The biggest limitation I find about prolog is that the model is limited to the vocabulary and knowledge of the author. If the author only speaks English, the prolog program only works in English. If the author doesn't know anything about medicine, it will be very hard to model all the medicinal knowledge in the world (which could even be in different language).

That's where Machine Learning and Neural Networks are much better : they can organize enormous amount of arbitrary data (not only text and words) way more faster that a programmer or an expert, given a sufficiently big enough dataset.

In other terms, prolog "doesn't scale" with the diversity and ambiguity of humans knowledge.