Table 1.
Breakdown of study encounters by BOAS functional grade.
Table 2.
Count of recordings with paired stertor and stridor labels.
Fig 1.
Flow chart showing nested cross validation procedure to train and evaluate models.
Fig 2.
Laryngeal stethoscope recording (a) and corresponding spectrogram (b) for a French bulldog with no audible stertor or stridor. No abnormal sounds are clear in the spectrogram, consistent with the quiet breathing that can be heard in the recording.
Fig 3.
Laryngeal stethoscope recording (a) and corresponding spectrogram (b) for a “BOAS positive” Pug with constant moderate stertor and no stridor. Four distinct stertor sounds are present with fundamental frequencies and harmonics visible in the spectrogram.
Fig 4.
Recurrent neural network attention mechanism weights plotted for a recording (a) with moderate, intermittent nasopharyngeal, stertor. The corresponding spectrogram (b) demonstrates that a combination of time and frequency information is needed to identify the stertor sounds. The RNN in this case is able to correctly use the 2D spectrogram information to address the key stertor sounds and discard silence and noise.
Fig 5.
Receiver operating characteristic curves for predicting significant stertor, with the ground truth given by the trained veterinarian’s label.
An example operating point (OP) that maximises the sum of sensitivity and specificity is shown.
Fig 6.
Receiver operating characteristic curves for prediction of ‘BOAS positive’ cases by averaging stertor predictions for a single patient encounter.
Two operating points (OPs) for the algorithm are shown which prioritise either sensitivity or specificity. Also shown is the performance of the expert stertor annotation at predicting BOAS, which provides an upper-bound for the algorithm accuracy.
Table 3.
Per-class precision and recall results.