A model that predicts binding affinity, reads a pathology slide or triages a waiting list is scientifically ready long before it is legally ready. The Chinese hospital that offers a cohort, the biotech that offers samples and the contract research organisation that offers a clean data pipeline are each offering something that a different regulator controls. The company then discovers that the data cannot lawfully leave, that the software needs a registration it has not filed, and that the collaboration agreement it signed gives the hospital joint ownership of the model.
The three questions are usually described as separate workstreams: device regulation, data protection and intellectual property. In China they are sequenced, and the sequence is set by the human genetic resources regime. An approval under that regime is tied to a named Chinese party, a stated research purpose and a work plan, and it cannot be obtained retrospectively for work already done. A model trained on Chinese genomic data without the approval is not merely exposed to a fine; the provenance problem travels with the weights into every later filing, publication and financing.
The conclusion this guide works towards is that the collaboration should be designed around the gates rather than around the science. Establish who holds the resource, obtain the approval for the purpose you actually intend, classify the product before anybody deploys it, choose the data route before the study begins, and paper the ownership before the first sample is sequenced. Each gate is cheap at the design stage and expensive, or impossible, afterwards.
The gate order and why it is fixed
Chinese life science regulation is a set of ministerial rules, technical guidelines and local practice rather than a single code, and five regulators each answer a different question. The order in which they are satisfied is not arbitrary. Each gate produces a document that the next gate reads.
| Gate | Authority | Question it answers | Can it be fixed after the event? |
|---|---|---|---|
| Human genetic resources | HGRAC, within the Ministry of Science and Technology | Is the resource held by an eligible Chinese party, and is this collaboration approved for this purpose? | No. An approval covers prospective work under a stated plan |
| Ethics and institutional approval | Hospital ethics committees, health commissions | May this study proceed at this site, on this consent basis? | Partly. Re-consent is possible and expensive |
| Data export | CAC, with sector regulators | Is there a lawful route to move the data, or information derived from it, offshore? | Sometimes. Restructuring to train in China is the usual repair |
| Device registration | NMPA and provincial medical products administrations | Is the product approved for the intended use it claims? | Yes, slowly |
| Foreign investment access | MOFCOM and NDRC | May a foreign investor hold equity in the provider? | Yes, by moving from equity to contract |
Two interlocks make the order binding rather than merely sensible. The human genetic resources approval states a research purpose, and the device registration states an intended use. If the two describe different things, the later filing exposes the earlier one. A study approved to investigate biomarkers of one condition does not authorise commercial development of a diagnostic for another, and an NMPA file that claims the second use invites the question of what authorised the data behind it.
The second interlock is between the data route and the clinical evaluation. NMPA accepts clinical evaluation based on existing literature or real-world data in some cases, which is attractive because it avoids a trial. The real-world data has to come from somewhere, and if it comes from a Chinese cohort it sits behind the human genetic resources gate and the personal information rules. A clinical evaluation strategy built on hospital data that cannot lawfully be used is a strategy that collapses at the review meeting.
Human genetic resources: what the gate catches
China treats human genetic resources as a national resource. The 2019 regulations and the Biosecurity Law prohibit foreign parties from collecting and preserving Chinese human genetic resources, and foreign-invested entities in China are generally treated as foreign for this purpose. International collaboration requires a Chinese partner and prior approval from the Human Genetic Resources Administration of China within the Ministry of Science and Technology. Exporting samples requires approval. Providing human genetic resources information to a foreign party generally requires a filing, and the information may be used only for the approved purpose.
The regime distinguishes materials from information, and AI companies usually trip on the second. Materials are the samples: blood, tissue, cells, cell lines and the like. Information is what is derived from them, including genomic sequence data, and the category is broad enough to cover the data types that modern discovery runs on. Whole genome and exome sequencing, transcriptomic data, single-cell data and many biomarker panels fall within it. Imaging on its own is generally not a human genetic resource, but imaging linked to a genomic dataset is caught through the link, which is precisely the structure of a multimodal model.
The harder question is what happens to a model trained on that data. Whether weights are themselves human genetic resources information is not settled, and the regulator’s concern is the transfer of information derived from the resource rather than the form it takes. The safe assumption for a foreign company is that a model trained on Chinese human genetic resources data is derived information, and that shipping the weights offshore, or training abroad on data that was moved offshore, is a transfer that the regime was designed to catch. Companies that plan to train in Singapore or California on a Chinese cohort should assume the plan requires the approval structure described below, not a contractual workaround.
