Figure 1.
Waveforms of a double-cough from a TB patient (upper) and of an asthma patient (lower).
Figure 2.
Bland-Altman plot comparing the number of epochs (definition epoch1) found by the nurses and the reviewed algorithm (i.e. semi-automated approach).
The mean of the two estimates (used in place of a gold standard) is plotted vs. the difference between nurse and semi-automated results. The mean bias and limits of agreement (+/−1.96 σ) are also shown. The plot shows the bias is not statistically significant and there is no evidence of changing agreement as a function of cough epoch count.
Figure 3.
Bland-Altman plot comparing repeatability of semi-automated results (blue circles) to repeatability of nurse findings (red squares).
In this comparison, the first and second files for each day (‘file1’ and ‘file2’) were compared. Repeatability is similar for nurse assignments and semi-automated algorithm results.
Figure 4.
Scatterplots showing the number of epochs found under ‘epoch1 ’ and ‘epoch2’ definitions, for semi-automated algorithm results.
Note that the ‘epoch2’ definition cannot be applied to nurse assignments in our dataset. The correlation coefficient between the two definitions is 0.97.
Figure 5.
Boxplot comparing semi-automated estimate of cough count at day 0 and day 14.
The plot shows 25th, 50th, and 75th percentiles, with outliers (1.5*IQR) are shown as ‘+’. At Day 14, the box collapses as 25th, 50th, and 75th percentiles are all zero.
Figure 6.
High-level flowchart of cough detection algorithm.
Figure 7.
Example issue with simple energy detector.
The threshold may miss the start of the acoustic event and is frequently crossed during speech events (shown above), increasing the chance of misclassification.
Figure 8.
Pseudo-code for event detection logic.
Figure 9.
Pseudo-code for clustering algorithm.