Thang, C. et al.

Paper:

Applying Fuzzy Logic and Neural Network to Rheumatism Treatment in Oriental Medicine
Cao Thang , Eric W. Cooper , Yukinobu Hoshino , and Katsuari Kamei
School of Science and Engineering, Ritsumeikan University 1-1-1 Noji-Higashi, Kusatsu, Shiga 525-8577, Japan E-mail: thangc@spice.ci.ritsumei.ac.jp College of Information Science and Engineering, Ritsumeikan University 1-1-1 Noji-Higashi, Kusatsu, Shiga 525-8577, Japan E-mail: cooper@is.ritsumei.ac.jp, kamei@ci.ritsumei.ac.jp Dept. of Electronic and Photonic System Engineering, Kochi University of Technology 185 Miyanoguchi, Tosayamada, Kami, Kochi 782-8502, Japan E-mail: hoshino.yukinobu@kochi-tech.ac.jp [Received October 31, 2005; accepted March 31, 2006]
Graduate

In this paper, we present an application of soft computing into a decision support system RETS: Rheumatic Evaluation and Treatment System in Oriental Medicine (OM). Inputs of the system are severities of observed symptoms on patients and outputs are a diagnosis of rheumatic states, its explanations and herbal prescriptions. First, an outline of the proposed decision support system is described after considering rheumatic diagnoses and prescriptions by OM doctors. Next, diagnosis by fuzzy inference and prescription by neural networks are described. By fuzzy inference, RETS diagnoses the most appropriate rheumatic state in which the patient appears to be infected, then it gives a prescription written in suitable herbs with reasonable amounts based on neural networks. Training data for the neural networks is collected from experienced OM physicians and OM text books. Finally, we describe evaluations and restrictions of RETS.

Keywords: medicine

neural network, fuzzy inference, oriental

1. Introduction
Rheumatism is an arthritis disease widespread in all Vietnamese population groups, unfortunately inuencing socioeconomic aspects of Vietnam. Among all soft tissue diseases, rheumatism accounts for 15% and the most common rheumatic type, joint degeneration, accounts for 10% [1]. In Vietnam, therapeutic treatments for rheumatism are physical methods, anti-inammatories and oriental medicine (OM). Among these, OM is an indispensable part because it has fewer side-effects than western medicine and gives good treatment results. Besides, herbal prescriptions are easy to nd and relatively cheap in comparison with western drugs. The number of Vietnamese patients treated by OM is about 50%.

Accurate diagnoses have an important role in treating diseases. Building a successful decision support system such as RETS based on knowledge from experienced OM physicians will help moderate evaluations in rheumatic diagnoses, which tend to be subjective. It will indirectly help physicians to provide the right treatment to the right patients, improving the quality of the health care services as a whole. It also will help qualied and experienced physicians in OM to maintain and share their profound knowledge with colleagues and to assist medical students or young physicians, especially those living and working in rural areas. In the last 50 years, the advent of the computer has greatly stimulated developments of Articial Intelligence (AI), especially Expert System (ES) and Decision Support System (DSS) which perform the roles of a specialist or assist people in carrying out works requiring specic expertise. Since the beginning of AI, ES and DSS have been successfully applied to Western Medicine (WM), and then to OM with reasoning techniques including Uncertainty Reasoning [5, 6], Fuzzy Logic [24, 7, 8], Casebased Reasoning [9, 10] and Neural Network [13, 14]. Most of the ES and DSS in OM focused on problems of disease diagnoses and analyses with the specic characteristics of OM such as four inspection steps, the six internal organs and the viscera, Yin and Yang [2, 7, 8]. A few recent researches are concerned with integrating WM and OM [11], and modifying treatment herbal prescription in OM [12] using Fuzzy Logic. This research presents an application of Fuzzy Logic and Neural Network (NN) into RETS: Rheumatic Evaluation and Treatment System in OM. In diagnosing stage of RETS, based on severities of observed symptoms on patients and importance degrees of clinical symptoms in each rheumatic state, fuzzy inference is used to decide which rheumatic states the patient has. Then in the prescribing stage, the fuzzy severities will be put into a corresponding NN to get an appropriate herbal prescription. The importance degrees of clinical symptoms are evalu-

Journal of Advanced Computational Intelligence and Intelligent Informatics

Vol.11 No.1, 2007

Applying Fuzzy Logic and NN to Rheumatism Treatment in OM

Fig. 1. Diagram of diagnosing and prescribing rheumatism by OM doctors.

Fig. 2. DSS for diagnosing and treating rheumatism in OM.

ated by rheumatism treatment experts. Training data for the NN is typical treatment prescriptions collected from experienced doctors and OM text books.