The approval itself is project-specific. It names the Chinese and foreign parties, states the research purpose, describes the work plan and runs for a defined period. Changes to the parties, the purpose or the plan require an amendment rather than an internal memo. Results that contain human genetic resources information are subject to rules on publication and on what may be provided to the foreign party, and the foreign party is restricted from using the information for any purpose outside the approval. A foreign parent that treats the approved dataset as a general-purpose asset for the group has stepped outside the approval even if every other document is in order.
The operating structure that works has three features. The Chinese party holds the resource: samples, raw data and the approvals are in its name, and it participates substantively rather than lending its name to a foreign study. The foreign party receives defined outputs: aggregated results, model parameters trained under the approved purpose, or analysis outputs specified in the collaboration agreement, rather than the raw resource. The work plan matches the commercial plan: if the company intends to develop a diagnostic, the purpose statement should describe that, because a purpose statement drafted narrowly for an academic paper will not stretch later.
Where the collaboration involves a contract research organisation or a data vendor, check its ownership before relying on it. A Chinese-incorporated company with foreign shareholders may be treated as foreign for this purpose, which means the entity the company thought was its Chinese partner is not an eligible one. The check is a corporate search and a question to the partner, not an assumption based on where the office is.
Clinical data: consent, ethics and the export decision
Patient data is sensitive personal information under the Personal Information Protection Law. That characterisation carries obligations that run through the whole collaboration: a specific and legitimate purpose, separate consent where the law requires it, a necessity assessment that limits collection to what the purpose needs, and an impact assessment for processing that poses a high risk. Hospitals are themselves personal information handlers with their own obligations, so the agreement has to allocate roles rather than assume the company is the only party processing data.
Consent is where the commercial plan most often outruns the paperwork. A consent form drafted for clinical care covers treatment, not the training of a commercial model, and not the transfer of derived data to a foreign parent. Secondary use of existing records for research generally requires ethics approval and, depending on the site and the study design, fresh consent. A consent that says “for research” without naming the commercial purpose, the categories of recipient and the cross-border element is a weak foundation for a product. The repair is a re-consent exercise, which costs months and, in a retrospective cohort, may be impossible for part of the population.
Ethics approval is institutional, so a multi-centre study needs approval at each site, usually with a lead site coordinating. The ethics committee is not a formality: it controls access to the data in practice, and it will ask what happens to the data after the study, who can see it and whether the model is a device. A protocol that describes a research-use tool while the commercial plan is a diagnostic product is the kind of mismatch that stalls a study halfway through.
The export decision is best made before the study design is fixed, because it constrains the design. There are three statutory routes for cross-border transfer: a security assessment by CAC, a standard contract filed with CAC, and certification. Which route applies depends on the volume and sensitivity of the personal information, whether the data is assessed as important data under sector rules, and whether the transfer falls within an exemption. The thresholds that separate the routes were relaxed in 2024 and remain subject to revision, so the analysis should be run on current figures rather than on a number remembered from an earlier deal.
Health data complicates the analysis because it may be assessed as important data under healthcare sector rules, which changes the route rather than merely the paperwork. De-identification helps and does not settle the question. Chinese law sets a high bar for anonymisation: the data must be incapable of identifying an individual and incapable of being re-identified, and the assessment is made on the dataset as a whole, including what could be achieved by linking it with other data. Pseudonymised data, where a key exists somewhere, remains personal information.
Where a transfer is not available, the practical structures are well known. Train inside China with a licensed partner and export the model rather than the data, subject to the human genetic resources analysis above. Run federated learning, where the data stays at the hospital and only model updates move, provided the agreement defines who owns the updates and the resulting model. Use a data enclave operated by the hospital or a licensed third party, with the company’s engineers working inside it under supervision. Synthetic data is sometimes proposed as a way around the problem, and it deserves a warning: data generated from real patient records may still be personal information, and if the source was human genetic resources, the derived dataset inherits the provenance question.
The NMPA file as a product gate
Software becomes a medical device when it is intended for diagnosis, monitoring, treatment or the relief of disease, or when it assists clinical decisions. A triage tool, a lesion detector or a radiotherapy planner is a device; a scheduling system, a billing tool or a pure data pipeline is not. The intended use statement in the labelling decides the analysis, which is why a sentence written for an investor deck can create a regulatory obligation. The statement should be drafted as a legal document, reviewed by the regulatory adviser, and kept consistent across the labelling, the instructions for use, the website and the clinical evaluation.
