Forecasting with Sparse but Informative Variables: A Case Study in Predicting Blood Glucose