Empirical evidence of the impact of study characteristics on the performance of prediction models: a meta-epidemiological study