When AI Puts on the White Coat: Who Will Wake Up Dormant Medical Data?

China has the world's largest patient population, the richest pool of clinical cases, and the most active large-model ecosystem. By all logic, AI combined with healthcare should be blooming everywhere.

Yet the reality is disheartening.

Hefei secured a national pilot base for AI application (primary-care direction) and set clear annual targets for medical-data aggregation in 2026, but months into the effort the results have fallen far short of expectations. It is not a lack of compute, nor weak algorithms, and certainly not a shortage of use cases. The only thing stuck is one thing: data. Medical data simply cannot be obtained.

This is no isolated case. Over the past two years, almost every team that has charged into the AI-healthcare race has said the same thing: the technology is ready, but the data is still asleep.

Every year China produces hundreds of millions of outpatient records, inpatient charts, imaging reports, and lab slips. By sheer "reserves" of data, we far surpass most countries. But when it comes to actually training models or improving diagnosis, this data is locked inside countless "drawers" that cannot talk to one another. Each hospital is a drawer; each department is another drawer; and data often does not even flow between different systems within the same department.

Why Is the Data Locked in the Drawers?

The most commonly cited explanation is privacy and security. Medical data involves people's most core personal information, and strict oversight is justified. But that is not the whole story.

A university researcher studying the AI industry put it bluntly. Their university wanted to merge the data of two departments, pushed for two years, and got nowhere. "Why? Information is power. Once you hand the data over, you lose the power."

The remark sounds sharp, yet it hits the root. Under the current institutional setup, data is first and foremost an organizational resource. Whoever holds the most complete data in a given field holds the power of discourse, the power of budget allocation, and irreplaceability. Handing over data is like cutting off one's own arm. This is not a problem of any individual, but of the entire incentive structure: those who give up data get no equivalent benefit, while those who withhold it pay no price. The system is naturally inclined to lock things down.

Economists call this "market failure": data has enormous use value yet cannot flow, because property rights are unclear, pricing is absent, and the rules of exchange are blank.

Zhejiang's Solution: Data Stays In-House, but Must Interoperate

A few years ago, Zhejiang began laying what later came to be called the "health-data high-speed rail." Its origin was not artificial intelligence but the "run at most once" reform. Ordinary people should not have to repeat medical tests; to achieve that, data between hospitals had to be connected.

The "Health Cloud" was built, and hundreds of public hospitals across the province gradually connected to the same mutual-recognition and interoperation system. Zhejiang became the only province in the country to achieve real-time interconnectivity of medical-institution data across its entire jurisdiction.

What is even more worth unpacking is how this road was built.

At the start there was debate. Some advocated pulling all the data up into a central repository for unified management. But it quickly became clear this would not work. On what basis should hospitals "surrender" their data? Who would bear responsibility if a security incident occurred? Every institution wanted to guard its own "drawer."

The eventually adopted plan made one key design choice: data would remain in each hospital, but the whole province unified the access protocols and mutual-recognition standards. The "Health Brain" handles scheduling—fetching whatever is needed, leaving traces after each fetch, fully traceable throughout. Hospitals retained control while being obligated to interoperate.

This is essentially building a club. Members share each other's data, set the rules together, and every operation leaves a trace. There is an admission threshold for outsiders and behavioral constraints inside. This design greatly reduces institutional friction. If institutions were forced to "surrender" data, round after round of bargaining could drag the whole project to death; the club model lets participants keep control while enjoying the dividends of collaboration, cutting resistance by a wide margin.

What Grows on Top of the Foundation

In 2024, Zhejiang launched the "Anzhen'er" AI triage system, with a digital assistant accompanying patients throughout the entire journey—from registration and waiting to examination, dispensing, and follow-up.

Technologically, there is no irreplaceable breakthrough. Its real value lies in the foundation. It is not the "in-hospital navigation" of any single hospital, but a unified entry point built atop the province-wide data network. Tests a patient had done in Hangzhou are recognized by "Anzhen'er" in Wenzhou as well. Once the platform effect kicks in, the more hospitals join, the more valuable it is to patients; the more patients use it, the more willing hospitals are to join. Once demand-side economies of scale form, latecomers face a much higher cost of catching up.

It goes beyond triage. AI-assisted diagnosis now covers all of Zhejiang's primary-care institutions. Community hospitals lack top-tier resources, yet AI can help doctors read scans and run initial screenings. Multi-cancer early screening is being piloted, and intelligent critical-care decision systems are running. In some grassroots areas, misdiagnosis rates have seen real, tangible declines. This metric is more convincing than how much money was raised or how many papers were published.

Taken apart, none of these applications is a "disruptive breakthrough." But run on the same data foundation, they constitute a kind of systemic efficiency: fewer trips to the doctor, no repeated tests, and primary care able to handle common illnesses. This is not the victory of any single product, but the return on an entire set of institutional designs.

From Local Experiment to National Agenda

In its "Action Plan for the Development of Trusted Data Spaces," the National Data Administration explicitly called for building "trusted data spaces" in key sectors such as healthcare, with the core principle of "raw data stays in-domain, data is usable but invisible." Zhejiang's "Health Brain" is essentially a prototype of an industry-grade trusted data space—using technology to guarantee data does not leak, and institutions to let data flow. On the technology side, companies like Ant are also rolling out "trusted agent swarms" that let multiple AIs verify and constrain one another, preventing large models from "talking nonsense" in serious medical scenarios. With trusted data spaces on the institutional side and trusted agent swarms on the technology side, only when the two come together can AI healthcare move from the lab to scale.

Why the "Cars" Won't Take to the Road

But we cannot tell only the pretty half of the story. The Zhejiang Provincial Health Commission candidly admits that, despite being the only province to achieve data interoperation, there remain many bottlenecks in data ownership, pricing mechanisms, and authorized operation of public data. In plain language: the road is built, but there are not yet enough "cars" willing to keep running on it.

The problem lies in incentives. A hospital puts its data out to be shared; the model gets trained, and commercialization lands—does any of that have anything to do with the hospital? If not, why should it keep cooperating next time? An administrative order can push things through once, but not sustainably.

The direction Zhejiang is exploring is to grant medical institutions that share data priority in using models and sharing R&D results, following the logic of "contribute data, get AI capabilities in return." The direction is right, but the details are hard.