Table 4.
Results for the leave-one-subject-out (LOSO) evaluation of the context classifier using different approaches.
| Data Representation | # of Features | Classifier | Sensitivity | Specificity | AUC |
|---|---|---|---|---|---|
| HAR [77] (baseline) | 290 | AdaBoost | 88.5% (5.56) | 88.6% (5.58) | 0.953 (0.037) |
| FFT | 64 | AdaBoost | 88.6% (4.88) | 88.6% (4.89) | 0.954 (0.029) |
| FFT | 64 | Gradient Boost | 89.8% (4.82) | 89.8% (4.82) | 0.961 (0.029) |
| Raw signal | 384 | CNN (with GAP) | 89.8% (5.00) | 89.8% (5.01) | 0.962 (0.029) |
| Raw signal | 384 | multitask CNN | 92.9% (3.42) 1 | 93.0% (3.58) 1 | 0.981 (0.017) 1 |
| FFT | 64 | multitask CNN | 89.0% (4.91) 1 | 89.1% (4.90) 1 | 0.960 (0.025) 1 |
1 Values obtained only from the context classifier output.