Evaluation metrics

Measures used to assess a model, dataset or review process against a defined objective.

The choice of metric determines what a result means. Overall accuracy, error detection recall, expert agreement and review throughput answer different questions. Define the positive class, denominator and sampling method, and report important slices so a strong average does not conceal weak performance on rare cases.

In practice

A pilot tracks missed annotation errors alongside turnaround time rather than reporting speed alone.