EVERBIZ | Original compilation. Please credit the source when republishing.
The Center for Food and Drug Inspection (CFDI) of China’s National Medical Products Administration (NMPA) has published the Key Inspection Points for Artificial Intelligence Medical Devices on its website. The document is aimed at inspectors, and also at registrants and contract manufacturers. It covers standalone AI software and products that contain an AI software component. In one sentence: the “production site” of an AI device is in its data and code, and quality system inspections are moving there too.
Status note: On 8 October 2026 the CFDI opened a public consultation on a draft of the Key Inspection Points for AI Medical Devices (comments due 28 October). The November text this article relies on carries no “draft” marking and states no effective date. Please confirm the effective date on the CFDI website.
1. The core idea of the document
Traditional device inspections look at buildings, equipment and batch records. An AI product is a virtual product. The document states that its “production” actually takes place in the design and development stage: building datasets, training algorithms and writing software code. Delivery is either physical or over the network.
The document therefore brings the whole chain “data → algorithm → verification → change → delivery → post-market” into the quality management system. It sets out 14 key risk points from requirements analysis to software delivery, each with control measures.
EVERBIZ view: This is a manual that applies GMP (Good Manufacturing Practice for medical devices) to AI software. Companies used to find their AI development process hard to explain. From now on, inspectors will have a clear yardstick.
2. The six new requirements worth most attention
1. Data labelling: three roles may not overlap
Labellers, reviewers and arbitrators may not hold each other’s roles. Where auto-labelling software is used, its output must be reviewed by a person. Labelling rules, staff qualifications, dispute arbitration and quality assessment all need documents and records.
2. Datasets: pairwise disjoint, no “reverse flow”
The training, tuning and test sets need a defined splitting method and ratio, and the three must be pairwise disjoint, with duplicate-check records kept. Changes in a dataset’s use must be traceable, and training and tuning sets may not be reused as test sets. The distribution of the test and tuning sets should also reflect real-world conditions.
3. Pretrained models must be “registered”
Where a pretrained model is used, you must list its name, version, source, licence and restrictions, and assess its compliance and security. This affects most directly companies that build on open-source large models or foundation models.
4. Performance evaluation: independent test set + five dimensions
Algorithm performance evaluation must be based on an independent test set and cover false negatives/false positives, repeatability, fairness, robustness and real-time performance. It must also produce an analysis report on the factors affecting performance. “Fairness” is explicitly written into the inspection points, which is worth noting.
5. Cybersecurity: vulnerability scanning is mandatory
Requirements, design, source code, test and risk management documents must be traceable and consistent. Vulnerability scanning is mandatory, while source code review, penetration testing and fuzz testing are optional. Data processing takes place in a controlled network, and a cybersecurity incident response procedure is required.
6. Network delivery and algorithm updates now have clear rules
- For network delivery, the upload record is the production record, and system logs may count as records;
- The UDI (unique device identification) is shown as human-readable text in the software interface, and products without an interface must be able to send it via an API (application programming interface);
- Network delivery needs tamper protection, identity authentication, integrity checks, and exception handling such as retransmission and resumable transfer;
- Continuous-learning products must have self-learning confirmed as switched off before release;
- Algorithm updates need an update study report, major changes must be filed as registration changes as required by law, and algorithm stability and accuracy must be assessed regularly after launch.
3. The 14 risk points at a glance
| Stage | Risk point |
|---|---|
| Requirements | ① Requirements analysis |
| Data | ② Data collection ③ Data preparation ④ Data labelling ⑤ Dataset construction ⑥ Dataset quality control |
| Algorithm | ⑦ Algorithm design ⑧ Algorithm selection ⑨ Algorithm training ⑩ Algorithm performance evaluation |
| Verification and change | ⑪ Software verification and validation ⑫ Change control |
| Security and delivery | ⑬ Network and data security ⑭ Software delivery |
The appendix also lists types of AI devices already approved, in four categories: triage support, detection support, diagnosis support and treatment support. Examples include software for CT (computed tomography) lung nodule detection, fundus image diagnosis of diabetic retinopathy, and radiotherapy contouring.
4. What companies should do now
- Back-fill records: go back over the process records for data collection, de-identification, labelling, splitting and training. In an inspection, a process without records counts as one that never happened.
- Cross-check: run a duplicate check on existing datasets and confirm the test set never took part in training or tuning.
- Take inventory of models: build a list of pretrained models and third-party databases, and complete the licences and assessment documents.
- Manage outsourcing: bring labelling vendors, cloud compute and third-party data providers under supplier management, with quality agreements.
- Set the rules: define version naming rules that separate major from minor updates, and link them to the decision on whether a registration change is needed.
5. How it fits existing Chinese rules, and what is still moving
How it connects to existing rules
- The inspection points build on GMP and its standalone-software annex and are used for on-site inspections. Beijing’s 2023 quality system guide for AI devices is built on the same documents;
- On the registration side there is the 41-page Guideline for Registration Review of AI Medical Devices issued by the CMDE in March 2022, and before it the 2019 Review Points for Deep Learning Decision-Support Software;
- The 2019 review points already asked for defined selection, training and assessment of labelling and arbitration staff. The new inspection points add that the three roles may not overlap.
