Taiwan FDA Updates AI Medical Device Technical Guidelines

The TFDA (Taiwan Food and Drug Administration) implemented new AI medical device guidelines in Taiwan on 14 August 2025.

The formal name of the revised technical guidelines issued by the TFDA is:

• "Technical Guidelines for Inspection and Registration of Computer-Aided Detection (CADe) and Computer-Aided Diagnosis (CADx) Medical Devices Based on Artificial Intelligence/Machine Learning" (hereinafter referred to as the "new guidelines").

Alongside this, the TFDA published:

• "Frequently Asked Questions on the Independent Performance Evaluation of Artificial Intelligence/Machine Learning Medical Devices".

Together, these publications aim to keep pace with rapid advances in AI healthcare technology and to standardise the regulatory review process to ensure their safety and effectiveness. Both came into effect immediately, replacing the previous version issued on 14 September 2023.

What’s Covered in the New Guidelines

The new guidelines mainly target Class II and III AI/ML CADe and CADx medical devices, whether embedded systems or independent software (SaMD). The document sets clear standards from multiple dimensions, such as product structure, performance assessment, and clinical evidence, including:

Product Disclosure Requirements

Manufacturers need to detail:

• Intended use of the product
• Applicable population
• Usage scenarios (e.g. real-time assistance or secondary reading)
• Data processing procedures
• Output formats

The new guidelines emphasise that the failure scenarios of the product and the potential clinical risks (e.g. impact of false positives/negatives) must be listed.

Algorithm and Training Data Requirements

The new guidelines require companies to disclose details such as:

• Algorithm type (e.g. CNN, RNN, SVM)
• Development framework
• Model structure
• Optimisation functions
• Loss functions

In terms of data, the new guidelines clearly require that training sets, tuning sets, and test sets are strictly separated. Manufacturers must also provide a detailed description of data sources, collection protocols, population distribution, disease characteristics, etc., to ensure the representativeness and quality of the data.

Standalone Performance Assessment

This is one of the key points of this revision. It requires manufacturers to quantitatively assess the accuracy, localization consistency, diagnostic ability, generalizability, and stress testing of the product's detection under conditions where no new annotation is used (e.g. retrospective clinical data or public databases). Common indicators include sensitivity, specificity, AUC-ROC, etc.

The Frequently Asked Questions document further clarifies the definition and application of standalone performance assessment, confirming that this method is not limited to CADe/CADx products but applies to the performance verification of all AI/ML medical devices.

Clinical Evidence Requirements

Depending on the product risk and whether it is a completely new product, manufacturers may need to provide clinical evaluation reports or clinical trial data. Multi-reader-multi-case (MRMC) study designs are recommended, especially for high-risk or independent diagnostic products.

The Q&A document specifically points out:

1. Testing using public databases (e.g. TCIA, NIH Chest X-ray, etc.) is acceptable but attention must be paid to their representativeness, quality, and compliance
2. Independent performance assessment cannot completely replace clinical evaluation, especially for new products without similar products;
3. Whether ethical review (IRB) is needed depends on whether the data source involves the scope of human research regulations.

Insights for International Manufacturers

The new guidelines raise several points of particular importance for overseas applicants aiming to bring AI/ML medical devices into Taiwan's market. Below are three key areas where overseas applicants will need to pay close attention:

Localisation of Data and Diversity Requirements

The new guidelines repeatedly emphasize that test data must represent the target population. International manufacturers who rely solely on European and American data to train models need to supplement local clinical data to prove the applicability of their products to the Taiwanese population.

Algorithm Interpretability and Output Transparency

The TFDA requires products to clearly state the clinical significance and limitations of the output results. For example, if CADx products provide a malignancy risk score, it must explain its clinical basis and grading criteria, otherwise it may affect the registration process.

Systematic Lifecycle Management

Although the new guidelines do not explicitly mention "algorithm iteration" or "continuous learning,", by emphasizing training data management, generalization testing, and quality documentation requirements, they effectively require companies to establish a quality management system from development, acceptance, to post-market monitoring throughout the lifecycle.

Strategies for Overseas Manufacturers

To meet the TFDA’s updated expectations, overseas manufacturers should prepare early and adopt a proactive compliance strategy.

Key steps include:

• Conduct a Gap Analysis: Compare existing technical documents with new guideline requirements, especially in data management, test coverage, and clinical validation areas.
• Optimise Data Strategy: Actively cooperate with local medical institutions to establish local data sets that meet population characteristics and quality requirements.
• Improve Performance Assessment Process: Incorporate independent performance assessment into the regular product development process and develop a verification strategy that combines with clinical evaluation.
• Track Regulatory Updates: Cooperate with professional regulatory consulting companies, such as Cisema, to obtain timely regulatory updates.

If you are a medical device manufacturer with questions about how to register AI/ML products for the Taiwan market, get in touch with Cisema today.

Further Information

Looking to register medical devices in Taiwan? Cisema’s Taiwan FDA services can guide you through every step.

References

For those who wish to review the source materials, the TFDA has made both documents available:

• Original Announcement (14 August 2025): Click here to read the official release of the revised Technical Guidelines for AI/ML CADe and CADx Medical Devices.
• Technical Guidelines: 人工智慧／機器學習技術之電腦輔助偵測(CADe)及電腦輔助診斷(CADx)醫療器材查驗登記技術指引
• Frequently Asked Questions (FAQ): 人工智慧／機器學習技術之醫療器材獨立性能評估常見問答集