Table 2.
Performance of the two proposed algorithms for quality control of clinimetric data and comparison with the randomised classifier. For the naive Bayes and the randomised classifiers, quality control predictions over all time points are evaluated using 10-fold cross-validation. We report the mean and standard deviation (in the brackets) of the balanced accuracy (BA), true positive (TP) and true negative (TN) rates across the different cross-validation trials changing the subsets of data used for training and testing. For the GMM-based approach, we report the BA, TP and TN rates using all the data for training and for testing since this approach is completely unsupervised; the standard deviation is not meaningful for a single trial (hence, standard deviations are omitted).
| Walking Tests | Balance Tests | Voice Tests | |
|---|---|---|---|
| Nonparametric switching AR + naive Bayes | |||
| BA | 85% (11%) | 81% (14%) | 89% (8%) |
| TP | 85% (18%) | 81% (16%) | 88% (9%) |
| TN | 90% (8%) | 88% (9%) | 91% (9%) |
| GMM + running median filtering | |||
| BA | 62% | 24% | 99% |
| TP | 80% | 74% | 86% |
| TN | 89% | 82% | 96% |
| randomised classifier | |||
| BA | 50% (1%) | 50% (0.2%) | 53% (24%) |
| TP | 1% (0.4%) | 0.4% (0.02%) | 99% (1%) |
| TN | 100% | 100% | 6% (23%) |