Assessing Predictive Ability of Dynamic Time Warping Functional Connectivity for ASD Classification

Figures

• 

  Figure 1. (a) Two plotted time series of a single BOLD signal; one plotted with a lag of 2 time points. The two time series have a PC of r = 0.49. (b) Two plotted time series of a single BOLD signal; one plotted with a lag of 3 time points. The two time series have a PC of r = 0.13.

• 

  Figure 2. Time series are extracted from regions of interests (ROIs), and functional connectivity is calculated (using Pearson's correlation or DTW) between each ROI pair resulting in an ROI × ROI functional connectivity matrix for each participant. The matrices are triangulated, and pairwise ROI distance/similarity measures are used for machine learning classification.

• 

  Figure 3. Visual demonstration of a nested CV algorithm with 5 outer folds and 5 inner folds. Notice that the fifth CV fold's training set (in green) is used to run an inner iteration of fivefold CV (the blue and yellow folds).

• 

  Figure 4. Box plots of CV-based accuracies comparing various models with the DTW-measured fcMRI data.

• 

  Figure 5. Box plots of CV-based accuracies comparing various models with the PC-measured fcMRI data.

• 

  Figure 6. Comparison of ASD prediction accuracies over 100 iterations between DTW-measured versus PC-measured fcMRI data.

• 

  Figure 7. Comparison of ASD prediction accuracies over 100 iterations between the top 30 PCA components of DTW-measured versus PC-measured fcMRI data.

• 

  Figure 8. The calculated global cost matrix between two example time series, s1 = (1, 2, 3, 3, 4, 1) and s2 = (1, 1, 3, 4, 3, 1). The optimal matches between time points are highlighted in gray, and the resulting DTW distance between s1 and s2 is highlighted in black.