3. Diagnosing by Fuzzy Inference

In OM, physicians usually give herbal prescriptions based on the severities of clinical symptoms such as high fever, slightly numb joints, moderately yellow urine etc. These ambiguous expressions of symptoms make it unsuitable for traditional quantitative approaches to build a DSS for OM. Fuzzy sets, known for their abilities to deal with vague variables using membership functions rather than with crisp values, have proven to be one of the most powerful approaches to resolve this problem. They also enable developers to use linguistic variables and build friendly user interfaces. OM physicians usually explain diagnosing procedures with such expressions as this patient has these typical symptoms with these severities, so I prescribe these herbs with these amounts. These expressions can be represented quite naturally in IF-THEN fuzzy rules. In addition, fuzzy rules can give expert-like explanations, making it easier for doctors to understand the DSS. So far, based on fuzzy logic many practical applications in both WM and OM have been built [2, 7, 8], including rheumatic disease [3, 4]. In RETS, from severities of observed symptoms on the patient and importance values of the symptoms, fuzzy inference is used to determine which rheumatic states the patient has

2. Outline of RETS
According to OM, each disease has specied states and each disease-state is cured by standard herbal prescriptions. Such a prescription is easily found in medical text books or OM reference books. With a patient, standard prescriptions become much more efcient and effective if physicians add more suitably additional herbs to it and then adjust all of its herbal amounts. The herbal adjustments are often based on the severities of observed symptoms on the patient, and their effects mainly depend on physicians treatment experience. Only experienced physicians can give patients suitable prescriptions with reasonable adjustments. In OM, rheumatism primarily consists of 12 disease states and 32 typical clinical symptoms. The number of rheumatic treatment herbs is 63 [15]. Based on the severities of observed symptoms, doctors diagnose and classify rheumatic states, then give a corresponding herbal prescription with reasonable amounts in grams. Fig. 1 shows the process of diagnosing and prescribing rheumatism by OM doctors. Such a process can be suitably assisted by a DSS or an ES as shown in Fig. 2 [17]. Roles of the functional parts in Fig. 2 are as follows: Knowledge Acquisition: Surveys symptoms, prescribing rules, explanations and sample prescriptions. Knowledge Base: Consists of symptoms, disease states, inference rules, training data and explanations. Fuzzy Inference: Checks rules, calculates weights and advises the most serious rheumatic state. Neural Networks: Gives prescriptions with reasonable herbal adjustments. Interface: Obtains symptoms and their severities from users and shows inferential results. Explanation: Helps users to understand OM, rheumatism, and explains the results. Vol.11 No.1, 2007

3.1. Symptom and Rule Expression Suppose that rheumatism has m clinical symptoms, l rheumatic states. A rheumatic state is determined by n clinical symptoms. O O Let SO S1 Sm be a set of observed symptoms O is a fuzzy proposition representing on a patient where S i a symptom. Let H H1 Hl be a set of the rheumatic states. Rj R Rj Let S S1 Sn j be a set of symptoms in l where R j is generally premise of rule R j j 1
5

Journal of Advanced Computational Intelligence and Intelligent Informatics

Thang, C. et al.

3.2. Fuzzy Inference Process If an observed symptom S iO is found in the premise of j rule R j , premise weight wSi of SiO in R j is calculated as:
wSi
j

SO
i

where is a t-norm operator, x y x y in RETS. If symptoms S R j of R j match with observed symptoms O , weight w of R is calculated as: S Rj j wR j
Si S

Si

Rj

. . . . . . . . . . . . (3)

R
j

SO

j wSi

. . . . . . . . . . (4)

where is a t-conorm operator, this t-conorm should be compatible with (2), x y x y in RETS. Then RETS nds the most serious state H having the largest wR j value among l rheumatic states H hm wRm max wR j
j

. . . . . . . (5)

Fig. 3. Diagram of inference procedure.

Figure 3 shows the diagram of the inference procedure in RETS. When the inputs, severities of observed symptoms, are matched with one or more rheumatic states, the system nds the most appropriate rheumatic state H corresponding with these inputs. If the inputs are not enough to match with any rheumatic state, RETS gives advice about the closest rheumatic state. In this case, the patient may have diseases other than rheumatism.

described in the following form: and S n j IF S1 j and S2 j and THEN the rheumatic state is H j
R R R

4. Prescribing by Neural Networks

. . . . .
R

(1) NN is an effective technique to help doctors to analyze, model and make sense of complex clinical data across a broad range of medical applications [13]. It enables intelligent systems to learn from experience, examples and clinical records, improving performance of the systems over time. Based on knowledge accumulated from experienced doctors and hospital information systems, NN can wisely give doctors good decisions, helping to moderate subjective evaluation in diagnosing and prescribing diseases. So far many useful NN applications have been developed [13, 14]. In RETS, trained by rheumatic treatment knowledge collected from skilled OM doctors, NN is used to give herbal prescriptions with reasonable amounts. An important point in preprocessing training data is to select the right sets of input and output features. Raw data are prescription rules and herbal treatment prescriptions with typical observed severities gathered from experienced doctors. Features should be reasonably chosen so that from trained NNs we can get appropriate prescriptions in accordance with observed symptoms and diagnosed rheumatic states. For the inputs, there are two types of symptoms. The rst type is associated with Boolean values: Yes (true, coded by 1) and No (false, coded by 0). Observed severities in the second type are associated with 5 linguistic values in company with fuzzy intervals: no (0.00), slightly (0.25), moderately (0.50), relatively (0.75) and clearly Vol.11 No.1, 2007

