Abstract
We developed a semi-automatic snore-detection algorithm, which only requires minimal human inputs to label a one-night-long snore recording. The algorithm performs hierarchical clustering on ten percent of snore signal segments extracted from the polysomnography (PSG) of a subject and has human users labeling the resulting clusters. The labeled clusters are then used to classify the remaining of the snore signal segments. This method enables fast snore classification by reducing significantly human manual labor. Since the data used to construct the cluster-based classifier are subject specific, it also allows the use of simple features to achieve reasonable accuracy. In our experiment, we use only the energy profile and the peak frequency of the signal segments to achieve an average F1-score of 0.9289, which is on par with the state-of-the-art methods. On average, it reduces the human labor time from 5.4 hours to 22 minutes (93.3 % reduction) to label a whole night snore record. The implication of the algorithm includes enabling wider use of snoring quantification in health studies. Given the speedup, the algorithm can be used to tap into vast retrospective PSG data. The use of hierarchical clustering and simple features make the classification more transparent and hence less resistance by the medical community to use it.