the market.

Xinhua reported in July 2026 that approvals have moved well past lung nodules: cervical cytology screening software, gastric lesion detection in endoscopy, orthopaedic surgical robots, chromosome karyotyping, and nuclear medicine imaging all cleared Class III in the first half of the year. United Imaging Intelligence has accumulated 20 Class III certificates, Shukun Technology 19. Lung imaging (32 certificates) and cardiovascular (22) are saturated. Pathology, endoscopy and ultrasound are still thin.

Why Chinese hospitals went looking for AI in the first place

The original argument still holds. Pairing AI with medical practice removes a large share of highly repetitive work and frees doctors to produce faster, more consistent reports. Machine learning reached healthcare and life sciences faster than most other regulated sectors for that reason alone.

The doctor supply has improved, the distribution has not. In 2018 this article quoted 2.22 licensed physicians per 1,000 people in China against an OECD average of 3.19. By 2024 China had reached 3.61 licensed and assistant physicians per 1,000 people, according to the National Health Commission’s annual statistical bulletin published in December 2025. On paper the gap closed. In practice patients still travel across provinces to Huashan or Xiehe because they do not trust the county hospital, and the county hospital does not have a radiologist who reads 200 chest CTs a day.

That mismatch is exactly the gap AI is pointed at. China carries one of the world’s heaviest imaging workloads, and scan volume grows faster than radiologist headcount. Software that flags a suspicious nodule and pushes it up the reading queue does not replace anyone. It changes the order in which the work gets done. That is a far easier sell to a hospital director than “AI diagnosis”. Cost is the second argument, and it bites harder at the bottom of the pyramid: a Shanghai tertiary hospital buys AI for throughput, a township health centre buys it because it has no specialist at all. Policy is now aimed squarely at the second group.

The five uses that are real in 2026

1. Diagnostic support and medical imaging

This is the mature segment and the one holding the certificates. Chinese vendors take roughly 85% of the domestic imaging AI market. Shukun Technology built its position on cardiovascular CT and sells by subscription rather than licence, which matters because hospital capex approval in China is slow and opex approval is not. United Imaging Intelligence bundles AI with its own scanners, a distribution advantage no software-only vendor can match. If your modality is lung or heart, you are entering a fight you will probably lose. Ultrasound, pathology, dermatology, endoscopy and ophthalmology sub-specialties still have room.

2. Medical large language models, including DeepSeek inside hospitals

This is the biggest change since 2018, and it happened fast. After DeepSeek released its reasoning models, Chinese hospitals started deploying them on their own servers rather than calling an API. Local deployment is the whole point: patient data never leaves the hospital network, which resolves the compliance objection that blocked cloud AI for years.

Hundreds of hospitals across more than twenty provinces have deployed DeepSeek-based systems, including Peking University First Hospital, Shanghai Sixth People’s Hospital and Tsinghua Changgung. The use cases are mundane, and that is why they stuck: drafting discharge summaries, retrieving drug interaction information, summarising a patient history before a consultation, quality control on medical records. Doctors save minutes per patient. Nobody claims the model diagnoses anything.

The consumer side moved just as fast. CNNIC reported in March 2026 that generative AI products reached 602 million users in China as of December 2025, up 141.7% in a year, a 42.8% penetration rate. Online healthcare users stood at 411 million. The two audiences overlap heavily: people now ask a chatbot about their symptoms before they ask a search engine.

The commercial consequence is direct. If DeepSeek, Doubao or Yuanbao cannot find a Chinese-language source describing your product, your certification and your clinical evidence, you do not exist in the answer. That is the mechanism behind generative engine optimisation in China. These models retrieve from indexed Chinese content: Baidu Baike, Zhihu threads, industry media. No Chinese footprint, no citation.

3. Triage, appointment booking and patient service

Tencent launched WeChat Intelligent Healthcare (微信智慧医疗) back in 2014. The platform let users book appointments, pay bills and handle guahao registration at hospitals through WeChat official accounts. By 2017 more than 38,000 medical facilities in China had a WeChat account, and most of them already offered some form of online consultation.

Eight years on, that layer is the default interface between Chinese patients and the health system. Alibaba built the same thing on Alipay through Ali Health, whose DoctorYou diagnostic system launched early but never reached the same adoption. iFlytek took another route and put voice AI into consultation rooms and rural general practice, which turned out to be the more defensible position.

What is new is the AI triage layer in front of all of it. A patient describes symptoms in plain language, the system routes them to the right department, pre-fills the intake form and hands the doctor a structured summary. Hospitals buy it because it shortens the queue at the registration desk, not because it is clever.

4. Drug discovery

China’s AI pharma sector stopped being a pitch deck category. XtalPi has moved several candidates into clinical stages and signed joint laboratory deals worth hundreds of millions of yuan with domestic pharma. Insilico Medicine’s platform serves a majority of the world’s twenty largest drug makers. In 2026 CSPC and Insilico signed partnerships with AstraZeneca and Servier respectively, which tells you multinationals now treat Chinese AI discovery platforms as suppliers rather than curiosities. Industry estimates put the cut in early discovery timelines at 30% to 40%. Treat those as vendor-side numbers: the honest proof will be approval rates in five years, not cycle times today.