Classification follows risk. Class III covers higher-risk uses, including software that supports diagnosis or treatment of serious disease or whose output a clinician is expected to rely on; Class II covers lower-risk functions. Class III registration is filed with NMPA nationally, Class II with the provincial medical products administration. An overseas manufacturer registers through a domestic agent appointed in China, and the agent carries obligations that continue after registration, so the appointment should be made before the file is drafted rather than after.
The file is heavier than software teams expect, and the heaviest part is the evidence that the algorithm behaves as claimed.
| Component | What the reviewer is testing | Where files usually fail |
|---|---|---|
| Product technical requirements | Whether the claimed performance is measurable | Performance described in marketing terms with no metric |
| Software research report | Whether the development lifecycle and version control are controlled | No traceability from requirement to test |
| Algorithm validation | Whether the model performs on the intended population | Validation set drawn from the same site as the training set |
| Clinical evaluation | Whether the clinical benefit is established | Literature argument asserted rather than constructed |
| Cybersecurity | Whether the device resists attack and protects data | Penetration testing done on a different build |
| Quality management system | Whether the manufacturer can make the device consistently | System documented for a different product line |
Clinical evaluation can sometimes rest on existing literature or real-world data rather than a new trial, but the argument has to be built and documented. For an AI diagnostic, the reviewer will want to know the intended population, the reference standard, the performance in subgroups and the behaviour at the operating point the product actually uses. A model that reports a high average accuracy and degrades on the subgroup the hospital cares about will be questioned at exactly that point.
Algorithm updates are governed by change control. NMPA approves a defined version. Major changes, such as a shift in intended use, a change of algorithm type or retraining that alters performance, require a change registration; minor changes can be handled inside the quality management system. Continuously learning models are difficult to approve because the approved product must be predictable, so the common solution is a locked model with a planned retraining cycle and documented change control rather than an algorithm that updates itself in the field.
The device gate and the data gate meet in the clinical evaluation. Real-world data used to support the file must be lawfully held, ethically approved and, where it is human genetic resources information, within an approval. A registration file that cites a hospital dataset which the hospital cannot lawfully provide to the company is a file that will not survive scrutiny, and the problem surfaces late, when the review is already underway.
Patents, exclusivity and the asset layer
A compound, a diagnostic method or a device can be patented. An AI system cannot be named as inventor, because Chinese practice requires a natural person. The human contribution to posing the problem, selecting the candidates and verifying the results determines inventorship and, in practice, whether the claim survives examination. A team that ran a generative model over a library and tested the top hundred candidates has made a contribution, and the patent should describe it in those terms rather than presenting the model as the author of the invention.
For AI-generated molecules the hard issues are sufficiency and inventiveness. The specification must disclose how the compound is made and support the claimed utility, and a candidate that emerges from a generative model still needs a credible technical advance over prior art. Computer-implemented claims need a technical effect, not an algorithm described in patent language. Where the invention involves a biological material that is not available to the public and cannot be described so that the invention can be carried out, Chinese practice provides for deposit with a recognised depositary institution, and the deposit becomes part of the disclosure strategy rather than an administrative detail.
Filing strategy should account for two features specific to pharmaceuticals. China provides patent term extension for new drug patents, which compensates part of the time lost to regulatory review, and operates a patent linkage system that connects drug approval to the patent status of the reference product. Both change the value of a filing made early in the development cycle, and both require the patent to be in the right form and the right status at the right time.
Approved innovative drugs can obtain a period of protection for undisclosed trial data, commonly described as up to six years and subject to conditions. That protection matters to a company whose value sits in clinical evidence rather than in a compound claim, and the scope has been adjusted through implementing rules, so it should be confirmed with counsel at the point of reliance rather than assumed from an earlier summary.
The asset layer that surrounds the patent is usually where an AI biotech’s value actually sits. Model weights, training corpora, labelling schemas and hyperparameters are better protected as trade secrets than as patents, because the Anti-Unfair Competition Law protects technical information that is secret, commercially valuable and subject to reasonable secrecy measures. Access control, logging, segmentation of the training pipeline and tight employment terms do real work here, and they are the evidence a court will look for. Software copyright registration is inexpensive evidence of authorship for the platform code and is worth filing. A target patent owned by a third party is the other side of the same coin: an AI company developing a molecule against a known target may be infringing before it has filed anything of its own, so a freedom-to-operate search on the target and the modality belongs at the start of the programme, not at the licensing stage.
