[fishbuff]

it's straightforward to train a decision tree to do the job, but too many exceptions which are difficult to handle, like medallists, low iq, mental and depression cases, those with a bright future and had contributed to society and etc. so it's better to leave it as is

not with decision trees. you have your use NLP and identify all the entities first (see below) you will need to vectorise the entities, and discover the relationship among the entities, for that  i use association rules mining. i prefer to use neural network to do the classification but if you triage/classify by history or would you want to refer to latest criteria to decide the latest classification. so, i did a comprehensive list of medical consideration critieria, then perform cosine similarity to find the nearest match. but this can be improved if we can build a model of the convict where we place both vectorised legal entities of defence and accuse. Sorry, didnt invest much time to think about this unless I get a commission to do so. I did want to do it for AU's migration law where AAT's appeal cases are available on AustLii website.