Where did the labels come from?
Tonight · ~20 min · read · energy: low · setup: none
A model is only as honest as its reference standard (its “ground truth”). This lesson is the reading question that catches the most insidious bias: a label that is circular, noisy, or not independent of the input. Get the reference standard wrong and every metric downstream is compromised.
The anchor: the label is a measurement, not a truth
“Ground truth” is a comforting phrase that hides a measurement process. Labels come from somewhere — a radiologist’s read, a pathology result, an NLP model on a report, a consensus vote — and each source carries bias and noise. The reading question is not “is the label correct?” but “where did it come from, and is it independent of the input the model sees?”
The three properties to check
- Independence from the input. A label derived from the same image the model reads is circular — the model can appear to predict the label simply by learning whatever feature the labeller used. A label derived from an independent source (pathology, a later adjudicated read, a different acquisition) is far stronger.
- Reference-standard quality. A single reader’s read is noisier than a consensus of experts; an NLP-derived label inherits the report’s inaccuracies; a retrospective label inherits whatever biases the original reporters had. Ask whether the label is the best available reference or merely the available one.
- Reader consensus and variability. How many readers, with what expertise, using what adjudication rule? Inter-reader variability (Ch. 6) contextualises what a score against this reference actually means.
The circularity trap, concretely
Suppose a model predicts “meniscal tear” on knee MRI, labelled from radiology reports. The reports were written by radiologists looking at the same MRI. If the model learns the textual/radiologist cue rather than the tear itself, it scores well against the report-derived label while measuring “agree with the radiologist”, not “detect the tear”. This is exactly the reference-standard issue the Tran et al. dissection raises (lesson 8): labels from reports via an NLP model, not from arthroscopy — flagged by the authors as a limitation, and the kind of thing a reader must carry into the claim. The stronger reference standard (arthroscopy, or adjudicated expert consensus) is harder to get but changes what the AUROC means.
Verification bias
A related trap: if only patients with a positive screening test (or a positive model) get the definitive reference test, the labelled subset is selected by the very process under evaluation — verification bias. The measured performance then applies to a non-representative subset. Ask whether everyone in the cohort received the same reference standard, or only a selected subset.
Stop and think — then reveal
A model predicts “pneumonia on chest X-ray”, labelled from radiology reports. It achieves AUROC 0.96. A clinician points out the reports themselves were the radiologists’ impression of the same X-ray. What does the 0.96 actually measure, and what reference standard would make it a stronger claim?
The 0.96 measures “agreement with the radiologists’ report-derived impression”, not “presence of pneumonia” — the label is partly circular because it came from readers looking at the same image. A model that mimics the radiologist’s cue (a text marker, a device watermark, a pattern the reporter favoured) scores well without detecting the disease. A stronger reference standard would be an independent one — CT confirmation, microbiology, or adjudicated expert consensus with defined criteria — applied uniformly (no verification bias). With a report-derived label, read the 0.96 as “reproduces the report”, which is a narrower and humbler claim than “detects pneumonia”.
What to retain
- The label is a measurement, not a truth; ask where it came from.
- Check independence from the input (circularity), reference-standard quality, and reader consensus/variability.
- A label derived from the same image the model reads is circular — the model may learn the labeller, not the disease.
- Watch verification bias: if only a selected subset gets the definitive reference, the performance applies to a non-representative sample.
Next: even with an honest label — did the test set really leave the development world?