Table 3 Relative area under the curve (AUC) values for the different perturbation schemes and models. Bold cells mark the highest score per row and italic cells the second-highest. Methods that have not reached first or second rank once were omitted for readability. LRP, layerwise relevance propagation; DTD, deep Taylor decomposition; ITG, input times gradient.

From: xECGArch: a trustworthy deep learning architecture for interpretable ECG analysis considering short-term and long-term features

Scheme

Model

LRP-\(\alpha \beta\)

LRP-Z

LRP-\(\epsilon\)

DTD

GradCAM+

ITG

Random

Interpolation

Short-term

0.248

0.217

0.220

0.546

0.427

0.270

0.129

Long-term

0.852

0.230

0.168

0.817

0.601

0.240

0.087

Zero

Short-term

0.261

0.490

0.459

0.209

0.232

0.529

0.252

Long-term

0.350

0.603

0.660

0.306

0.297

0.712

0.303