⚠️ This tool is an educational illustration of how CheXpert's published uncertainty-handling policies affect reported labels and a stand-in performance metric. It does not run a real trained model and is not medical advice or a diagnostic device.
CheXpert (Irvin et al., Stanford, 2019) is a large chest radiograph dataset whose labels for 14 findings were extracted automatically from free-text radiology reports. Because radiologist language is often hedged ("cannot exclude", "possible", "likely represents"), many labels come out as genuinely uncertain rather than a clean positive or negative. The original CheXpert study compared several ways of handling that uncertain class during training and evaluation, most notably treating uncertain labels as positive (U-Ones), as negative (U-Zeros), or simply dropping them (U-Ignore), and found the best policy differs by finding.
Each slider sets a simulated raw model confidence score, from 0 to 1, for one of five CheXpert findings. Scores between 0.40 and 0.60 sit in a simulated "uncertain band" representing genuinely ambiguous cases. Outside that band the classification is fixed at positive (≥ 0.5) or negative. Inside the band, the selected policy decides the outcome: U-Ones always calls it positive, U-Zeros always calls it negative, and U-Ignore excludes it from the reported bar chart and metric entirely. An illustrative AUC-style score, built from fixed per-finding sensitivities to each policy, then shifts to show why label-handling strategy materially changes reported performance, not just reported labels.
Policy selector (U-Ones / U-Zeros / U-Ignore) and five confidence sliders, one per finding: Cardiomegaly, Edema, Consolidation, Atelectasis, Pleural Effusion.
Did you know that in the original CheXpert study, no single uncertainty policy won for every finding? U-Ones tended to help some findings and hurt others, which is exactly why the paper reports results under multiple policies side by side rather than picking one "best" approach for the whole dataset.
This simulator illustrates how the CheXpert chest X-ray dataset's three uncertainty-label policies, U-Ones, U-Zeros and U-Ignore, change both individual finding labels and an illustrative model-performance score whenever a finding's confidence sits in the ambiguous 0.40–0.60 band.
This simulator illustrates how the CheXpert chest X-ray dataset's three uncertainty-label policies, U-Ones, U-Zeros and U-Ignore, change both individual finding labels and an illustrative model-performance score whenever a finding's confidence sits in the ambiguous 0.40–0.60 band.
Drag each of the five finding sliders to set a simulated raw confidence score, then switch the policy selector between U-Ones, U-Zeros and U-Ignore and watch the bar chart relabel any finding sitting inside the shaded uncertain band, while the illustrative AUC score and its delta from baseline update to show the downstream effect on reported performance.
Did you know that in the original CheXpert study, no single uncertainty policy won for every finding? U-Ones tended to help some findings and hurt others, which is exactly why the paper reports results under multiple policies side by side rather than picking one "best" approach for the whole dataset.