By the end of the development phase, you should have the following in place.
Data
Training, validation and test sets that reflect the patients and settings of intended use, split at patient level
Documented data sources, preprocessing steps and annotation protocols
Privacy measures in place (pseudonymisation, access control, access logging)
Model
A trained model with documented architecture, hyperparameters and training procedure
Performance measured on the validation set and compared with a simple baseline model
Calibration checked and a first fairness analysis across patient subgroups
Tested behaviour for missing and implausible inputs, and a written threat model; the full robustness evaluation follows in the evaluation phase
Human-AI interface
Interface prototypes designed and tested with representative users
Mechanisms for human oversight and correction designed
Documentation
Draft model card or technical documentation
Data lineage documented
Version control in place
Design decisions and their reasons recorded