5. Health wearables

The consumer end matters more in China than in Europe because it feeds the same data pipeline. IDC counted 73.9 million wrist-worn devices shipped in China in 2025, up 20.8% year on year, split between 50.6 million smartwatches and 23.3 million bands, with about 79.6 million units forecast for 2026. Growth has slowed to single digits, which is why the differentiation fight moved entirely to health sensing: blood oxygen, ECG, sleep apnoea screening, blood pressure trends, and the prize nobody has won yet, non-invasive glucose monitoring. Huawei leads and has been the most aggressive at pushing health features through medical device registration instead of shipping them as wellness gadgets. For a foreign brand, that is the bar now.

The rules you have to work inside

NMPA and AI software as a medical device

Any AI product that produces a diagnostic or treatment output in China is regulated as a medical device, almost always Class III, which means clinical evaluation. For AI imaging software the NMPA has consistently required multi-centre clinical trials run in China on Chinese patients, even when the product already holds FDA clearance or a CE mark. Algorithm performance depends on the training population, so the requirement is medically defensible. The practical effect is a two to four year timeline and a bill most mid-size foreign vendors will not sign off. A draft NMPA guideline on clinical evaluation of AI-assisted diagnostic devices, published in June 2026, reduces the ambiguity but not the requirement.

Health data, PIPL and localisation

Medical and health information counts as sensitive personal information under China’s Personal Information Protection Law. That triggers separate consent, a stated necessity for processing, and stricter security duties. National standard GB/T 45574-2025 spells out what is covered: disease conditions, medical history, consultation records, test data, plus biometric and genetic identifiers.

In February 2026 five agencies, including the National Health Commission, the Cyberspace Administration and the Ministry of Public Security, jointly issued trial measures on data security and personal information protection for medical institutions. They set a three-tier data classification (core, important, general), assign liability to whoever administers, operates or uses the data, and impose conditions on providing data externally or moving it across borders. Add the Human Genetic Resources rules if you touch genomic data.

The conclusion for a foreign vendor is simple. Assume training data stays in China, assume inference runs in China, and design for that from day one instead of asking for an exemption later.

The line AI is not allowed to cross

The 2022 Internet Diagnosis and Treatment Supervision Rules from the National Health Commission state that a prescription must be issued by the attending physician in person, and explicitly prohibit using artificial intelligence to generate prescriptions automatically. Several municipal health commissions, Beijing included, restated the ban in their own rules. AI in China assists. It does not prescribe and it does not sign a diagnosis. Any marketing claim implying otherwise gets your account pulled before it gets you a customer.

Policy points at the bottom of the pyramid

On 1 December 2025 several ministries led by the National Health Commission issued an implementation opinion on promoting and regulating “AI + healthcare” applications. The stated target is that by 2030, AI-assisted diagnosis reaches essentially full coverage in primary care facilities, with secondary and higher hospitals widely running AI imaging and clinical decision support. It follows the November 2024 reference guide from the same commission, which mapped 84 AI application scenarios across medical service management, public health, health industry development and medical research.

The procurement signal is clear. Money is going to county hospitals, township clinics and chronic disease management, not to another lung nodule algorithm for a Shanghai tertiary hospital. If your product only works in a top-tier hospital, you are aiming at the segment with the least budget growth.

Market access for a foreign AI health company

Three realistic routes exist, and only one of them is “sell your software to Chinese hospitals”.

• Direct registration and direct sales. Full NMPA Class III process, a local legal agent, Chinese clinical trials, then a hospital sales team. Highest cost, longest timeline, full control. Works if you have a genuinely differentiated modality and patient capital.
• Licence the algorithm to a Chinese holder. A domestic partner registers the product, holds the certificate, sells it, and pays you a royalty. You lose margin and some control. You gain three years and several million euros. This is what most mid-size European vendors end up doing.
• Sell around the hospital. Private clinics, health check-up chains, insurers, corporate health programmes and pharma services buy AI tools with less procurement friction than public hospitals. Lower ticket, faster cash.

Whichever route you take, the marketing problem is identical: Chinese partners, distributors and hospital procurement staff have to find you and have to believe you. They will not read your English website. They will search your company name in Chinese, and they will ask an AI assistant.

How to be visible in a regulated category

Baidu is still where the diligence happens

The old claim that Baidu handles 75% of Chinese search no longer holds, and anyone still quoting it is reading a 2018 deck. Baidu lost ground to in-app search, to Douyin, and now to AI assistants. It has not lost its role in B2B diligence. When a hospital equipment manager or a provincial distributor checks whether your company is real, Baidu is the first stop, and an empty first page reads as a red flag in a category where trust is the product. You need a mainland-hosted Chinese site, indexed content on your registration status and clinical evidence, and a clean brand SERP. Paid search works, but medical categories carry heavy ad restrictions on Baidu and account review is strict: expect to justify every claim with a document. Our sector detail sits in the guide to medical, pharmacy and healthcare marketing in China.

