基于心率变异性分析的睡眠分期算法研究和验证

Sleep Stage Classification Based on Heart Rate Variability Analysis and Model Performance Validation

为实现睡眠分期,为穿戴式生理参数监测技术在慢病监测领域的应用提供技术支撑,发展基于心率变异性和支持向量机模型的睡眠分期算法。从心率时间间期序列中提取时域、频域和非线性等86个特征,将多导睡眠图仪的三分类结果(醒、快速眼动期、非快速眼动期)作为"金标准",采用支持向量机作为多分类器模型;为保证训练集数据质量,使用开放睡眠数据库SHHS中由专家确认挑选的67例PSG样本作为训练集,实现特征筛选和模型参数训练。为验证模型的泛化性能,从SHHS数据库中进一步随机提取939例PSG样本,对模型性能进行测试。睡眠分期模型在训练集上的五折交叉验证的准确率为84.00%±1.33%,卡帕系数为0.70±0.03;在939例测试集上的准确率为76.10%±10.80%,卡帕系数为0.57±0.15。剔除RR间期异常(110例)和明显睡眠结构异常(29例)的样本后,测试集(800例)的准确率为82.00%±5.60%,卡帕系数为0.67±0.14。所提出的基于心率变异性分析的睡眠分期算法具有较高的准确性,大样本人群测试结果表明,该模型具有较好的普适性。

To conduct sleep stage classification and provide technical support for the application of wearable physiological monitoring technology in the field of chronic disease monitoring and management,a sleep stage classification algorithm based on heart rate variability(HRV)analysis and support vector machine(SVM)was developed.In order to ensure the quality of the training set,67 polysomnography(PSG)records were extracted by experts from the SHHS(Sleep Heart Health Study)PSG database for model training and internal validation.The sleep stages(wake,rapid eye movement and non-rapid eye movement)classified by EEG signals were used as labels to train the SVM model.Totally 86 features were derived from HRV analysis,including time domain,frequency domain and nonlinear domain.To test the generalization of the model,another 939 PSG records were further randomly extracted from the SHHS PSG database for model external validation.The accuracy of the 5-fold cross-validation on the training dataset of the 67 PSG records was 84.00%±1.33%,with a Kappa coefficient:0.70±0.03,and the accuracy of the algorithm on the 939 PSG records was 76.10%±10.8%,with a Kappa coefficient:0.57±0.15.The accuracy and the Kappa coefficient increased to 82.00%±5.6%and 0.67±0.14 when some records were excluded from the test dataset,including 110 records with abnormal RR intervals and 29 records with apparent abnormal sleep structures.These results showed that the model of heart rate variability analysis based sleep stage classification proposed in this paper exhibited a good performance,and the external validation by a dataset with large sample size demonstrated the generalization of the model.