Online glucose prediction which can be used to provide important information of future glucose status is a key step to facilitate proactive management before glucose reaches undesirable concentrations. Based on frequency-band separation (FS) and empirical modeling approaches, this article considers several important aspects of on-line glucose prediction for subjects with type 1 diabetes mellitus. Three issues are of particular interest: (1) Can a global (or universal) model be developed from glucose data for a single subject and then used to make suitably accurate on-line glucose predictions for other subjects? (2) Does a new FS approach based on data filtering provide more accurate models than standard modeling methods? (3) Does a new latent variable modeling method result in more accurate models than standard modeling methods? These and related issues are investigated by developing autoregressive models and autoregressive models with exogenous inputs based on clinical data for two groups of subjects. The alternative modeling approaches are evaluated with respect to on-line short-term prediction accuracy for prediction horizons of 30 and 60 min, using independent test data. (c) 2013 American Institute of Chemical Engineers AIChE J 60: 574-584, 2014