Survival analysis (time-to-event modeling)

Many clinical outcomes, such as recurrence, recanalization, reoperation or death, are about when something happens, not only whether it happens. Survival analysis models the time to an event and produces curves showing the probability of remaining event-free over time.

Its key feature is handling censoring: a patient last seen at two years without the event still contributes information up to two years, even though their later outcome is unknown. Ignoring censoring or treating such patients as event-free biases results.

Classic methods include Kaplan-Meier curves and Cox proportional hazards models. Machine-learning versions include random survival forests and gradient-boosted survival models, which are judged with the concordance index and time-dependent calibration.