Choosing leading indicators that actually predict something

Lost time injury frequency is a lagging indicator and a poor one at low injury counts — a site can run for two years without a recordable and still be one procedural failure away from a fatality. So everyone adds leading indicators, and most of the ones chosen measure activity rather than risk: number of inspections done, number of talks held, percentage of training complete. An indicator earns its place if a bad reading would make you change something. Which of yours have ever done that? And has anyone found a leading indicator that flagged a serious event before it happened?