Huiyan Wang, Yiyu Cheng Pharmaceutical Informatics Institute, Zhejiang University, 310027, Hangzhou, China Abstract-The pulse diagnosis is one of the most important examinations in Traditional Chinese Medicine (TCM). Due to the subjectivity and fuzziness of pulse diagnosis in TCM, quantitative systems or methods are needed to modernize pulse diagnosis. But up to now, the effective models that can classify pulse types according to pulse waves automatically have not been reported, which undoubtedly limits the practical applications of pulse diagnosis in clinical medicines. In this article, a novel quantitative system for pulse diagnosis was constructed based on Bayesian networks (BNs) to build the mapping relationships between pulse waves and pulse types. The results show that the system obtains relative reliable predictions of pulse types, and its predictive accuracy rate reach 84%, which testifies that the model used in our system is feasible and effective and can be expected to facilitate popular applications of TCM.
I. INTRODUCTION Traditional Chinese Medicine is one of the most important complementary and alternative medicines. Pulse diagnosis is one of the four examinations, namely inspection, auscultation and olfaction, inquiry and palpation. The doctors diagnose the patient by feeling his pulse beating at the measuring point of the radial artery. Traditional pulse diagnosis requires long experiences and a high level of skill, and is subjective and deficient in quantitative criteria of diagnosis. This affects the reliability and repeatability of pulse diagnosis. So quantitative methods or models are needed to classify the pulse types automatically. Recently, many researchers have made great efforts in pulse analysis. [1-6]. But the mapping relations between pulse waves and pulse types is not considered, which undoubtedly limits their practical applications in clinical medicines. In TCM, fuzziness and uncertainty are inherent issues in the procedure of pulse diagnosis. An attraction tool for managing various forms of uncertainty is Bayesian networks BNs [7,8], which is able to represent knowledge with uncertainty and efficiently performing reasoning tasks. Comparing with other methods that also can handle uncertainty quite adequately, such as fuzzy logic [9], belief functions [8] etc., BNs have several advantages. Firstly, it is based on a rigorous theory with a vast amount of known results, which prompted some researchers [9-11] to claim that
This work was sponsored by a key grant from the State Administration of Traditional Chinese Medicines of Peoples Republic of China (No. 2000-J-Z-03) probability is the only sensible description of uncertainty and is adequate for all purposes. Secondly, BNs describe causal relationships in graphics mode, which is prone to comprehend and can be used to predict the consequences of intervention. Thirdly, it is often insensitive to imprecision in the numerical probabilities. In view of the advantages, we utilize BNs to construct a system for quantitative pulse diagnosis in TCM. In this article, we develop a new quantitative system for pulse diagnosis based on Bayesian networks (BNs) to build the mapping from pulse parameters to pulse types. This system constructs the mapping relationships between pulse wave parameters and pulse types. The performance of the system was evaluated. The results show that the system performs well in the differentiation of pulse types. II. DATABASES OF PULSE WAVE In our system, the pulse transducer is pressure-adjusting and we can regulate the pressure from 25~150 g . We record the pulse wave when the amplitude reaches the maximum. The pulse sample database (PS-DB) consists of two parts: (1) a total of 407 pulse waves, collecting from 298 patients and 109 healthy volunteers, and (2) diagnostic results of pulse type. The patients were from several hospitals of TCM and healthy volunteers were chosen from the campus students of universities and children from certain kindergartens. The diagnostic results were all given by one group of clinical physicians with three members. In addition, the cases in database did not contain missing values, or in other words, the missing values were regarded as absence, so we did not consider handling missing values in the experiments. Fig.1. Pulse recognition contents Proceedings of the 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference Shanghai, China, September 1-4, 2005 0-7803-8740-6/05/$20.00 2005 IEEE. 5676 The pulse parameters are continuous variables. For convenience, we use the equal-width-intervals discretization method [13], and the interval is set to 5. III. FEATURE EXTRACTION OF PULSE WAVE Generally, the pulse types that often appear in clinic can be classified into eleven types see Fig. 1 according to seven factors, viz. depth, width, length, frequency, rhythm and strength. Fig.2 