Fig. 3: Visual explanations of SleepXViT on KISS dataset.
From: Explainable vision transformer for automatic visual sleep staging on multimodal PSG signals

Heatmaps emphasize the key regions of the input image that significantly affect the outcomes. The colors represent the relevance of each pixel to the model’s decision. a The average heatmap, created by overlaying 10,000 maps for each class, demonstrates which signal parts influence the decisions for each class across numerous samples. b Representative samples that vividly highlight distinctive features characteristic of each class.