中醫脈搏診斷系統之開發

Abstract

本研究利用ANSWatch儀器收集了10種脈種，其中包含5種單一脈種以及5種複合脈種，並在訊號處理時綜合8種特徵產生方法，其中包括時(time domain)、頻域(frequency domain)、伽瑪密度方程式(Gamma density function)、希爾伯特黃轉換(Hilbert Huang transform)、近似熵(approximate entropy)、多模態樣本熵(multiscale sample entropy)、小波轉換(wavelet transform)及小波包轉換(wavelet packet transform)來分析量測到的手腕脈搏訊號，並搭配主成分分析(principal component analysis)將冗餘的資訊去除，保留含有重要資訊的特徵。最後，使用類神經網路(artificial neural network)將 10 種脈搏進行分類。 在實驗中，利用類神經網絡對五個模型進行訓練: 模型一使用的數據為利用主成分分析取得含有原始特徵資料的70%的累積變異數資料，模型二使用的數據為利用主成分分析取得含有原始特徵資料的80%的累積變異數資料，模型三使用的數據為利用主成分分析取得含有原始特徵資料的90%的累積變異數資料，模型四使用的數據為利用主成分分析取得含有原始特徵資料的95%的累積變異數資料，而模型五使用的數據為沒有經過主成分分析降維的原始高維度特徵資料。實驗結果顯示，模型一可達到92.8%的平均分類準確率，模型二可達到94.6%的平均分類準確率，模型三可達到97.4%的平均分類準確率，模型四可達到98.2%的平均分類準確率，模型五可達到96.6%的平均分類準確率。實驗結果發現，原始的高維度特徵資料可達到不錯的分類結果，然而透過主成份分析篩選重要特徵能夠再提升分類結果。 本研究提出了高維度特徵方法對中醫脈搏進行分類，實驗結果顯示，本研究的方法分類脈搏能力佳，並且避免了多數研究只適用診斷特殊疾病的模式，預計本結果能有效提升中醫脈診之診斷能力。

This research used ANSWatch device to collect 10 kinds of pulse patterns, including 5 single pulse patterns, and 5 complex pulse patterns. In terms of signal processing, eight feature generation methods were used to analyze the recorded wrist pulse patterns, including time domain, frequency domain, Gamma density function, Hilbert Huang transform (HHT), approximate entropy (ApEn), multiscale sample entropy (Multi-SampEn), wavelet transform (WT), and wavelet packet transform (WPT). Then, principal component analysis (PCA) was used to remove redundant information and retain features with important information. Finally, the 10 pulse patterns were classified using artificial neural network (ANN). In the experiment, five models were trained by ANN: model 1 used the data applying PCA with 70% of cumulative variance, model 2 used the data applying PCA with 80% of cumulative variance, model 3 used the data applying PCA with 90% of cumulative variance, model 4 used the data applying PCA with 95% of cumulative variance, and model 5 used the data with original features without applying PCA. The results showed that model 1 reached 92.8% average accuracy of pulse classification, model 2 reached 94.6% average accuracy of pulse classification, model 3 reached 97.4% average accuracy of pulse classification, model 4 reached 98.2% average accuracy of pulse classification, and model 5 reached 96.6% average accuracy of pulse classification. The results illustrated that the original high-dimensional features achieved good classification results, but through PCA could further improve the results of the experiment. This study proposed a high-dimensional features method to classify traditional Chinese medicine (TCM) pulse patterns. The experimental results showed that the method proposed by this research had better pulse classification ability than other prior research. The results of this research are expected to effectively improve the accuracy of Chinese medicine pulse diagnosis (TCPD).