The Taiwan Food and Drug Administration (TFDA) has issued a new guidance document, "Good Machine Learning Practice (GMLP): Development and Management Principles for Artificial Intelligence Medical Devices". The guidance applies to medical devices utilizing machine learning technologies and establishes expectations for the entire product lifecycle, from design and development to post-market monitoring.
Purpose of the Guidance
As AI and machine learning technologies become increasingly integrated into healthcare, TFDA aims to ensure that AI-enabled medical devices remain safe, effective and clinically meaningful throughout their lifecycle. The guidance is based on the IMDRF's Good Machine Learning Practice principles and other international standards, emphasizing risk management, data quality, transparency, and continuous performance monitoring.
Scope
The guidance applies to medical devices that incorporate machine learning technologies. Manufacturers are encouraged to determine the applicability of each principle according to the device's intended use, target population, performance claims, and use environment.
Key Principles
1. Multidisciplinary Approach Throughout the Product Lifecycle
Manufacturers should clearly define the device's intended use, clinical context, expected benefits, and patient risks. Leveraging multidisciplinary expertise throughout the product lifecycle helps ensure clinically meaningful performance and supports the safety and effectiveness of AI-enabled medical devices.
2. Robust Software Engineering and Cybersecurity Practices
Robust software engineering, quality management, cybersecurity, and systematic risk management should be implemented throughout the product lifecycle. Main points that should be considered include software traceability, reproducibility, authenticity, confidentiality, integrity, and availability.
AI-specific risks should be addressed using a total systems safety approach, supported by standards such as ISO 14971, AAMI TIR 34971, and ISO/IEC 23894. Ethical use of patient data, including de-identification and appropriate consent, should also be ensured.
3. Representative Clinical Data
Datasets used for model training, testing, and monitoring should be sufficiently representative of the intended patient population and use environment. Models trained on non-local data should demonstrate applicability to Taiwanese populations. Manufacturers should proactively address unintended bias and dataset drift to ensure robust and generalizable performance, with reference to standards such as ISO/IEC TR 24027 and the ISO/IEC 5259 series. Clinical investigations conducted in Taiwan should comply with applicable TFDA Good Clinical Practice requirements.
4. Independent Training and Test Datasets
Training and test datasets should be maintained independently, with potential dependencies arising from patients, sites, or data acquisition carefully controlled. The scope and rigor of external validation should be proportionate to the device's risk profile.
5. Appropriate Reference Standards
Reference standards should be based on clinically accepted methods and aligned with the device's intended use. Manufacturers should document the rationale for their selection and, where recognized standards exist, apply them during model development and validation to support robust and generalizable performance. Reference standard selection should be guided by expert knowledge and, where possible, broad professional consensus.
6. Model Selection and Design Based on Intended Use
Model selection and design should be evaluated and demonstrated to be appropriate for the available data, while supporting the proactive mitigation of known risks, such as overfitting, performance degradation, and cybersecurity risks. Performance objectives should be clinically meaningful and support the device's intended use. Manufacturers should also consider the impact on patient subgroups and account for variability and uncertainty in inputs, outputs, and clinical conditions.
7. Focus on Human-AI Interaction
AI-enabled medical devices should be evaluated within their intended clinical workflows, with particular attention to human-AI interaction and usability factors. Manufacturers should consider users' expertise, understanding of model limitations, and the risk of overreliance on AI. Critical medical decisions and interventions should remain under the supervision of healthcare professionals, and key processes requiring human oversight should be clearly identified.
8. Testing Under Clinically Relevant Conditions
Manufacturers should establish and execute test plans based on sound methodological and statistical principles to generate clinically relevant performance data using datasets that are independent of the training data. Testing should take into account factors including the intended patient population and relevant subgroups, the clinical environment, the real-world use of the "human-AI team," measurement inputs, and potential confounding factors.
9. Transparent Information for Users
Manufacturers should provide intended users with clear and context-appropriate information regarding the device's intended use, benefits, risks, model performance, training and testing data characteristics, acceptable inputs, and known limitations. Users should understand how AI outputs are integrated into clinical workflows and, where possible, the rationale behind those outputs. Information on product updates and channels for reporting concerns should also be made available.
10. Continuous Real-World Performance Monitoring
Manufacturers should continuously monitor the real-world performance of deployed AI models using a risk-based approach to maintain device safety and effectiveness. Model retraining and software updates should be controlled within the quality management system to mitigate risks such as overfitting, unintended bias, and model degradation. Post-market surveillance programs should track adverse events and real-world performance, while AI governance mechanisms may be integrated into existing quality systems with reference to ISO/IEC 42001. Manufacturers may also consider implementing a Predetermined Change Control Plan (PCCP) to proactively manage lifecycle changes.
Regulatory Impact
This guidance reinforces TFDA's alignment with international regulatory trends and underscores the importance of lifecycle management for AI-enabled medical devices. Manufacturers should consider:
- Strengthening quality management and AI governance systems
- Ensuring representative and clinically relevant datasets
- Evaluating the applicability of foreign-trained models to Taiwan's population
- Implementing robust post-market monitoring mechanisms
- Considering a Predetermined Change Control Plan (PCCP) for future model updates
Conclusion
TFDA's new GMLP guidance represents a significant step toward harmonization with global regulatory expectations for AI-based medical devices. By emphasizing data quality, human oversight, transparency, and continuous monitoring, the guidance provides manufacturers with a framework to support the safe and effective development and lifecycle management of machine learning-enabled medical devices.
Manufacturers developing AI/ML-based medical devices for the Taiwanese market are encouraged to review their current development processes and quality systems to ensure alignment with these newly established principles.
Planning to introduce AI-enabled medical devices into Taiwan? With a local office in Taipei and 26+ years of regulatory experience, Qualtech supports organizations in navigating TFDA requirements, product registration, and ongoing compliance for medical devices across Taiwan and the Asia-Pacific region. Contact us now!