Dataset

Cohort size: 1200 cases Label noise: 1.0×

Training

Learning rate: 0.30

Decision threshold

Threshold: 0.50

Drag the yellow marker directly on the ROC curve, or use this slider. Both control the same decision threshold.

Live stats (test set)

Epoch0

Train log-loss—

AUC—

TPR (sensitivity)—

FPR—

Precision—

Accuracy—

The model is a real logistic regression: weights train via batch gradient descent on true binary cross-entropy loss over standardised age / length-of-stay / prior-admissions / comorbidity-count features. The ROC curve sweeps 201 real thresholds over the held-out test set and AUC is the trapezoidal-rule integral of those points.