is an example of pulse wave obtained through the pulse transducer. As the contact pressure increases, the amplitude of the pulse signal first increases, reaching a maximum and then decreases. The pulse wave is composed of three sub-waves, namely the wave SP , the wave EF and the wave FG see Fig. 5. The parameters of pulse wave that utilized in our system are as follows [14]: (a) the contact pressure P 1 , at which the maximum amplitude of pulse wave is attained, (b) the height of the maximum amplitude of SP ( sp h , shown as ' PP in Fig.5), (c) the height of the maximum amplitude of EF ( ef h , shown as ' KK in Fig.5), (d) the height of the maximum amplitude of FG ( fg h , shown as ' LL in Fig.5), (e) the height of the start point of FG ( ff h , shown as ' FF in Fig.5), (f) the height of the start point of EF ( ee h , shown as ' EE in Fig.5), (g) sp ee ep h h r = , (h) sp ff fp h h r = , (i) sp ef es h h r = , (j) sp fg fs h h r = , (k) the cycle of pulse wave 1 C , (l) the ascending time of pulse wave ( 1 t ), (m) the area of the pulse wave ( A ). Thereof ' P , ' E , ' F , and ' L are the projections of P , E , F , K and L onto X axis, respectively. These parameters are meaningful in TCM [16] and testified to be effective in the analysis of pulse [14,15]. According to the pulse definition in TCM, a floating pulse is palpable on light touch and grows faint on hard pressure, while sinking pulse can only be felt on pressing hard, so we conclude that Floating or Sinking (FS) can be distinguished reliably by 1 P . In like manner, the pulse frequency can be reflected by 1 C , which can be used to differentiate Moderate or Rapid (MR). Besides, it were known that the knotted pulse pausing at irregular intervals and intermittent pulse pausing at regular intervals, according to which we define several new parameters as following, which can be representative of the pulse rhythm. (n) 2 1 m m ma N N r = , thereof 1 m N and 2 m N denoting the number of cycles of pulse wave in two adjacent minutes, (o) 1 m c mb N N r = , thereof 1 60 C N c = . For slippery or hesitant or wiry (SHW) and feeble or replete (FR), there are not corresponding parameters yet. So we assume that SHW and FR are influenced by all the parameters defined above. The effective of these parameters will be validated in the experiments as below. IV. QUANTITATIVE DIAGNOSIS MODEL BASED ON BNs The architecture of the system for pulse diagnosis is given (see Fig. 3). The probabilistic reasoning module consists of discovering dependency relationship module, parameter learning and reasoning module (see Fig.4). A new hybrid leaning algorithm [12] for dependency relationship discovery and parameter learning was used, which was developed by our research group and named as * GBPS algorithm. Given the parameterized model, the reasoning module is implemented via Clique Tree Propagation algorithm (CTP) [16], which allows computation sharing among multiple queries and can satisfy the requirement of pulse diagnosis. In addition, we use Markov blanket [8] to perform causal inference. A Markov blanket is the minimum set of nodes that renders node X conditionally independent of all other nodes in the directed graph. The Markov blanket of a node X consists of the parents of X , the children of X , and the parents of the children of X . We can get rid of all other nodes outside Markov blanket of X to obtain simplified BNs without influencing the predictive PAR Fig.3. The architecture of the system for pulse diagnosis Fig.2. An example of pulse wave 133 1 = P E F ' F ' E P S G ' P K L ' L ' K 1 C 1 t 5677 V. EXPERIMENT RESULTS The sample size of some items is relative small, such as knotted, intermittent, hesitant in Table 2. This situation is popular in medical statistics due to the difficulty in obtaining both sufficient cases and reliable data for each case. To solve this problem, we used a stratified fold cross validation technique [17] ( equals to 5 in all of the experiments, named as CV-5) to estimate the predictive accuracy rate (PAR) of the system. Besides, the PAR for each item is defined as follows: samples Total correctly recognized being samples of Number rate Accuracy = The performance of the quantitative system is evaluated through experiments. In the experiments, we trained 5 BNs from PS-DB, which are used for recognize the five pulse types, respectively, based on corresponding parameters of pulse wave. The graphical structures of Markov blanket of SHW and FR are illustrated in Fig.5. From SHW-BNC, we can see that only 5 parameters, namely fp r , es r , sp h , ef h and 1 t , are selected and contribute to the discrimination of SHW. From FR-BNC, it can be noted that 6 parameters, namely es r , ff h , P 1 , fp r , fs r and sp h , survive and are most relevant to the diagnosis of FR. The predictive results are tabulated as below (see Table 1). The average