Machine Learning Classification to Identify the Stage of Brain-Computer Interface Therapy for Stroke Rehabilitation Using Functional Connectivity

Figure 1. Study paradigm. The time-points at which neuroimaging data were collected are represented by: T1: Control baseline 1, T2: Control baseline 2, T3: Control baseline 3, T4: Therapy baseline T5: Mid-therapy, T6: Post-therapy, and T7: 1-month post-therapy. While the crossover control group completed visits T1 through T7, the BCI therapy group completed visits T4 through T7 only.

Figure 2. The 236 seeds regions involving motor and non-motor regions include 13 major brain networks color coded according to Table 2 and visualized using BrainNet Viewer (Xia et al., 2013). The seed regions falling outside the template of cerebrum were part of the cerebellum.

Figure 3. Methodology for single-participant analysis: (A) raw structural T1 scan (top) was preprocessed and spatially normalized to MNI space (bottom); (B) raw functional scan (top) was preprocessed up to smoothing (bottom); (C) smoothed fMRI was temporally filtered to obtain the low frequency oscillations within the range of 0.01–0.1 Hz using a bandpass filter; (D) 236 seeds comprising of 13 major brain networks were used to extract BOLD time courses at each seed region; (E) 236 × 236 rs-FC matrix was computed using the BOLD time courses; (F) unique pairwise correlations contained in the lower triangle of the rs-FC matrix were extracted and vectorized into a 27,730-dimensional vector.

Figure 4. Methodology for group-level analysis: (A) vectorized form of rs-FC matrix for each participant aggregated for T4, i.e., pre-therapy and T6, i.e., post-therapy time points. Each group had 20 participants with 27,730-dimesional features; (B) outliers (marked in yellow) at pre- and post-therapy were identified using MAD approach; (C) reduced rs-FC matrix after cumulative outliers were removed, i.e., each stage consisted of 20 participants and 17,614 features; (D) 679 features that were significantly different between pre- and post-therapy stages as identified by a paired t-test were retained and data across the two stages were combined together for a feature transformation step; (E) feature transformation using PCA was performed that resulted in data with 40 participants and 39 low-dimensional principal components features. Of them 25 features accounted for more than 85% variance and were used as final features for classification; (F) the selected features were fed to the binary SVM classifier that labels each test participant to either pre-therapy or post-therapy stage using LOOCV.

Figure 5. First three principal components corresponding to pre-therapy rs-FC and post-therapy rs-FC for all participants were visualized. Each point in the 3-D plot corresponds to a participant. There appeared to be an almost clear separation between the two stages just with three principal components. Adding higher number of components better explained the variance in the data. Our analysis used 25 components that explained over 85% of the variance in the dataset.

Figure 6. The number of principal components are arranged in order of importance so that the first component accounts for the largest proportion of variance in the rs-FC data. Of the 39 principal components, 25 were chosen as marked in the graph as they cumulatively explained over 85% of the variance in the data, represented by the shaded area under the curve.

Figure 7. The ROC for the learned SVM classifier was compared to that of a random classifier. The SVM classifier with optimized model parameters showed the best performance. The area under the curves for unoptimized and optimized SVM are specified in Table 4.

Figure 8. Visualization of (A) 441 strengthening functional connections and (B) 238 weakening functional connections. The overall number of connections involved in the motor and non-motor networks can be found in Table 5. A detailed list of individual connections can be found in the Supplementary Tables 1, 2, respectively. All brain visualizations were performed using BrainNet Viewer Toolbox (Xia et al., 2013).

Figure 9. Number of discriminating connections per network is plotted below: (A) shows the distribution of involvement of various networks in discriminating features; (B) shows the involvement of various networks when normalized with respect to the number of seeds found in each network. The two networks primarily associated with motor functions are highlighted.

Figure 10. Involved seed regions were weighted as per their contribution in classification. The size of each seed was directly proportional to assigned weight. The top weighted seeds belonged to fronto-parietal, hand motor, default mode, and visual networks. A detailed list of the networks and labels of ROIs ranked as per their weights are presented in Supplementary Table 3.