Table 7.

COVID-19 breath classifier performance: For breaths, the best performance was achieved by an SVM using bottleneck features (AUC = 0.942). The Resnet50 classifier trained by transfer learning achieves a similar AUC of 0.934.

Dataset	ID	Classifier	Best Feature Hyperparameters	Best Classifier Hyperparameters (Optimised inside nested cross-validation)	Performance
					Spec	Sens	Acc	AUC	σAUC
Coswara	B1	Resnet50 + TL	Table 4	Default Resnet50 (Table 1 in Ref. [39])	87%	93%	90%	0.934	3 × 10−3
	B2	LSTM + TL	”	Table 4	86%	90%	88%	0.927	3 × 10−3
	B3	CNN + TL	”	”	85%	89%	87%	0.914	3 × 10−3
	B4	SVM + BNF	”	α1 = 102,α4 = 10−2	88%	94%	91%	0.942	4 × 10−3
	B5	MLP + BNF	”	α3 = 0.45, α7 = 50	87%	93%	90%	0.923	6 × 10−3
	B6	KNN + BNF	”	α5 = 70, α6 = 10	87%	93%	90%	0.922	9 × 10−3
	B7	LR + BNF	”	α1 = 10−4, α2 = 0.8, α3 = 0.2	86%	90%	88%	0.891	8 × 10−3
	B8	Resnet50 + PF	$\mathcal{M}=39,\mathcal{F}=2^{10},\mathcal{S}=150$	Default Resnet50 (Table 1 in Ref. [39])	92%	90%	91%	0.923	34 × 10−3
	B9	LSTM + PF	$\mathcal{M}=26,\mathcal{F}=2^{11},\mathcal{S}=120$	β3 = 0.1, β4 = 32, β5 = 128, β6 = 0.001, β7 = 256, β8 = 170	90%	86%	88%	0.917	41 × 10−3
	B10	CNN + PF	$\mathcal{M}=52,\mathcal{F}=2^{10},\mathcal{S}=100$	β1 = 48, β2 = 2, β3 = 0.3, β4 = 32, β7 = 256, β8 = 210	87%	85%	86%	0.898	42 × 10−3