[samesense]

Yes. I've seen several examples of decision support since 2010. It's not easy to train models because the codes for procedures can be ambiguous. Furthermore, since hospital classifiers favor medical error recall over precision, docs can be swamped with so many warning messages that they tend to ignore them.

Take adverse drug interactions as an example. The training data for drug interactions mostly come from adults, so the resulting models do not apply in a pediatric setting. When the models are let loose in pediatric hospital, a high percentage of the drug interaction warnings are false positives, so these type of warnings tend to be ignored.

It seems that the trend is to use decision support with a lot of human oversight and investigation of the raw data to see if the model conclusions are correct.

[cmdrfred]

What you describe to me seems to be a simple UI problem. I'd add a button "False Positive, Don't show this warning again" and flag the "drug interaction" for review. If I get enough of these, I'd change the behavior for all users in the next update.