Figure 3.

Tensor component analysis of swim behavior during SC regeneration. (A) Model selection parameters. Plots on the left show model error various ranks (x-axis). Plots on the right show model similarity for various ranks (x-axis). Unlike PCA, TCA is not deterministic. Thus, multiple models were trained and compared. Three model types are shown. To achieve an optimal balance between low model error and high model similarity, we chose a rank-7 canonical polyadic decomposition by alternating least squares (cp_als). (B) Selecting an appropriate distance threshold for clustering. We clustered fish according to their TCA factors using agglomerative clustering. To maximize the number of clusters while keeping a high minimum cluster size, we chose 1.7 as our distance threshold with five clusters identified. (C) Recovery outcomes for the five identified fish clusters, sorted by average glial bridging, low to high, from left to right. Each row’s color is normalized such that the darkest blue corresponds to the maximum observation among all fish for that measurement. (D) Factors from a rank 7 tensor decomposition corresponding to each axis of the full data tensor: fish, assays (wpi), and functional measurements. Fish are sorted on the x-axis by their 8 wpi measured percent glial bridging and colored according to the rank of the cluster with respect to average glial bridging. Assays are sorted temporally. Measurements are listed below the plots in descending order. In this figure, 44 out of 60 total fish were analyzed. Sixteen fish were omitted either due to early death (6 fish) or because cruise waveform statistics were insufficient for at least one assay (10 fish).