Ground truth
The answer a model's output is scored against during development and testing. In medicine it is usually a human judgment of some kind: one radiologist's read, a consensus panel, a pathology result, or what happened to the patient at follow-up. Each is imperfect in its own way, and a model cannot be shown to be more accurate than the truth it was measured against. If experts disagree on a third of cases, a claim of matching expert performance means something different from the same claim on a task where they agree. When a tool is said to match or beat clinicians, the first question is what the truth was and who decided it.
Two mammography tools both report performance equal to radiologists. In the first study the truth was biopsy results plus two years of follow-up, so cancers that surfaced later counted against whoever missed them, tool or reader. In the second the truth was the original screening reports, with no biopsy or follow-up, and a separate panel of radiologists was compared with the tool against those reports. The second design cannot show the tool or the panel finding a cancer the original readers missed, because any such finding is scored as a false positive.
Terms like this come up in real clinical scenarios across the HelloAI courses: bite-sized modules with verifiable certificates. An account takes one minute, no password needed.
Sign in →