An efficient method for snore/nonsnore classification of sleep sounds

A new method to detect snoring episodes in sleep sound recordings is proposed. Sleep sound segments (i.e., 'sound episodes' or simply 'episodes') are classified as snores and nonsnores according to their subband energy distributions. The similarity of inter- and intra-individual spectral energy distributions motivated the representation of the feature vectors in a lower dimensional space. Episodes have been efficiently represented in two dimensions using principal component analysis, and classified as snores or nonsnores. The sound recordings were obtained from individuals who are suspected of OSAS pathology while they were connected to the polysomnography in Gülhane Military Medical Academy Sleep Studies Laboratory (GMMA-SSL), Ankara, Turkey. The data from 30 subjects (18 simple snorers and 12 OSA patients) with different apnoea/hypopnea indices were classified using the proposed algorithm. The system was tested by using the manual annotations of an ENT specialist as a reference. The accuracy for simple snorers was found to be 97.3% when the system was trained using only simple snorers' data. It drops to 90.2% when the training data contain both simple snorers' and OSA patients' data. (Both of these results were obtained by using training and testing sets of different individuals.) In the case of snore episode detection with OSA patients the accuracy is 86.8%. All these results can be considered as acceptable values to use the system for clinical purposes including the diagnosis and treatment of OSAS. The method proposed here has been used to develop a tool for the ENT clinic of GMMA-SSL that provides information for objective evaluation of sleep sounds.