WeChat is the private domain, not the acquisition channel

With over 1.3 billion monthly active accounts, WeChat is not optional. For a B2B health company it is not where leads come from. It is where they are kept. A service account plus a mini-program hosts your Chinese documentation, certification files and distributor onboarding. Sales staff add prospects on WeCom and the relationship lives there for years. Tencent also pushed its own health products deep into the app, from bookings to WeSure insurance, which trained a generation of users to handle medical matters inside WeChat without thinking about it.

Zhihu and the professional forums now feed the AI answers

Chinese professionals bring questions to the internet and answer each other in real detail. Zhihu, Baidu Zhidao and specialist medical communities like DXY are where a radiologist actually asks whether a given AI tool is worth the workflow disruption. Those threads used to matter for reputation alone. Now they matter twice, because they are retrieval material for the assistants your buyers query. A well-argued answer from a named expert can be quoted back to a prospect by a chatbot six months later. Weibo still works for announcements and media pickup, and Xiaohongshu matters if you sell anything consumer facing, wearables included.

Case study: Pieter, a Dutch AI screening vendor

Pieter runs a Dutch company selling AI retinal screening software for diabetic retinopathy. CE marked, deployed in several European screening programmes, solid clinical publications. He had been trying to enter China for 18 months when we met. The score at that point: about 140,000 euros spent, two CMEF trade fair stands, a translated English website, a LinkedIn campaign, three inbound enquiries in a year, zero contracts. His conclusion was that Chinese hospitals did not want foreign technology.

That was not the problem. Two other things were. Diabetic retinopathy is one of the most crowded AI modalities in China, and several domestic vendors already held Class III certificates with Chinese training data and hospital relationships behind them. And he had no NMPA registration at all, so no public hospital could legally buy the product however much a department head liked it. Every trade fair lead was dead on arrival.

We changed the offer, not the marketing volume. He stopped selling software to hospitals and started looking for a Chinese partner to hold the registration and distribute under licence. That reframed the whole content brief. Instead of product brochures we built Chinese material aimed at partners: the clinical evidence package, the European screening results, and a frank explanation of how the licensing structure would work. Mainland-hosted Chinese site, technical arguments seeded on Zhihu and DXY, narrow Baidu SEM on partner-intent keywords rather than product keywords, WeChat service account for follow-up.

It worked because the audience was right. Ophthalmology distributors and mid-size Chinese medtech firms were actively hunting for foreign algorithms to license, and nobody was talking to them in their own language about it. Over nine months: 47 qualified partner enquiries, 11 serious conversations, two signed evaluation agreements, and one licensing deal with a Chinese medtech company that opened its NMPA file in month eleven. Pieter never sold a licence to a hospital directly, which was the point.

FAQ

Can a foreign company sell AI diagnostic software directly to Chinese hospitals?

Only with an NMPA registration certificate, and for diagnostic AI that means Class III. Of the 127 AI medical software certificates issued to date, seven went to imported products. Public hospitals cannot legally purchase an unregistered diagnostic device, so a department head’s enthusiasm counts for nothing without the certificate. Most foreign vendors either license the algorithm to a Chinese registration holder or sell into private clinics, check-up chains and insurers, where the procurement route is shorter.

How long does NMPA approval take for AI medical software?

Plan for two to four years from decision to certificate. The long pole is the multi-centre clinical trial the NMPA requires in Chinese patients, even for products already cleared by the FDA. You need to line up several hospital imaging departments, run the study, and compile the file through a local legal agent. A draft NMPA guideline on clinical evaluation of AI-assisted diagnostic devices published in June 2026 clarifies the expectations, but it does not remove the local trial requirement.

Does our patient data have to stay in China?

Assume yes and design accordingly. Health and medical information is sensitive personal information under PIPL, which means separate explicit consent and stricter security duties. The trial measures issued in February 2026 by five agencies add a three-tier classification and specific conditions on any external provision or cross-border transfer. Hospitals overwhelmingly prefer on-premise deployment for exactly this reason, which is why locally deployed large models took off so quickly.

Should we build content for DeepSeek and other AI assistants?

If your buyers are Chinese, yes. With 602 million generative AI users in China as of December 2025, a growing share of first-stage research happens in a chatbot, not a search box. These models retrieve from indexed Chinese sources: Baidu Baike, Zhihu, industry media, official registration databases. If your certification status, clinical evidence and product category are not documented in Chinese anywhere those crawlers reach, the model will answer with your competitors’ information instead of yours.

Working with us

Gentlemen Marketing Agency has been running China digital marketing out of Shanghai since 2012, with a dedicated practice for medical devices, diagnostics, pharma services and health brands.

We handle the parts that decide whether a regulated product gets traction: a compliant Chinese site that ranks on Baidu, content built for AI assistants like DeepSeek and Doubao, partner and distributor lead generation, and WeChat private domain management for long B2B cycles.

If you are weighing direct registration against licensing, or you simply need Chinese buyers to find you, contact us and we will tell you honestly what the segment looks like before you spend anything.