This doesn't prevent fingerprinting. If you send me enough data, even if it is not associated with an ID, I can still use that data to deanonymize or track you. People should read the papers on DP instead of just randomly commenting "you could just do X". The security researchers studying the threat models aren't just doing things in an overly complicated way for no reason.
> ML locally
ML training depends heavy on the training data set. You may not have all of the training data, and the training data may involve other people's private data.
Also, your phone is not the best place to run training for a model that takes 8 hours of training time and gobs of memory. However, you can slice up the problem, and people's phones can work together -- Federated Learning, which goes hand in hand with differential privacy.
You could run ML locally.
If you want to know how long patients are waiting, you can just record wait time without recording details of the patient.