Section F
Health Data & AI
What data and algorithms can and cannot say.
What Is Health Data?
Health data includes many kinds of information about health, care and people's experiences, each with its own strengths and limits.
F02The Health Data Lifecycle
The health data lifecycle covers how information is collected, used, stored, shared, retained and eventually disposed of.
F03Data Quality and Missingness
Data quality depends on fitness for purpose, while missing information can reflect both recording problems and meaningful patterns in care.
F04Labels and Ground Truth
Reference labels provide targets for training and evaluation, but their meaning and reliability depend on how they are created.
F05Training, Validation and Test Sets
Separating data by purpose helps estimate how a model may perform on new cases and reduces misleading results from information leakage.
F06Clinical AI: Prediction Is Not a Decision
A model output can inform care, but deciding what to do requires clinical context, evidence, patient preferences and clear accountability.
F07Model Performance Measures
Different performance measures describe different strengths and errors, so no single score can establish whether a clinical model is useful.
F08Calibration and Decision Thresholds
Calibration concerns the reliability of predicted probabilities, while decision thresholds determine when an output leads to a particular action.
F09Generalisability and Dataset Shift
A model's performance can change when the people, measurements or care processes in use differ from those in its development data.
F10Fairness and Subgroup Evaluation
Subgroup evaluation can reveal unequal model performance, but fairness also depends on data, access, clinical consequences and how a system is used.
F11Explainability and Its Limits
Explanations can help people inspect model behaviour, but they do not establish that an output is correct, causal or clinically useful.
F12Human Oversight and Automation Bias
Effective human oversight requires clear responsibilities and practical support, because simply placing a person in the process does not prevent over-reliance.
F13Monitoring and Updating Clinical AI
Ongoing monitoring and controlled updates help identify changing performance, investigate problems and reassess whether a clinical AI system remains suitable.