AI Driven Snore Source Localization
Researchers have developed a heterogeneous deep learning framework that can classify the anatomical origin of snoring by analyzing audio signals alone. Snoring is a hallmark of obstructive sleep apnea, where vibration or obstruction occurs in structures such as the soft palate, tongue base, epiglottis, or lateral oropharyngeal walls. Currently, identifying the specific site of obstruction often requires invasive procedures. The proposed model uses short-time Fourier transform (STFT) to convert snore sounds into spectrograms, preserving both time and frequency information. A pretrained convolutional neural network (VGG19 or AlexNet) then extracts high-level features, which are classified by a support vector machine (SVM) into four categories corresponding to different upper airway structures.
Performance and Clinical Implications
Tested on the Munich-Passau Snore Sound Corpus (MPSSC), the pipeline achieved an unweighted average recall of 67.1% on the test set, outperforming conventional methods such as MFCC+SVM, end to end CNNs, and advanced audio models like wav2vec 2.0 and audio spectrogram transformers. Ablation studies showed that removing any module (STFT, pretrained CNN, or SVM) reduced performance by 7.5 to 21.3 percentage points, confirming the complementary role of each component. While the results are promising, the model requires independent validation on external clinical datasets before it can be applied in routine sleep medicine. If confirmed, this AI approach could enable noninvasive, low cost triage of snoring patients and guide personalized treatments such as targeted surgery or continuous positive airway pressure therapy.
Source: News-Medical
