MEG does not have the ability to be trained locally for missed detections. a previous version (IronMail 6.7.2) had a bayesian engine that could be trained, however over time the bayesian engine would cause about as many false positives as true detections and would ultimately need to be reset to the default table by support. due to this happening, it was not implemented in version 7. I would recommend you submit a PER if you would like this feature added. thanks.
Understood, thanks for the quick feedback!