Structuring the hospital collaboration
Hospitals are powerful counterparties. They control data, ethics approval and clinical access, and their research agreements often claim joint ownership of results and broad rights to use data. The agreement is negotiated with the hospital’s research management office as well as the principal investigator, and a deal struck with the investigator alone may not bind the institution.
Five clauses carry most of the value. Background IP should be listed rather than described, so that the model the company brought, the data the hospital holds and the assays each party contributes are identifiable. Foreground IP should be allocated by category: patents on compositions and methods, copyright in software, rights in the trained model and rights in the data derivatives are different assets and need not follow the same rule. Joint ownership should be understood before it is accepted, because the default position lets each party exploit the patent and grant non-exclusive licences, which means the hospital can license the joint result elsewhere. Commercialisation rights should state who may sell what, in which field, and on what royalty. Sample custody and data access after termination should say what happens when the collaboration ends, including whether the model may continue to be trained on the hospital’s data and whether the hospital retains a copy of the weights.
Publication review is the clause foreign companies most often regret. The hospital and the investigators have an interest in publishing, the foreign parent has disclosure obligations to investors and obligations to customers, and the patent strategy requires filing before publication. A review period long enough to file, with a defined process for clearing the text, protects all three.
The equity alternative deserves a short warning. Medical institutions are restricted for foreign investors, generally to joint venture form, and internet-based diagnosis is separately regulated. Most AI companies are better served by a wholly foreign-owned entity for software research and development and licensing, with collaboration contracts rather than equity in the provider. Taking equity in a hospital-adjacent business also brings the transaction within the foreign investment regime, which adds a review that the contract route avoids.
Documents, filings and timeline
| Step | Authority | Typical timing | What foreign companies miss |
|---|---|---|---|
| Intended use and classification review | Internal, with a regulatory adviser | Two to four weeks | The statement is drafted for investors, and the claim makes the product Class III |
| Human genetic resources collaboration approval | HGRAC, within MOST | Months, and the clock starts only when the Chinese partner’s file is complete | The foreign parent is named as a party, or the work plan describes a broader purpose than the approval |
| Ethics approval at each site | Hospital ethics committees | Weeks to months per site | The consent form does not cover commercial model training or transfer |
| Data export assessment, standard contract filing or certification | CAC | Weeks to months | Volume thresholds assumed rather than calculated, and important data not assessed |
| Class II device registration | Provincial medical products administration | Months | Local review practice differs from the national guidance |
| Class III device registration | NMPA | Well over a year, including clinical evaluation | The algorithm validation file is incomplete, so the review restarts |
| Domestic agent appointment | NMPA | Before filing | Appointed after the file is drafted, so the agent cannot correct the labelling |
| Patent filing for the compound, the method or the device | CNIPA | Two to three years to grant | Filed after the conference abstract that disclosed the structure |
The sequence in the table is the sequence in which the documents depend on each other. Classification comes first because it determines what the clinical evaluation has to prove. The human genetic resources approval comes next because it defines who may hold the data and for what purpose. Ethics and export follow the study design. Registration comes last because it consumes everything before it.
Where foreign companies get this wrong
The first failure is running the pilot before classifying the product. A hospital puts the model into a workflow, a clinician begins to rely on the output, and the deployment becomes an unapproved use of a device. The registration file then has to explain a clinical use that was never authorised, and the company has to decide whether to stop the pilot, which damages the relationship, or continue it, which compounds the exposure.
The second is assuming that de-identified data is exportable. The hospital removes names and identifiers, the company ships the file, and the data remains personal information because a linkage key exists at the site and the genomic component is human genetic resources information. The export is then unauthorised twice over, and the model trained on it carries the problem.
The third is naming the foreign parent as the collaborating party. The approval is given to a Chinese party. A foreign entity cannot hold the resource, and a Chinese company with foreign shareholders may not be an eligible partner either. The error is usually made in the term sheet, where the foreign parent signs as the party of interest, and it is expensive to unwind once the study has begun.
The fourth is accepting the hospital’s template intellectual property clause. Joint ownership, a broad licence back, no field-of-use limit and no obligation on the hospital to prosecute produce a result in which the company has funded the work and does not control the asset. The clause is in the template because it is the hospital’s default, and it is negotiable when the negotiation starts early.
The fifth is delivering the weights to the hospital as a project deliverable. A well-meaning engineering team ships the trained model so that the clinicians can use it, and the hospital’s research office treats the weights as foreground IP. The company has then given away the asset it needs for the next product, and the gift is documented in an email.
