Table 2. Performance across datasets.

Average ROC-AUC score, sensitivity, and specificity with SD across five folds using different training data (coughs from all devices versus coughs from smartphone) and test sets (T1, T2, and T3). Two variations of the classifier are tested on three different test sets: T1: subject balanced passive cough dataset (for gender and number of subjects) and used for fivefold training and testing of the classifier; T2: expanded T1 consisting of all non-TB subjects and TB cough data not included for training the fivefold classifier; T3: a voluntary cough dataset consisting of coughs from TB and non-TB subjects. Table S2 represents aggregated result from multiple folds using bootstrapping.

	Model training parameters	Test set	ROC-AUC score (average of 5 folds ± SD across folds)	Sensitivity (average of 5 folds ± SD across folds, threshold = 0.5)	Specificity (average of 5 folds ± SD across folds, threshold = 0.5)	Sensitivity at 70% specificity (average of 5 folds ± SD across folds)	Combined ROC- AUC score of 5 folds (average after combining results from all five folds (DeLong’s CI)
TBscreen	Device: All Scalo gram: 10 Hz to 4 kHz Sampling rate: 44.1 kHz	T1: Subject balanced CV	0.79 ± 0.06	0.70 ± 0.11	0.71 ± 0.10	0.72 ± 0.10	0.80 (0.79–0.80)
		T2: Expanded T1, unbalanced set	0.82 ± 0.03	0.74 ± 0.02	0.72 ± 0.10	0.76 ± 0.04	0.80 (0.79–0.80)
		T3: Voluntary cough, unbal anced set	0.64 ± 0.05	0.34 ± 0.13	0.81 ± 0.12	0.47 ± 0.06	0.64 (0.62–0.66)
TBscreen trained/ evaluated on coughs from smartphone	Device: Smart phone Scalogram: 10 Hz to 4 kHz Sampling rate: 44.1 kHz	T1 subset: Sub ject balanced CV	0.83 ± 0.11	0.76 ± 0.12	0.74 ± 0.10	0.76 ± 0.20	0.85 (0.84–0.85)
		T2 subset: Expanded T1, unbalanced set	0.86 ± 0.03	0.80 ± 0.03	0.74 ± 0.10	0.83 ± 0.05	0.86 (0.85–0.87)
		T3 subset: Voluntary cough, unbalanced set	0.61 ± 0.14	0.16 ± 0.11	0.95 ± 0.05	0.51 ± 0.18	0.66 (0.62–0.70)