Establishment And Application Of A Digital Diagnostic System Based On Pulse Recognition By Machine Learning

Traditional Chinese medicine has a history of more than two thousand years.The pulse diagnosis has played a very important role in the syndrome differentiation system of TCM.However,in the clinical practice of pulse diagnosis,the pulse is mainly acquired by the doctor’s finger to sense the pulsation of the radial artery.There are many influencing factors,and there is no objective evaluation standard.The accuracy of the diagnosis is closely related to the doctor’s clinical experience.From the perspective of the development of traditional Chinese medicine,it is of great significance to construct a digital pulse diagnosis system through modern instruments and techniques,and is also an important way to promote the modernization of Chinese medicine.According to the theory of traditional Chinese medicine,there are obvious differences in the pulse patterns between pregnant and non-pregnant women.Pulse diagnosis is an important means of early diagnosis of pregnancy.Therefore,we use pregnancy detection as an application scenario,combined with signal processing and machine learning techniques to construct digital pulse diagnosis system.In this study,the integrated pulse signal acquisition,feature extraction and pattern recognition are integrated.Pulse signals from 495 women were measured and collected by a modern digital pulse waveform device.This device converted the pulse waveform into a one-dimensional time domain signal.Both time domain and wavelet domain features were extracted for further waveform analysis.These features were sent to different supervised learning models for pattern recognition,including extreme gradient boost,support vector machines,and probabilistic neural networks.Grid search and bayesian optimization were used for hyperparametric optimization during model training.We use the leave-one-out cross-validation to rigorously evaluate the performance of the recognition model.The results show that the Bayesian optimization based on the time&frequency domain features of the XGBoost model has the best detection accuracy.The rate reached 87.5%and the area under the receiver operating characteristic curve was 0.94.We found that the recognition rate of pregnancy was related to the pregnancy stage.The recognition rate in the second trimester was 91.3%,and that in the late pregnancy was 90.7%,which was significantly higher than 85.4%in the early pregnancy.We also compared the difference between the leave-one-out cross-validation and the traditional method to separate the test sets in terms of evaluation stability,the former has better stability.The digital diagnosis and diagnosis system constructed by this research can objectively collect and analyze traditional Chinese medicine pulse,and has the advantages of non-radiation,non-invasion and precision.With the continuous accumulation of data and system optimization,the digital pulse diagnosis system will bring great potential to the social and economic value of Chinese medicine.