The safest way now to use LLMs is for simple entity extraction of signs, symptoms, investigation summaries, then translate them into the inputs for an understan

The safest way now to use LLMs is for simple entity extraction of signs, symptoms, investigation summaries, then translate them into the inputs for an understandable linear/logistic model, e.g. Wells criteria, Canadian CT head, or Centor score. I feel that a comprehensive but explainable model that supports multiple diagnoses will be developed in the future, but no such model currently exists.