PAR is 84%, which demonstrates that the system performs well, especially for the identification of MR with 97% PAR, and FS with 95% PAR. For SHW, FR and KI, the PAR is relative low. The error might come from three parts. Firstly, the information contained in pulse wave was not adequate, which can be conquered by amelioration of pulse transducer. Secondly, the pulse parameters that utilized in our system were not very discriminating for diagnosing these pulse types. Thirdly, the corresponding pulse samples were insufficient. Further improvement of diagnostic performance of the system could be expected in the near future. Pulse types SF MR KI FR SHW Average PAR 0.95 0.97 0.76 0.84 0.69 0.84 TABLE 2 LISTS OF VALUES OF PULSE TYPES AND CORRESPONDING NUMBER OF SAMPLES Pulse type Value ID Number of samples Depth Regular depth Sinking Floating 0 1 2 FS 184 204 19 Frequency Regular frequency Moderate Rapid 0 1 2 MR 288 82 37 Rhythm Non-knotted and Non-intermittent Knotted Intermittent 0 1 2 KI 370 14 23 Strength Feeble Replete 0 1 FR 132 275 Quality Slippery Hesitant Wiry 0 1 2 SHW 185 32 190 VI. CONCLUSION Pulse diagnosis is the most important components in TCM. But the complexity and elusiveness of diagnosis method confine its development and generalization. In order to classify pulse types automatically, we propose a quantitative system based on BNs to build the mapping relationships between pulse waves and pulse types. The experiment results validate that the system for pulse diagnosis is effective. Note that, the feature of our system is that it can keep enhancing the predictive accuracy of pulse types by learning the experiences of experts of TCM. REFERENCES Fig.4. The architecture of probabilistic reasoning model TABLE 1 PREDICTIVE RESULTS OF THE SYSTEMFOR PULSE Fig.5 .The Markov blanket of SHW and FR 5678 [1] H. L. Lee, S. Suzuki, Y. Adachi, M. Umeno and Shan K, Fuzzy Theory in Traditional Chinese Pulse Diagnosis, Proceeding of International Joint Conference on Neural Networks, Nagoya, Japan, vol. 10, pp. 774-778, 1993. [2] Y. Z. Yoon, M. H. Lee and K. S. Soh., Pulse Type Classification by Varying Contact Pressure, IEEE Engineering in Medicine and Biology Magazine, vol. 19, pp. 106-110, 2000. [3] B. H. Wang and X. J. Lin, Detecting system and power-spectral analysis of pulse signals of human body, Fourth International Conference on Signal Processing, vol. 2, pp. 1646 - 1649, 1998. [4] S. E. Fu and S. P. Lai, A system for pulse measurement and analysis of Chinese medicine, Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society, vol. 5, pp. 1695-1696, 1989. [5] Y. C. Yu, S. J. Yeh and C. C. Chiu, Analysis of autonomic nervous system by applying artificial neural network to categorize pulse signals from Finapres (Abstract), Clinical Autonomic Research, vol. 5, pp. 295, 1996. [6] C. C. Chiu, S.J. Yeh and Y.C. Yu, Classification of the pulse signals based on self-organizing neural network for the analysis of the autonomic nervous system, Chinese Journal of Medical and Biological Engineering, vol. 16, pp. 461- 476, 1996. [7] J. Pearl, Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference, Morgan Kaufmann Publishers, Inc. San Francisco, California, 1988. [8] R. E. Neapolitan, Probabilistic Reasoning in Expert Systems, John Wiley and Sons, New York, NY, 1990. [9] D. Dubois, H. Prade and Yager RR, eds, Fuzzy information engineering: A guide tour of applications, Wiley, New York, pp. 149-162, 1997. [10] L. N. Kanal, J. F. Lemmer, ed, Uncertainty in artificial intelligence., NorthHolland, Amsterdam, 1986. [11] D.V. LINDLEY, The probability approach to the treatment of uncertainty in artificial intelligence and expert systems, Statistical Science, vol. 2, pp. 17- 24, 1987. [12] X. W. Wang, H. B. Qu, P. Liu and Y. Y. Cheng, A self-learning expert systems for diagnosis in traditional Chinese medicine, Expert Systems with Applications, vol. 26, pp. 557-566, 2004. [13] J. Dougherty, R. Kohavi and M. Sahami., Supervised and unsupervised discretization of continuous features, In Proceedings of Twelfth International Conference on Machine Learning, pp. 194-202 1995. [14] M. Z. Xie, S. Z. Li and B. X. Li, Observation on ordinary persons parameter of pulse condition and pulse map, Hunan Guiding Journal of TCMP, vol. 6, pp. 9-11, 2000, in Chinese [15] H. Y. Wang and J. Zhu, Edge detection based on wavelet transform and its application in porcelain bottles fault monitoring, Automation of Electric Power Systems, vol. 28, pp. 76-79, 2004. In Chinese. [16] S. L. Lauritzen and D. J. Spiegelhalter, Local computations with probabilities on graphical structures and their application to expert systems, Journal of the Royal Statistical Society, Series B, vol. 50, pp. 157-224, 1988. [17] M. Mullin and R. Sukthankar, Complete cross-validation for nearest neighbor classifiers, Proceedings of ICML, pp. 639-646, 2000. 5679