Cost and commercial impact
The costs in this sector are dominated by time and by the structure of the study rather than by official fees. The human genetic resources process costs what it costs to design a compliant collaboration and to run a Chinese partner’s work programme; the fee itself is not the item that matters. Ethics and data costs scale with the number of sites and the amount of re-consent. Device registration cost is driven by clinical evaluation and by the quality management system, and a Class III file that has to restart because the validation data was inadequate costs a year.
The commercial impact of the gate order is asymmetric. A collaboration structured correctly at the outset costs some months and some legal fees. A collaboration structured incorrectly costs the study, the dataset, the model trained on it and, in an extreme case, the ability to use the result in a regulatory file. Companies that treat the gates as a compliance exercise at the end of the project are buying the most expensive version of the same work.
Data exclusivity and patent term extension change the arithmetic on the other side. A product that reaches approval with a compound patent, a period of data protection and a device registration has a defensible commercial position; a product that reaches approval with a trade secret in the weights and no registered right has a position that depends on nobody reverse-engineering it.
Practical checklist
- Write the intended use statement and test whether it makes the product a medical device, then keep every document consistent with it.
- Classify the software and confirm the Class II or Class III route before building the validation file.
- Map every dataset to its source, its consent basis and its human genetic resources status.
- Confirm that the Chinese partner is an eligible entity and not a foreign-invested company before naming it in an approval.
- Draft the research purpose to match the commercial plan, because the approval will not stretch later.
- Choose the data route, whether local training, federated learning, an enclave or a filed transfer, before the study begins.
- Put background IP, foreground IP, joint ownership, commercialisation rights and post-termination data access into the hospital agreement.
- Adopt a written trade secret programme covering weights, corpora and pipeline access, and keep the records that prove it.
- Check foreign investment restrictions before taking equity in a provider, and prefer the contract route where the restriction applies.
- Confirm current NMPA guidance, human genetic resources practice and data thresholds with counsel before filing.
The firm advises AI biotech and medtech companies on China entry, hospital collaborations, human genetic resources approvals, device registration strategy and the intellectual property side of those deals. The relevant service page is intellectual property, and companies setting up a Chinese research or licensing entity should also read the note on incorporation. Two related guides go further: intellectual property strategy for AI companies covers the portfolio and filing routes that sit behind the asset layer described here, and the note on cross-border data transfer sets out the export routes in more detail.
Frequently Asked Questions↑ Back to article
Is our AI diagnostic software a medical device in China?
It usually is if the output is used for diagnosis, monitoring or treatment decisions, or to assist a clinician's judgement. Software that merely stores, transmits or displays data without clinical interpretation generally falls outside device regulation, so the intended use statement in your labelling decides the classification.
Do we need NMPA approval before running a hospital pilot?
If the pilot uses an unapproved device to influence patient care, yes, or it must run as an approved clinical trial. A research-use-only deployment that does not affect treatment may avoid device registration, but hospital ethics review and data approvals still apply, and the boundary is easy to cross in practice.
Can we train a model on Chinese patient data outside China?
Generally not without approvals. Chinese human genetic resource rules restrict foreign parties from collecting or holding Chinese samples and data, and personal information rules restrict cross-border transfer. Training abroad usually requires a Chinese partner, ethics approval, a human genetic resource approval or filing, and a lawful data export route.
Who owns the intellectual property from a project with a Chinese hospital?
It depends on the research agreement, and hospitals increasingly ask for joint ownership or a broad licence. Joint ownership under Chinese law lets each party exploit the patent and grant non-exclusive licences, so define background IP, foreground IP and commercialisation rights before the study starts.
Can a foreign company hold Chinese human genetic resource data?
No. Foreign parties are prohibited from collecting and preserving Chinese human genetic resources, and foreign-invested entities in China are generally treated as foreign for this purpose. Access is structured through a Chinese partner, with approvals or filings, and the foreign party receives agreed data rather than control of the resource.
How long does NMPA registration take for AI medical software?
Class III software commonly takes well over a year once clinical evaluation and quality system review are included, and Class II is faster but still measured in months. Timelines vary with the completeness of the algorithm validation file, so confirm the current review position with counsel and your regulatory adviser.
Are AI-generated molecules patentable in China?
A compound can be patented, but an AI system cannot be named as inventor, because Chinese practice requires a natural person. The human contribution to selecting and verifying the candidate determines inventorship, and you also need sufficient disclosure of how the compound is made and why it works.