Labels, Codes & Leakage

'Ground truth' in clinical ML is usually a billing code, a problem-list entry, or a registry — all proxies for disease. Was 'diabetes' coded when a clinician suspected it, billed for it, or the A1c crossed a threshold? Different definitions select different populations.

And mislabeled positives aren't random: coding concentrates in patients who get tested — so your label inherits your system's screening patterns. A 'confirmed case' often means 'a case someone had a reason to look for.'