Fig. 5: Evaluation of Explainability via Clustering.
From: Explainable vision transformer for automatic visual sleep staging on multimodal PSG signals
Heatmaps are clustered into five groups corresponding to the five sleep stages using K-means clustering with K = 5. a Distribution of predicted classes across clusters and results of cluster pseudo labeling. b Randomly sampled heatmaps from each cluster showing that similar maps cluster together, supporting the model’s consistent explainability. The displayed sleep stage is a pseudo labeled sleep stage. c Normalized confusion matrix comparing the predictions of the Intra-epoch ViT with the clustering outcomes, illustrating the alignment and discrepancies between model predictions and clustering.
