Data¶
Training data reflects the demographic and clinical diversity of the intended patient population
Data sources documented with provenance (site, period, system, equipment)
Preprocessing pipeline written as code and under version control
Annotation guidelines written and applied consistently
Agreement between annotators measured and documented
Attributes needed for fairness analysis (such as age and sex) recorded, with a lawful basis
Train, validation and test split made at patient level, with no leakage between sets
Privacy and security¶
Patient data pseudonymised or anonymised
Role-based access controls in place, with access to patient records logged (NEN 7513)
Lawful basis under GDPR Art. 6 and condition under Art. 9 documented, and DPIA completed
Re-identification risk assessed, including for any model you plan to share
Threat model written: which attacks are in scope and which measures address each
Model¶
Model architecture and training procedure documented
Hyperparameters and training configuration under version control
Simple baseline model (clinical rule or logistic regression) trained and compared
Performance measured on the validation set
Calibration checked and, if needed, corrected
First subgroup performance analysis completed
Uncertainty quantification implemented or its absence justified
Behaviour with missing or implausible inputs tested
Human-AI interface¶
Interface designed with input from representative users
Mechanisms for human oversight, override and error reporting designed
Explanations integrated where users said they need them
Documentation¶
Model card or equivalent technical documentation drafted
Data documentation (datasheet) drafted
Version control repository in use for code, configuration and documentation
Design decisions and their reasons recorded