This paper presents a methodology for the comparative analysis of human and machine translation at a lexical-terminological level. This proposal is applied to an English-Spanish parallel corpus of specialized texts in the medical domain. The main aim of the study is to explore systematic linguistic differences between machine and human translation in the light of the problem of automatic system evaluation. The specific objectives are: a) detect differences in the distribution of terminological units between human and machine translation, b) identify the conditions under which such differences occur considering original texts and strategies for human and machine translation. The study methodology involves, on the one hand, the use of stylometry techniques to characterize the language of machine translation versus human translation and, on the other hand, the classification of translation shifts performed by human translators and modifications made by machine translation systems to the original text. The research results indicate that the differences between machine translation and human translation related to optional translation shifts performed by translators and the differences related to the lack of obligatory changes in machine translation are not equally important to assess the quality of the latter.