Predicting the risk of emergency admissions: development of machine learning models for a large cohort of linked electronic health records