The European Water Framework Directive (WFD 2000) brought the need in European Union countries to establish consistent quantitative methods for the water quality assessment of streams, using aquatic communities. With this work we aimed to develop predictive models using macroinvertebrate communities that could be used in Portugal as an alternative to the more traditional indices and metrics. We used data from 197 reference sites and 174 sites suspected of being impaired, which were obtained in a national survey conducted in 2004–2005 by the Instituto da Água (INAG, Portugal). The spatial scale at which to develop predictive models was an issue to address because the Portuguese territory covers a wide variety of landscapes in a small area. We built three models using the AUSRIVAS methods, a national and two regional (North and South) models that produced acceptable assessments. However, the regional models, predicted more taxa than the National model, were more accurate and had lower misclassification errors when placing sites into pre-defined groups. The regional models were also more sensitive to some disturbances related to water chemistry (e.g., nutrients, BOD5, oxidability) and land use. The exception was for the northern costal area, which had few reference sites. In the northern costal area the National model provides more useful results than the regional model. The 5-class WFD quality assessment scheme, adapted from the AUSRIVAS bands, appears to be justified because of the good correspondence between the human disturbance level and the classes to which test sites were allocated. Elimination of the AUSRIVAS X band in the WFD scheme has produced a clearer relationship. The predictive models were able to detect a decline in river health, responded to several causes of degradation and provided site-specific assessments.