Fig. 4: Visual description of sampling method and generation of reference distributions for connectivity differences.
From: Dissecting the heterogeneity of “in the wild” stress from multimodal sensor data
Multivariate time series features (multicolors: temperature, HR, HRV, etc…) change over time and a binary stress label is shown underneath (red: stress, blue: non-stress). Bootstrap sampling is used to generate iid observations from the time series data, samples are taken from Stress and Non-stress as well as at random (agnostic to stress label), in order to compare stress to a random sampling. Causal graphs representing the causal structure underlying the observations (blue circles: Oura Ring features, purple circles: Survey features) are then generated across the 100 bootstrap iterations, which are then compared using a graph similarity measure. Graph similarity refers to a series of metrics we used to compare graphs during Stress and Non-stress (see Methods).
