Snoring Sound Classification in Patients with Cerebrovascular Stenosis Based on an Improved ConvNeXt Model

Caijian Hua¹Zhihui Liu¹Liuying Li^2*Xia Zhou^2*Caorong Xiang¹

• ¹Sichuan University of Science and Engineering, Zigong, China
• ²Zigong First People's Hospital, Zigong, China

Introduction: Snoring is a common symptom of Obstructive Sleep Apnea (OSA) and has also been associated with an elevated risk of cerebrovascular disease. However, existing snoring detection studies predominantly focus on individuals with Obstructive Sleep Apnea-Hypopnea Syndrome (OSAHS), with limited attention given to the specific acoustic characteristics of patients with concomitant cerebrovascular diseases. To address this gap, this paper proposes a snoring classification method integrating dynamic convolution and attention mechanisms, and explores the acoustic feature differences between patients with cerebrovascular stenosis and those without stenosis. Methods: First, we collected nocturnal snoring sounds from 31 patients diagnosed with OSAHS, including 16 patients with cerebrovascular stenosis, and extracted four types of acoustic features: Mel spectrogram, Mel-frequency cepstral coefficients (MFCCs), Constant Q Transform (CQT) spectrogram, and Chroma Energy Normalized Statistics (CENS). Then, based on the ConvNeXt backbone, we enhanced the network by incorporating the Alterable Kernel Convolution (AKConv) module, the Convolutional Block Attention Module (CBAM), and the Conv2Former module. We conducted experiments on snoring versus non-snoring classification and stenotic versus non-stenotic snoring classification, and validated the role of each module through ablation studies. Finally, the Mann-Whitney U test was applied to compare intergroup differences in low-frequency energy ratio, snoring frequency, and snoring event duration. Results: This method achieves the best performance on the Mel spectrogram, with a snoring classification accuracy of 90.24%, compared to 88.16% for the ConvNeXt baseline model. It also maintains superiority in classifying stenotic versus non-stenotic snoring. Ablation analysis indicates that all three modules contribute to performance improvements. Moreover, the Mann–Whitney U test revealed significant differences (p < 0.05) between the stenotic and non-stenotic groups in terms of low-frequency energy ratio and nocturnal snoring frequency, whereas snoring event duration showed no significant difference. Discussion: The proposed method demonstrates excellent performance in snoring classification and provides preliminary evidence for exploring acoustic features associated with cerebrovascular stenosis.