EHR Data Is Not a Dataset

Electronic health records were built to get paid, not to do science. The origin is visible in every column: diagnoses recorded are the ones that justify billing, structured fields exist where reimbursement requires them, and the richest clinical reasoning lives in unstructured notes written for other clinicians — full of shorthand ('c/o', 's/p'), templated bloat ('ROS negative' copied forward), and copy-paste drift that makes today's note partly yesterday's.

Model consequence: your features encode documentation habits and billing incentives alongside physiology. 'No recorded allergy' might mean no allergy — or a field nobody filled.