Table 4.
Confusion matrices obtained after majority voting (MV). Results are reported for two different strategies: (1) solving the classification problem using HMM based information only; (2) solving the problem using a SVM classifier. In the latter case, the contributions of HMM-based features only (2A), of time and frequency domain features only (2B) and of the full features set (2C) are evaluated separately. Each entry in the matrix corresponds to a subject. Post stroke (PS) subjects were reported twice, separating the two sides contributions. Correct classifications are in bold.
| Classification Output | ||||
|---|---|---|---|---|
| EL | PS | HD | ||
| 1. HMM-based information (maximum log-likelihood) | ||||
| Actual Label | EL | 7 (70%) | 0 (0%) | 3 (30%) |
| PS–not imp. side | 0 (0%) | 5 (33.3%) | 10 (66.7%) | |
| PS–imp. side | 0 (0%) | 10 (66.7%) | 5 (33.3%) | |
| HD | 1 (5.9%) | 0 (0%) | 16 (94.1%) | |
| Overall accuracy 76.2% of subjects | ||||
| 2A. SVM classifier (HMM-based features only) | ||||
| Actual Label | EL | 9 (90%) | 0 (0%) | 1 (10%) |
| PS–not imp. side | 1 (6.7%) | 13 (86.7%) | 1 (6.7%) | |
| PS–imp. side | 0 (0%) | 15 (100%) | 0 (0%) | |
| HD | 1 (5.9%) | 4 (23.5%) | 12 (70.6%) | |
| Overall accuracy 85.7% of subjects | ||||
| 2B. SVM classifier (time and frequency domain features only) | ||||
| Actual Label | EL | 8 (80%) | 0 (0%) | 2 (20%) |
| PS–not imp. side | 0 (0%) | 13 (86.7%) | 2 (13.3%) | |
| PS–imp. side | 0 (0%) | 14 (93.3%) | 1 (6.7%) | |
| HD | 3 (17.6%) | 0 (0%) | 14 (82.4%) | |
| Overall accuracy 83.3% of subjects | ||||
| 2C. SVM classifier (all available features) | ||||
| Actual label | EL | 9 (90%) | 1 (10%) | 0 (0%) |
| PS–not imp. side | 0 (0%) | 13 (86.7%) | 2 (13.3%) | |
| PS–imp. side | 0 (0%) | 15 (100%) | 0 (0%) | |
| HD | 0 (0%) | 2 (11.8%) | 15 (88.2%) | |
| Overall accuracy 90.5% of subjects | ||||