Taking pulse as a disease diagnosis process has a long history in traditional Chinese
medicine (TCM). Ancient physicians used the common attributes of pulse conditions and
finger-feeling characteristics as a basis for pulse classification, which " position, rate,
shape and tendency " is the principle for pulse differentiation. However, it is not easy to
express feelings of hands in a scientific way and not easy for clinical teaching and
practice. The modernization of pulse diagnosis in Taiwan originated in the 1970s. By using
pressure waves of the radial artery, two methods were developed : time-domain analysis and
frequency domain analysis. Dr. Huang used time-domain analysis combined with frequency-domain
analysis of 6-sec pulse waves, to quantify 28 pulse patterns in TCM. Professor Wang measured
a single pulse wave and performed Fourier transformation to obtain the corresponding 12
meridian frequency spectrum, but it is very different from the clinical practice of pulse
diagnosis. Our team found that the frequency-domain and the tim-domain analysis can be
integrated if Fourier transformation integral formula is applied. Because the extracted data
is big, the characteristic values of time and frequency domain analysis are calculated and
judged by deep learning method.
The purpose of this study is to use the " Integration analysis of time-domain" method to
extract the characteristic values of the radial pulse, and then use deep learning for model
training. That is, after measuring the pulse waves at different positions and depths of the
bilateral radial arteries, by using the pulse diagnostic instrument, to initial signal
processing and to get a single pulse. Then Fourier transformation is performed to obtain the
magnitude and phase parameters of the 12 harmonics (24 variables in total), and then extract
7 time-domain characteristic parameters of a single pulse. The next step to perform Fourier
transformation again using the 6-second pulse waves to obtain high and low frequency spectrum
by using above parameters. The feature parameters obtained by the above two analysis methods
are simultaneously sent to the deep learning-convolution neuron network (CNN) training. Since
the pulse wave changes of the radial artery are related to time, CNN combined with
long-short-term memory work (LSTM) is also used to do the above-mentioned model training. It
is set to compare the differences between the pulse waves of healthy subjects and subjects
with the suboptimal health status. It is also proved whether the frequency-domain analysis
analysis method by Professor Wang and the time-domain analysis method by Dr. Huang is the
same through the deep learning training process. It is possible to develope a new direction
of pulse diagnosis in TCM by deep learning and integrative time-frequency domain analysis.