Relevance data for: Cross-Domain Transfer Learning Strategy Enhances Interpretability of Deep Learning Model Explanations

This dataset contains pre-computed relevance data extracted from deep learning models trained for atrial fibrillation (AF) detection in 10-second single-lead Einthoven lead II ECGs. Relevance values were computed using Deep Taylor Decomposition (DTD). The data supports the findings of the study "Cross-Domain Transfer Learning Strategy Enhances Interpretability of Deep Learning Model Explanations", which investigates the effect of an inductive transfer learning pipeline on the interpretability of the xECGArch deep learning architecture. xECGArch consists of two branches: a short-term model (STM) processing morphological ECG features, and a long-term model (LTM) processing rhythmic features. 

The dataset contains the ECG test subset used in the paper in af_testset.json, 4 json and 12 MATLAB .mat files, three per model (MSTN, MLTN, RSTN, RLTN), storing the following pre-computed variables:

rhythmAnalysis_*.mat: per-ECG relevance statistics aggregated over clinically defined beat intervals, model predictions, ground truth labels, and signal identifiers

results_*.mat: ANOVA and Tukey-Kramer post-hoc comparison results, effect sizes (Cohen's d), and boxplot summary statistics

templateVals_*.mat: average beat template relevance and standard deviation per model configuration and class

*_norm_relevance.json: Relevance values obatined by DTD for each ECG in the test subset.