The Ministry of Food and Drug Safety (MFDS) of Korea and the Health Sciences Authority (HSA) of Singapore have jointly released guiding principles for conducting clinical trials for machine learning-enabled medical devices (MLMDs). These guidelines aim to address the unique challenges posed by MLMDs in clinical research, ensuring they meet rigorous standards for safety, efficacy, and regulatory approval.
The Ministry of Food and Drug Safety (MFDS) of Korea and the Health Sciences Authority (HSA) of Singapore have jointly released guiding principles for conducting clinical trials for machine learning-enabled medical devices (MLMDs). By establishing clear principles for clinical trial design, patient selection, data integrity, and results analysis, this framework facilitates the efficient market entry of MLMDs while maintaining patient safety and clinical reliability.
Key Considerations in Clinical Trial Design
Clinical trials for MLMDs require careful planning to ensure scientific validity, ethical compliance, and meaningful clinical insights. Key factors include:
- Trial Design Selection – Single-arm, parallel, crossover, or other configurations should align with the study objectives.
- Statistical Hypothesis Formulation – Clearly defining primary and secondary endpoints is crucial.
- Randomization & Blinding – Helps reduce bias and improve study reliability.
- Sample Size Calculation – Ensures sufficient statistical power to detect meaningful effects.
- Regulatory Compliance – Trials must adhere to local laws, human subject protections, and data privacy regulations.
Unlike traditional medical device trials, retrospective studies using pre-existing datasets may be considered for MLMDs. However, they cannot fully assess usability issues, unintended workflow consequences, or real-world clinical performance. Additional studies, such as usability testing, may be required.
Patient & Test Dataset Selection
Ensuring that clinical participants and test datasets accurately reflect the intended patient population is essential for producing generalizable and unbiased results.
- Inclusion & Exclusion Criteria: These must align with the device’s intended use, considering target demographics, disease conditions, and clinical workflows.
- Independence of Datasets: Training datasets used for ML model development must not overlap with retrospective or prospective test datasets to prevent bias.
- Minimizing Bias: Randomization and blinding techniques should be used in both prospective and retrospective trials to ensure unbiased patient assignment and outcome assessment.
By addressing these factors, clinical trials can generate robust and meaningful evidence for regulatory approval.
Clinical Reference Standards & Data Interpretation
A clinical reference standard is an objectively determined benchmark used for comparison, validation, and training of MLMDs.
When available, reference standards should be derived from recognized clinical protocols. In novel use cases without established guidelines, expert panels may define reference benchmarks. Disagreements in clinical data interpretation should be documented and resolved transparently, using expert adjudication or consensus-building methods. To maintain trial integrity, reference standard evaluators should be independent from clinical investigators to reduce bias in outcome assessments.
Primary Endpoints and Results Analysis
The primary endpoint in an MLMD clinical trial is the main outcome used to evaluate device performance.
- Sensitivity & Specificity – Measures the device’s accuracy in detecting true positives and negatives.
- Positive & Negative Predictive Values (PPV/NPV) – Assesses the likelihood of correct diagnoses.
- Number Needed to Treat (NNT) – Evaluates clinical effectiveness.
- Area Under the Curve (AUC) – Determines the overall diagnostic ability of the ML model.
- Statistical Analysis & Justification.
Sponsors are responsible for defining acceptance criteria for trial outcomes and providing scientific justification for their selection. This ensures regulatory transparency and clinical validity.
Conclusion
The MFDS-HSA guiding principles represent a major step forward in regulating machine learning-enabled medical devices. By establishing a structured framework for clinical trials, these guidelines support: regulatory harmonization across jurisdictions, faster market entry for AI-driven medical innovations, improved patient safety and clinical efficacy. As MLMD technology continues to evolve, ongoing collaboration between regulators, researchers, and industry stakeholders will be essential to refine these principles and ensure the safe, ethical, and effective use of AI in healthcare.
Link to the document:
