一种基于集成学习的OSA鼾声监测与分级方法

A Method for OSA Snore Monitoring and Grading Based on Ensemble Learning

摘要: 针对阻塞性睡眠呼吸暂停(Obstructive Sleep Apnea, OSA)居家筛查需求，设计分区气囊枕鼾声监测系统，提出基于Stacking集成学习的鼾声分类方法。系统集成嵌入式鼾声检测模块，实现鼾声信号采集。针对采集的鼾声信号提取梅尔频率倒谱系数(Mel Frequency Cepstral Coefficients, MFCC)特征，构建融合卷积神经网络(Convolutional Neural Network, CNN)、双向门控循环单元(Bidirectional Gated Recurrent Unit, BiGRU)及极端梯度提升(Extreme Gradient Boosting, XGBoost)的异构Stacking模型，并采用逻辑回归进行二层融合分类。实验结果表明，所提方法在OSA鼾声四分类任务中取得95.91%的准确率和95.87%的加权平均F1值，整体性能优于单一模型，可为OSA家庭化智能筛查提供技术支持。

Abstract: To address the need for home-based screening of Obstructive Sleep Apnea (OSA), this paper designs a snore monitoring system based on a partitioned airbag pillow and proposes a snore classification method using Stacking ensemble learning. The system integrates an embedded snore detection module for snore signal acquisition. Mel Frequency Cepstral Coefficients (MFCC) are extracted from the collected snore signals, and a heterogeneous Stacking model is constructed by fusing Convolutional Neural Network (CNN), Bidirectional Gated Recurrent Unit (BiGRU), and Extreme Gradient Boosting (XGBoost), with Logistic Regression employed for second-level fusion classification. Experimental results demonstrate that the proposed method achieves 95.91% accuracy and 95.87% weighted average F1-score on the four-class OSA snore classification task, outperforming individual models overall, and thus can provide technical support for intelligent home-based OSA screening.