A hospital will train a custom model in Microsoft Foundry to suggest discharge codes. Clinicians can label a curated note set. The live notes store changes all day and those notes are unlabeled. Which training design should the architect approve?
Select an answer to reveal the explanation.
Short Explanation
A chart still in pencil is not a textbook. You train this discharge-code model only on notes clinicians have labeled, and you keep the live notes store out of training. The live store is the record of today's work, not a stable labeled set.
Full Explanation
A custom model in Microsoft Foundry needs stable labeled data, and discharge codes are only as good as the clinician who marked them. The approved design trains on that curated labeled set and keeps the live notes store out of training, because the store changes all day and its notes have no code label. Training on the live notes store so each new unsigned note becomes a row with no clinician label teaches the model unlabeled production text and does not supply the code a clinician would have chosen. Skipping clinician labels and folding the live notes store into the model weights from raw notes has the same hole: raw notes do not state the discharge code. Using the live notes store as both the training source and the validation source for the same notes inflates the score, because the model is graded on rows it already trained on, and those rows are still unlabeled. Exam caveat: a labeled extract is a copy made for training, not a live connection that keeps learning from today's unsigned notes. Operational check: confirm the training job's input is the clinician-labeled set, and that the live notes store is absent from the training inputs.