What is overfitting?

Why

A flexible model can achieve nearly perfect training performance by capturing incidental features rather than stable predictors. In small biological studies this is especially risky because correlated features, repeated donors and technical batches greatly outnumber independent units. Diagnose overfitting by comparing training and correctly separated validation performance, examining learning curves and using external cohorts where possible.

Control complexity with appropriate regularization, feature selection performed inside each training fold and adequate sample design. Strong internal validation still may not predict deployment at another hospital because the population and laboratory process differ. Report uncertainty rather than selecting a lucky split.

The core issue is poor generalization, not merely that the model has many parameters or that its training accuracy is high.