During development you turn the design-phase decisions into a working AI system: you manage the data, train the model, protect patient privacy, design the interface for clinicians and document what you did.
Prerequisites¶
| Prerequisite | Importance | Level |
|---|---|---|
| Completed design phase checklist | Necessary | n/a |
| Data access agreements in place | Necessary | n/a |
| Approval from an ethics committee or local review board, if your project needs one (see ethical review) | Necessary if applicable | n/a |
| Understanding of machine learning | Necessary | Intermediate |
| Understanding of GDPR and data governance | Helpful | Beginner |
Learning outcomes¶
This section has seven chapters. When you have worked through them, you should be able to:
Collect training data that represents the patients and settings where the system will be used, and record where every data point came from
Check data quality and the representation of patient groups, and choose a bias correction where one is needed
Apply pseudonymisation, access control and logging that meet the GDPR and Dutch healthcare security standards
Train a model, compare it with a simple baseline and check that its risk estimates are calibrated (a predicted 20% risk should match about 20% observed events)
Design how the output reaches clinicians so that they can review, question and override it
Keep the model card, data documentation and code history that later evaluation and regulatory review depend on