Table 6.
COVID-19 cough classification performance. For the Coswara, Sarcos and ComParE datasets the highest AUCs of 0.982, 0.961 and 0.944 respectively were achieved by a Resnet50 trained by transfer learning in the first two cases and a KNN classifier using 12 primary features determined by sequential forward selection (SFS) in the third. When Sarcos is used exclusively as a validation set for a classifier trained on the Coswara data, an AUC of 0.954 is achieved.
| Dataset | ID | Classifier | Best Feature Hyperparameters | Best Classifier Hyperparameters (Optimised inside nested cross-validation) | Performance |
||||
|---|---|---|---|---|---|---|---|---|---|
| Spec | Sens | Acc | AUC | σAUC | |||||
| Coswara | C1 | Resnet50 + TL | Table 4 | Default Resnet50 (Table 1 in Ref. [39]) | 97% | 98% | 97% | 0.982 | 2 × 10−3 |
| C2 | CNN + TL | ” | Table 4 | 92% | 98% | 95% | 0.972 | 3 × 10−3 | |
| C3 | LSTM + TL | ” | ” | 93% | 95% | 94% | 0.964 | 3 × 10−3 | |
| C4 | MLP + BNF | ” | α3 = 0.35, α7 = 50 | 92% | 96% | 94% | 0.963 | 4 × 10−3 | |
| C5 | SVM + BNF | ” | α1 = 104, α4 = 101 | 89% | 93% | 91% | 0.942 | 3 × 10−3 | |
| C6 | KNN + BNF | ” | α5 = 20, α6 = 15 | 88% | 90% | 89% | 0.917 | 7 × 10−3 | |
| C7 | LR + BNF | ” | α1 = 10−1, α2 = 0.5, α3 = 0.5 | 84% | 86% | 85% | 0.898 | 8 × 10−3 | |
| C8 | Resnet50 + PF [20] | Table 4 in [20] | Default Resnet50 (Table 1 in Ref. [39]) | 98% | 93% | 95% | 0.976 | 18 × 10−3 | |
| C9 | CNN + PF [20] | ” | Table 4 in [20] | 99% | 90% | 95% | 0.953 | 39 × 10−3 | |
| C10 | LSTM + PF [20] | ” | ” | 97% | 91% | 94% | 0.942 | 43 × 10−3 | |
| Sarcos | C11 | Resnet50 + TL | Table 4 | Default Resnet50 (Table 1 in Ref. [39]) | 92% | 96% | 94% | 0.961 | 3 × 10−3 |
| C12 | LSTM + TL | ” | Table 4 | 92% | 92% | 92% | 0.943 | 3 × 10−3 | |
| C13 | CNN + TL | ” | ” | 89% | 91% | 90% | 0.917 | 4 × 10−3 | |
| C14 | MLP + BNF | ” | α3 = 0.75, α7 = 70 | 88% | 90% | 89% | 0.913 | 7 × 10−3 | |
| C15 | SVM + BNF | ” | α1 = 10−2, α4 = 104 | 88% | 89% | 89% | 0.904 | 6 × 10−3 | |
| C16 | KNN + BNF | ” | α5 = 40, α6 = 20 | 85% | 87% | 86% | 0.883 | 8 × 10−3 | |
| C17 | LR + BNF | ” | α1 = 10−3, α2 = 0.4, α3 = 0.6 | 83% | 86% | 85% | 0.867 | 9 × 10−3 | |
| Sarcos (val only) | C18 | Resnet50 + TL | ” | Default Resnet50 (Table 1 in Ref. [39]) | 92% | 96% | 94% | 0.954 | – |
| C19 | LSTM + PF [20] | Table 5 in [20] | Table 5 in [20] | 73% | 75% | 74% | 0.779 | – | |
| C20 | LSTM + PF + SFS [20] | ” | ” | 96% | 91% | 93% | 0.938 | – | |
| ComParE | C21 | Resnet50 + TL | Table 4 | Default Resnet50 (Table 1 in Ref. [39]) | 89% | 93% | 91% | 0.934 | 4 × 10−3 |
| C22 | LSTM + TL | ” | Table 4 | 88% | 92% | 90% | 0.916 | 4 × 10−3 | |
| C23 | CNN + TL | ” | ” | 86% | 90% | 88% | 0.898 | 4 × 10−3 | |
| C24 | MLP + BNF | ” | α3 = 0.25, α7 = 20 | 85% | 90% | 88% | 0.912 | 5 × 10−3 | |
| C25 | SVM + BNF | ” | α1 = 10−3, α4 = 102 | 85% | 90% | 88% | 0.903 | 6 × 10−3 | |
| C26 | KNN + BNF | ” | α5 = 70, α6 = 20 | 85% | 86% | 86% | 0.882 | 8 × 10−3 | |
| C27 | LR + BNF | ” | α1 = 104, α2 = 0.3, α3 = 0.7 | 84% | 86% | 85% | 0.863 | 8 × 10−3 | |
| C28 | KNN + PF + SFS | α5 = 60,α6 = 25 | 84% | 90% | 92% | 0.944 | 9 × 10−3 | ||
| C29 | KNN + PF | α5 = 60, α6 = 25 | 78% | 80% | 80% | 0.855 | 13 × 10−3 | ||
| C30 | MLP + PF | α3 = 0.65, α7 = 40 | 76% | 80% | 78% | 0.839 | 14 × 10−3 | ||
| C31 | SVM + PF | α1 = 10−4, α4 = 10−1 | 75% | 78% | 77% | 0.814 | 12 × 10−3 | ||
| C32 | LR + PF | α1 = 10−2, α2 = 0.6, α3 = 0.4 | 69% | 73% | 71% | 0.789 | 13 × 10−3 | ||