Modelling

• Risk prediction, diagnosis support and phenotyping as distinct problems
• Outcome definition, look-ahead bias and the label leakage endemic to EHR work
• Clustering for patient subgroups and the fragility of the resulting clusters

Validation

• Discrimination against calibration, and why calibration decides clinical usefulness
• External and temporal validation, and performance drop across sites
• Decision curve analysis and net benefit over existing practice

Safety

• Documented cases where a clinical model encoded a resource proxy as need
• Subgroup performance reporting rather than a single aggregate metric
• Automation bias and the effect of a wrong recommendation on a clinician

Deployment

• Software as a Medical Device, CDSCO and FDA pathways in outline
• Workflow integration, alert fatigue and the reason good models go unused
• Post-deployment monitoring, drift and the governance that owns it