Good AI Practice Principles 5–7: Development

Principles 5–7 address who should be involved in AI development, how data must be managed, and how models should be built. Together they ensure AI outputs contributing to drug evidence are traceable, fit-for-purpose, and safe for patients.

Principle 5: Multidisciplinary expertise

Multidisciplinary expertise covering both the AI technology and its context of use are integrated throughout the technology's life cycle.

Relevant expertise may include:

• Clinicians and clinical pharmacologists
• Statisticians and data scientists
• Regulatory affairs and quality professionals
• Informatics and software engineers
• Pharmacovigilance specialists

Expertise should not be confined to initial model building. It must inform deployment, monitoring, re-evaluation, and communication to users.

For Hong Kong research teams

Local trial sites should ensure clinical investigators and pharmacists participate in defining context of use and reviewing whether AI outputs are clinically meaningful — not only data science teams remote from patient care.

Principle 6: Data governance and documentation

Data source provenance, processing steps, and analytical decisions are documented in a detailed, traceable, and verifiable manner, in line with GxP requirements.

Appropriate governance is maintained throughout the life cycle, including privacy and protection for sensitive data.

Key expectations:

• Know where data came from and how it was transformed
• Maintain audit trails for regulatory inspection readiness
• Protect patient confidentiality in line with PDPO and study protocols

Principle 7: Model design and development practices

The development of AI technologies follows best practices in model and system design and software engineering and leverages data that is fit-for-use, considering interpretability, explainability, and predictive performance.

Good model and system development promotes:

• Transparency
• Reliability
• Generalizability
• Robustness

…for AI technologies contributing to patient safety.

Practical implications

When evaluating an AI vendor or internal model:

• Is the training data fit-for-use for your indication and population?
• Can the team explain key model decisions and limitations?
• Does design documentation support reproducibility and future audits?

A high-accuracy model built on poorly governed or misaligned data remains a regulatory and clinical risk.

Source: FDA & EMA — Guiding Principles of Good AI Practice in Drug Development (January 2026)