Going from the body of your question:
but are there metrics or models that I can use to help predict which words are likely to decay over time?
No, none, since words do not generally change much individually: what happens is that sequences of sounds undergo joint mutation, and generally all words containing those sequences will undergo the changes together. This covers strictly the phonetic part of language change (even ignoring the maldefinedness of "decay" in a linguistic context).
Further changes of word properties may be morphological and paradigmatic (morphological levelling, analogy, etc.), as well as individual (or of a narrow group) and semantic (changes in meaning of a word or several words in one of several directions, including amelioration and generalisation). This too is generally unpredictable.
The best tendencies (read: only nearly accurate predictions) can be draughted for languages with taboo word replacement tendencies: in such systems, words considered taboo will be replaced on a regular basis. This works for names of certain predatorial animals, diseases, and names of dead relatives (among other possible taboo categories). This is generally poorly quantifiable and predictable, just like the rest of these processes.