| 摘要: |
| 目的 为改进基于人工智能技术高血压时域脉搏波分类模型精度低、泛化性能差的问题，提出一种基于融合注意力机制的频域脉搏波预测模型。方法 首先将时域脉搏波转换为频域梅尔频率倒谱系数特征，增强脉搏波区分度，采用时间卷积网络与Transformer 结构提取脉搏波深层特征，并将自注意力机制与选择性内核注意力进行决策融合，提取脉搏波关联特征，并采用Floodings正则化方法间接控制训练损失，防止过拟合发生。针对上海中医药大学附属龙华医院及上海市中西医结合医院提供的527例临床脉诊数据，进行5折交叉验证实验。此外，采用梯度提升决策树算法统计脉搏波频域特征的贡献率排名，分析影响模型分类精度的关键因素，为中医临床辅助诊断提供参考价值。结果 本研究提出的模型分类评估指标准确度、F1值、精确率、召回率和AUC值分别为0.939 6、0.924 9、0.940 9、0.929 5和0.993 4。脉搏波的静态特征、一阶差分和二阶差分系数的贡献率相对均衡，说明高血压危险程度不仅与脉搏波的静态特征相关，也应当考虑脉搏波的动态特征。结论 与典型脉搏波分类模型相比，本研究提出的模型具有较高的分类精度和泛化性能。 |
| 关键词: 高血压 危险分层 梅尔频率倒谱系数 时间卷积网络 Transformer |
| DOI：10.16781/j.CN31-2187/R.20230243 |
| 投稿时间：2023-04-29修订日期：2023-09-04 |
| 基金项目:国家自然科学基金（81973749），中国中医科学院科技创新工程重大攻关项目（CI2021A01503）. |
|
| Hypertension risk stratification prediction model based on frequency-domain pulse wave Mel-scale frequency cepstral coefficient features |
| QI Chenhao1,YANG Jingdong1*,QIU Zehao1,YAO Minghui2,YAN Haixia2 |
(1. Autonomous Robot Lab, School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China;
2. Department of Chinese Medicine Diagnosis, School of Traditional Chinese Medicine, Shanghai University of Traditional Chinese Medicine, Shanghai 201203, China
*Corresponding author) |
| Abstract: |
| Objective To propose a frequency-domain pulse wave prediction model based on fusion attention mechanism, improving the low classification accuracy and poor generalization performance of hypertension time-domain pulse wave classification based on artificial intelligence technology. Methods Firstly, the time-domain pulse wave was transformed into frequency-domain Mel-scale frequency cepstral coefficient features to enhance its discriminability. Then, temporal convolutional network and Transformer structures were employed to extract the deep features of pulse waves, and self-attention mechanism and selective kernel attention were combined for decision fusion to extract relevant features. Floodings regularization method was used to indirectly control the training loss and prevent overfitting. A 5-fold cross-verification experiment was conducted based on 527 clinical pulse diagnosis data provided by Longhua Hospital, Shanghai University of Traditional Chinese Medicine and Shanghai Traditional Chinese Medicine-Integrated Hospital. Additionally, the extreme gradient boosting algorithm was employed to calculate the contribution rate ranking of frequency-domain pulse wave features, and the key factors affecting the classification accuracy of the model were analyzed to provide reference for the clinical auxiliary diagnosis of traditional Chinese medicine. Results The evaluation metrics accuracy, F1 score, precision, recall rate and area under curve value of the model proposed in this study were 0.939 6, 0.924 9, 0.940 9, 0.929 5, and 0.993 4, respectively. The static characteristics of the pulse wave, the contribution rate of the first-order difference and the second-order difference coefficients were relatively balanced, indicating that the degree of hypertension risk was not only related to the static characteristics of the pulse wave, but also to the dynamic characteristics of the pulse wave. Conclusion The proposed model has higher classification accuracy and generalization performance compared to typical pulse wave classification models. |
| Key words: hypertension risk stratification Mel-scale frequency cepstral coefficient temporal convolutional network Transformer |
