Table 3 Performance comparison of ViT models trained with different class ratios for predicting hematoma expansion

From: An end-to-end deep learning pipeline for hematoma expansion prediction in spontaneous intracerebral hemorrhage based on non-contrast computed tomography

Cohorts

Models

AUC (95%CI)

Accuracy (95%CI)

Sensitivity (95%CI)

Specificity (95%CI)

NPV (95%CI)

PPV (95%CI)

Training set

ViT-1:1

0.805 (0.776–0.836)

0.758 (0.733–0.783)

0.656 (0.593–0.719)

0.783 (0.756–0.810)

0.902 (0.881–0.923)

0.427 (0.374–0.480)

 

ViT-1:2

0.815 (0.790–0.841)

0.771 (0.746–0.795)

0.679 (0.617–0.741)

0.793 (0.767–0.820)

0.909 (0.889–0.930)

0.447 (0.394–0.501)

 

ViT-1:4

0.802 (0.771–0.833)

0.770 (0.745–0.795)

0.633 (0.569–0.697)

0.803 (0.777–0.830)

0.899 (0.878–0.920)

0.442 (0.387–0.497)

External validation set1

ViT-1:1

0.752 (0.701–0.806)

0.715 (0.676–0.754)

0.634 (0.545–0.723)

0.738 (0.695–0.781)

0.878 (0.843–0.913)

0.403 (0.331–0.476)

 

ViT-1:2

0.793 (0.750–0.837)

0.747 (0.709–0.784)

0.661 (0.573–0.748)

0.771 (0.729–0.812)

0.890 (0.858–0.923)

0.446 (0.370–0.521)

 

ViT-1:4

0.714 (0.661–0.770)

0.696 (0.656–0.736)

0.625 (0.535–0.715)

0.716 (0.672–0.760)

0.872 (0.836–0.908)

0.380 (0.310–0.451)

External validation set2

ViT-1:1

0.738 (0.681–0.792)

0.691 (0.646–0.736)

0.642 (0.538–0.746)

0.703 (0.653–0.753)

0.887 (0.848–0.926)

0.351 (0.274–0.428)

 

ViT-1:2

0.781 (0.733–0.827)

0.730 (0.687–0.773)

0.630 (0.524–0.735)

0.755 (0.709–0.802)

0.891 (0.854–0.927)

0.392 (0.308–0.476)

 

ViT-1:4

0.728 (0.667–0.783)

0.708 (0.664–0.752)

0.617 (0.511–0.723)

0.731 (0.682–0.779)

0.884 (0.845–0.922)

0.365 (0.284–0.446)

  1. ViT-1:1 trained on the UKAN-balanced dataset (1:1), ViT-1:2 trained on the UKAN-semibalanced dataset (1:2), ViT-1:4 trained on the original real-imbalanced dataset (1:4), AUC area under the curve, CI confidence interval, PPV positive predictive value, NPV negative predictive value.
Back to article page