Related items still in progress
- In September 2026 the CMDE opened consultation on a draft guideline for clinical evaluation of multi-disease AI decision-support devices (comments due 14 October 2026);
- A June 2026 media report says Class III imaging AI that gives a benign/malignant judgement must validate its core algorithm in a clinical trial (media report, official text to be confirmed);
- A brokerage research note mentions the national standard GB/T 47941.1-2026, AI medical device quality requirements and evaluation, Part 1: Terminology (to be confirmed).
6. Compared with the EU: where it differs
The EU requires dual compliance for AI medical devices: both MDR/IVDR and the AI Act. The application date of the AI Act’s high-risk rules for Annex I products (including medical devices) has moved from 2 August 2027 to 2 August 2028 (Regulation (EU) 2026/1744, published in the EU Official Journal on 24 July 2026 and in force since 27 July).
This does not loosen MDR/IVDR requirements: existing products are still reviewed mainly under MDR/IVDR, and the AI Act requirements are layered on top later.
| Area | China inspection points | EU (MDR/IVDR + AI Act) |
|---|---|---|
| Data requirements | Specific and directly checkable: three labelling roles kept separate, datasets pairwise disjoint with duplicate-check records | Principles: data must be representative, unbiased and documented; training, validation and test sets kept separate and justified |
| Fairness / bias | One of the performance evaluation metrics | Broader: risk management must cover impact on fundamental rights and systemic bias |
| Human oversight / explainability | Mainly through use restrictions and instructions for use | Separate AI Act requirement: transparent design, explainable output, human oversight |
| Continuous learning | Self-learning must be off before release | Allowed through a “predetermined change” and a change control plan setting the acceptable range |
| Who assesses | Registration review + on-site inspection by CFDI or the provincial authority | One notified body can assess both laws, but each set of criteria must be met |
What companies going abroad can do (EVERBIZ analysis):
- Data and labelling records built to the Chinese inspection points can largely serve as the “evidence layer” for EU data governance;
- Add three more things: bias and fundamental-rights assessment, human oversight design, and a predetermined change control plan;
- Plan continuous-learning products by market: self-learning off for China, and a change control plan to lock the update range for the EU.
EU certification route and standards
- For AI medical devices that need a notified body, the AI Act conformity assessment follows the MDR/IVDR procedure: one combined assessment, one technical file, one declaration of conformity, one CE mark;
- The AI Act adds requirements for notified bodies, such as AI expertise, and some analysts expect their AI assessment capacity to be tight, so raise the timeline with your notified body early;
- 2 August 2028 is not extra time for MDR/IVDR work: clinical, quality system and CE marking work continues under the current requirements;
- EN 18286:2026 is the first harmonised standard supporting the AI Act (AI quality management system). Commission Implementing Decision (EU) 2026/2274 (12 October 2026, in force 14 October) concerns harmonised standards under the AI Act. Several commentators read it as giving a presumption of conformity with the Article 17 quality management requirements. The standard is voluntary and does not cover Article 17(2) to (4); the decision’s full text prevails;
- Clause 0.1 of EN 18286 says companies can extend an existing sector quality system rather than build a separate one, with EN ISO 13485 for medical devices as the example.
US comparison: In December 2024 the US FDA issued final guidance letting manufacturers get certain modifications to AI-enabled device software functions pre-authorised through a predetermined change control plan (PCCP), so each planned change does not need a new submission. The plan includes a description of modifications, a modification protocol and an impact assessment. The idea is close to the EU’s “predetermined change”, so companies targeting the US as well can plan for both.
EVERBIZ view: The two systems point the same way but are not identical: China leans towards an operational checklist, the EU towards principles with two laws layered together. Records built to the Chinese requirements can be reused, but the three items above still need to be added before “one set of documents, two markets” works.
Based on this document, EVERBIZ has prepared the Self-Inspection Checklist for Quality Management Systems of AI Medical Devices (14 risk points + 48 clauses by chapter, editable version). To get the checklist, or for support with AI medical device registration in China, quality system inspection coaching, and EU / Southeast Asia registration, please contact us.
Sources (official first)
- CFDI: 人工智能医疗器械检查要点 (Nov 2026) (official)
- CFDI consultation notice on the draft, 8 Oct 2026 (repost) (consultancy or media)
- Commission Implementing Decision (EU) 2026/2274, EUR-Lex (official)
- White & Case: EU AI Omnibus enters into force (Reg. (EU) 2026/1744) (law firm or certification body)
- Cuatrecasas: Digital Omnibus on AI has been published (law firm or certification body)
- OpenRegulatory: MDCG 2025-6 explained (consultancy or media)
- BSI: conformity assessment of AI-enabled medical devices (slides) (law firm or certification body)
- Pure Global: Annex I high-risk dates move to 2 Aug 2028 (consultancy or media)
- Modulos: EN 18286 summary (consultancy or media)
- Beijing Medical Products Administration: AI device QMS guide (2023) (official)
- Jiemian: 2022 NMPA AI registration review guideline (consultancy or media)
- CMDE draft clinical-evaluation guideline notice, 16 Sep 2026 (repost) (consultancy or media)
- Foley: FDA final guidance on PCCPs for AI-enabled devices (law firm or certification body)
Source: CFDI, Key Inspection Points for Artificial Intelligence Medical Devices (cfdi.org.cn, November 2026). This article is EVERBIZ’s interpretation of the document and is not an official interpretation; the original text prevails. Unofficial English translation of the Chinese original.
EU sources: Regulation (EU) 2026/1744 in the EU Official Journal (24 July 2026) and law-firm commentary; the Official Journal text prevails on dates.