AI software that reads X-ray images, analyses endoscopy video or predicts disease risk is entering Vietnamese hospitals faster than the management framework. The first legal question is not about the algorithm but about characterisation: is software that never touches the patient a medical device requiring market authorisation? This article answers under Decree 98/2021/ND-CP and goes deep into the three issues specific to SaMD (software as a medical device): risk classification, training data and post-registration algorithm change management.
Legal basis
- Decree 98/2021/ND-CP on medical device management: the definition of a medical device includes software meeting a medical purpose determined by its owner; classification principles by risk level in Articles 4 and 5.
- Circular 05/2022/TT-BYT on medical device classification: classification rules for medical software by intended use and the severity of the health condition the software affects; the A, B, C, D framework is retained in Circular 24/2026/TT-BYT (effective 01 July 2026).
- Law on Personal Data Protection 91/2025/QH15: health data is sensitive personal data, the frame governing patient data used in developing and operating medical software.
Is AI diagnostic software a medical device?
Yes, where the software meets the medical device definition in Decree 98/2021: software determined by its owner for a specific medical purpose such as diagnosis, disease prevention, monitoring, treatment or alleviation for humans is a medical device even without accompanying hardware — the group known as SaMD (Software as a Medical Device). The characterisation boundary sits in the purpose the manufacturer declares: software managing appointment schedules or storing purely administrative medical records is not a device; software analysing medical images to suggest a diagnosis, dose calculation software, or complication risk prediction software affecting clinical decisions is a device and must be classified and registered or declared in the correct risk group before being placed in a healthcare facility. Supplying AI software as a cloud service (SaaS) rented by a hospital does not change this characterisation: the supply form is a service but the product nature is a medical device, and the software owner remains the subject responsible for market authorisation in Vietnam (directly or through a representative office or authorised domestic unit).
Which risk class does medical AI fall into?
Medical devices in Vietnam are classified into four classes of increasing risk: class A low risk, class B low-medium risk, class C medium-high risk and class D high risk, under the principles in Articles 4 and 5 of Decree 98/2021 and the classification rules in Circular 05/2022/TT-BYT. For software, the deciding factors are the significance of the information the software provides for a clinical decision and the severity of the health condition: software merely storing or displaying existing measurement data is usually in the low group; software supporting diagnosis or treatment of ordinary conditions is in the middle groups; software providing information that decides diagnosis or treatment of a serious condition, or directly driving a treatment device, falls into class C or D. In practice, AI software assisting image reading for cancer or stroke diagnosis or emergency risk stratification is usually classified in the high risk groups (C or D) because an algorithm error can lead to a missed dangerous disease. The establishment in whose name the authorisation stands must self-classify under the rules and bears responsibility for the classification result (Clause 6, Article 5 of Decree 98/2021); classifying lower than the correct level to take the declaration route instead of registration is a violation sanctioned on post-check. Procedure by group: classes A and B go through standards declaration at the Department of Health; classes C and D require marketing registration at the Ministry of Health with a full technical dossier and clinical evaluation documentation.
Are there specific rules on AI training data?
It should be stated plainly: current medical device law has no dedicated provision governing training data of AI models; this is a management gap that the enterprise must fill with three adjacent frames. The first frame is the marketing registration dossier: the technical documentation and clinical performance evidence of the software must describe the data used to train and validate the model well enough for the appraising authority to assess representativeness for the Vietnamese patient population; a model trained entirely on foreign data risks a request for supplementary performance evidence on domestic data. The second frame is personal data protection: health data used for training is sensitive personal data under the Law on Personal Data Protection 91/2025, so collection from hospitals needs a lawful basis, anonymisation and a clear data processing agreement; using real, non-anonymised medical records to train a commercial model is the largest legal risk of medical AI projects in Vietnam. The third frame is post-registration change management: a continuously learning AI model, or a manufacturer algorithm update that changes the registered performance or intended use, requires the owner to carry out the change registration procedure as prescribed instead of silently updating through the cloud; this is a fundamental difference between SaMD and traditional devices that the hospital supply contract must state clearly on version control rights.
How long does market authorisation for medical AI take?
Timing depends first on the classification group: class A and B software takes the standards declaration route and is fast, depending mainly on the enterprise dossier preparation time; class C and D software must pass marketing registration appraisal at the Ministry of Health, where appraisal of the technical documentation and clinical evidence of an AI product usually generates supplementation requests because the authority must also assess the algorithm validation method. The enterprise should plan on a scenario with at least one dossier supplementation round and prepare the performance validation document set on independent data from the time of filing. An AI product approved by the FDA, the EU or a reference authority has a documentation advantage but is not exempt from registration in Vietnam.
Frequently asked questions
Does AI diagnostic software count as a medical device?
Yes. Under Decree 98/2021, software determined for a medical purpose such as diagnosis, monitoring or treatment is a medical device (SaMD) even without hardware, and must be classified and declared or registered before being placed in a healthcare facility. Cloud SaaS supply does not change this characterisation; purely administrative software such as appointment management is not a device.
Which risk class does medical AI fall into?
Classes A to D under Articles 4 and 5 of Decree 98/2021 and the rules in Circular 05/2022, based on the significance of the software information for clinical decisions and condition severity. AI supporting diagnosis of serious diseases such as cancer or stroke is usually class C or D; classes A and B are declared at the Department of Health, classes C and D are registered at the Ministry of Health. The named establishment self-classifies and is responsible for the result.
Are there rules on medical AI training data?
There is currently no dedicated provision on training data in device law; this is an open gap. Governance runs through three frames: the registration dossier must describe training and validation data and representativeness for Vietnamese patients; health data is sensitive personal data under the 2025 Law on Personal Data Protection requiring a collection basis and anonymisation; and algorithm updates changing registered performance or purpose require change registration.
How long does market authorisation for medical AI take?
Classes A and B take the declaration route and are fast, depending mainly on enterprise dossier preparation. Classes C and D require registration appraisal at the Ministry of Health, and AI dossiers usually face at least one supplementation round on algorithm validation and clinical evidence, so planning should assume that scenario. FDA or EU approval is a documentation advantage but does not exempt the Vietnam procedure.
This article has been quality-reviewed by the FLAW AI Legal Council