Predicting the risk of emergency admission with machine learning: Development and validation using linked electronic health records