Spatiotemporal activity estimation for multivoxel pattern analysis with rapid event-related designs

Abstract

Despite growing interest in multi-voxel pattern analysis (MVPA) methods for fMRI, a major problem remains--that of generating estimates in rapid event-related (ER) designs, where the BOLD responses of temporally adjacent events will overlap. While this problem has been investigated for methods that reduce each event to a single parameter per voxel (Mumford et al., 2012), most of these methods make strong parametric assumptions about the shape of the hemodynamic response, and require exact knowledge of the temporal profile of the underlying neural activity. A second class of methods uses multiple parameters per event (per voxel) to capture temporal information more faithfully. In addition to enabling a more accurate estimate of ER responses, this allows for the extension of the standard classification paradigm into the temporal domain (e.g., Mourão-Miranda et al., 2007). However, existing methods in this class were developed for use with block and slow ER data, and there has not yet been an exploration of how to adapt such methods to data collected using rapid ER designs. Here, we demonstrate that the use of multiple parameters preserves or improves classification accuracy, while additionally providing information on the evolution of class discrimination. Additionally, we explore an alternative to the method of Mourão-Miranda et al. tailored to use in rapid ER designs that yields equivalent classification accuracies, but is better at unmixing responses to temporally adjacent events. The current work paves the way for wider adoption of spatiotemporal classification analyses, and greater use of MVPA with rapid ER designs.

Figure 1

Schematic representation of the design matrix for a single iteration of each of the three methods, shown with only a single class (and hence only one set of nuisance columns for FS); the parameters of interest are the β_Ns for LS1 and FS, and the βs for MM.

Figure 2

Classification accuracy results for the simulated data for six methods. Signal-to-noise ratio decreases across rows, while ISI increases across columns. The solid horizontal line indicates chance (50%), while the dashed line indicates accuracy significantly better than chance according to the binomial distribution.

Figure 3

Decrease in accuracy caused by using a shifted HRF to generate simulated data. Each line runs from 0 s shift to 2 s shift in steps of 0.5 s. The solid horizontal line indicates chance (50%), while the dashed line indicates accuracy significantly better than chance according to the binomial distribution.

Figure 4

a. Beta weights (±95% confidence intervals) showing the influence of neighboring events for 50 simulated runs with 0–4s ISI and noise standard deviation = 0.8. The ordinate shows change in classifier margin (y-axis of each frame) as a function of the presence of a neighbor at each 2s lag ±18s (x-axis), and whether that neighbor was of the same (blue) or opposite (red) class. b. Average classifier accuracy (± 95% confidence interval) for each method as a function of number of same- or opposite-class neighbors within ±4s, for 50 simulated runs with 0–4s ISI and noise standard deviation = 0.8.

Figure 4

a. Beta weights (±95% confidence intervals) showing the influence of neighboring events for 50 simulated runs with 0–4s ISI and noise standard deviation = 0.8. The ordinate shows change in classifier margin (y-axis of each frame) as a function of the presence of a neighbor at each 2s lag ±18s (x-axis), and whether that neighbor was of the same (blue) or opposite (red) class. b. Average classifier accuracy (± 95% confidence interval) for each method as a function of number of same- or opposite-class neighbors within ±4s, for 50 simulated runs with 0–4s ISI and noise standard deviation = 0.8.

Figure 5

a. Exponentiated logistic regression weights (±95% confidence intervals) showing the influence of neighboring events for real data. The regression predicts event-by-event accuracy of an SVM classifier as a function of the presence of a neighbor at each 2s lag ±18s (x-axis), and whether that neighbor was of the same (blue) or opposite (red) class. The ordinate (log scaled for visualization) gives the change in odds ratio. For example, a value of 2 above zero corresponds to a doubling in p(correct)/p(incorrect) with presence as compared to absence, while a value of 2 below zero corresponds to a doubling in p(incorrect)/p(correct) for presence versus absence. b. Average SVM classifier accuracy (± 95% confidence interval) for each method as a function of number of same- (blue) or opposite- (red) class neighbors within ±4s.

Figure 5

a. Exponentiated logistic regression weights (±95% confidence intervals) showing the influence of neighboring events for real data. The regression predicts event-by-event accuracy of an SVM classifier as a function of the presence of a neighbor at each 2s lag ±18s (x-axis), and whether that neighbor was of the same (blue) or opposite (red) class. The ordinate (log scaled for visualization) gives the change in odds ratio. For example, a value of 2 above zero corresponds to a doubling in p(correct)/p(incorrect) with presence as compared to absence, while a value of 2 below zero corresponds to a doubling in p(incorrect)/p(correct) for presence versus absence. b. Average SVM classifier accuracy (± 95% confidence interval) for each method as a function of number of same- (blue) or opposite- (red) class neighbors within ±4s.