Fig. 2

Performance of agonal breathing classifier. a Demographic summary of subjects with agonal breathing during 9-1-1 calls showing distribution of age and gender. b ROC curve for our support vector machine classifier, cross-validated on sounds collected from a sleep study, and domestic interfering sounds. c Confusion matrix of agonal breathing and sleeping/domestic interfering sounds indicating the operating point on the ROC curve. d t-SNE algorithm is applied to visualize the audio embeddings in 2-D. The point clouds show clusters representing the abstract features learned to represent both agonal breathing and negative sound instances. e The false positive rate when running the classifier across an 82-h stream of sleep data without and with the frequency filter. By applying a frequency filter to check if the rate of positive predictions matches the rate of agonal breathing, we can decrease the false positive rate