After addressing the challenges of historical spelling variants and OCR errors, we show that document classification attains high accuracy, and that the feature weights can be interpreted historically and linguistically, although with a high level of noise. Further, we were surprised how accurately topic models allow us to trace socio-historical changes, for example the change from scholastic thinking to empirical science in medical studies, and how professional health care replaced medieval quackery. Both of these approaches are robust to parameter details, as long as stopword lists are used and OCR errors or historical spelling variants are addressed.