Let the two following fuzzy values in S iO and Si j be dened:

SO 0 1 : truth value of S iO given by doctors when

diagnosing. SO patient, SO
i i i

1 means SiO clearly appears on the

0 means SiO does not appear on the

patient, and 0 SO 1 means SiO appears on the i patient with severity SO .

i

state H j given by skilled doctors via survey in advance, where

n

Rj Si

0 1 : importance value of S i j for rheumatic

i 1

SiR j

1 . . . . . . . . . . . . (2)

Si

Rj

0 means Si j totally does not affect H j ,

R Si j

Si

Rj

is the only symptom affecting H j , and 0 means R Rj 1 means Si j affects H j with fuzzy importance Si Rj .
Si

Journal of Advanced Computational Intelligence and Intelligent Informatics

Applying Fuzzy Logic and NN to Rheumatism Treatment in OM

(1.00). For the outputs, there are two main kinds of herbs in a prescription, treating and conducting herbs. Treating herbs directly cure the infected disease while conducting herbs help patients organisms to easily absorb herbal effects. The total number of herbs in a prescription is from 9 to 15. Amounts of treating herbs are often adjusted by severities of the observed symptoms whereas amounts of conducting herbs are normally unchanged as in the standard prescriptions. The number of observed symptoms used to adjust herbs in a standard prescription is often from 6 to 12. In training data for NNs of RETS only symptoms that affect herbal adjustments are used for inputs and all of treating and conducting herbs are used for outputs. For the same rheumatic state, prescriptions by different doctors might not look similar because some doctors use some herbs but others prefer equivalent herbs that also give the same effects but come in different amounts. To avoid using many equivalent herbs for the same prescriptions in training data, lists of herbs in the standard prescriptions from text books of Hanoi Medicine University are used and claried by experienced doctors. For user reference, equivalent herbs are written in the description part of sample prescriptions. Depending on each rheumatic state, the amounts of herbs in prescriptions vary from 2 to 60 grams. The error in the adjusted amounts of an herb accepted by doctors is usually 0.5 gram for small amounts and 1.0 gram for large amounts. Since the output range of NN is chosen from 0 to 1, amounts of herbs are normalized as coefcients. The coefcient ck of amounts of herb k in training data, and the actual amount WkP of herb k in the prescriptive results is calculated as: ck WkP where WkT WkT W . . . . . . . . . . . . .

Fig. 4. One neural network in RETS.

severities of observed symptoms and same infected rheumatic states with the diagnosed patient, then shows prescriptions of these cases and their explanations from experienced doctors.

6. Implementation
Based on the text books, a preliminary survey and real rheumatic prescriptions from experienced doctors in Thaibinh OM College, we have assessed important fuzzy values of symptoms in rheumatic states, chosen standard prescriptions from the text books, claried additional and equivalent herbs, selected specic symptoms that affected herbal adjustments, then generated 5,000 doctor-like prescriptions with combinations of severities of the statespecied symptoms using doctor-prescribing rules and linear methods with ranges of herbal adjustments. Training data for NN are the generated prescriptions together with 460 real rheumatic prescriptions from the experienced doctors. Since most of herbs are written in Chinese and Vietnamese, some herbs have English or French names and some do not, in RETS the names of herbs are written in Vietnamese. Symptoms and general explanations are written in both Vietnamese and English. Detailed explanations with OM terms are written in Vietnamese. There are 12 networks corresponding with 12 rheumatic states. Each NN has 3 layers as shown in Fig. 4. Inputs to NNs are state-specied symptoms SiO i 1 2 m with SO and outputs from NNs are i coefcients ck k 1 2 p of amounts of herbs. The number of neurons in the hidden layer is equal to the number of output neurons l p. NNs are back-propagation networks adopting sigmoid or hyperbolic tangent activated functions. To accelerated training, adaptive learning and momentum term are also used. Figs. 5 and 6 show the interfaces of RETS for diagnosis and prescription results, respectively. 7

(6) (7) is

ck W

. . . . . . . . . . . . .

is amount of herb k in training data the maximum amount of a herb in the prescription.

and W

5. Explanations
One of indispensable features of ES and DSS is capabilities to offer explanations. Logical explanations of RETS can help users, especially young doctors or medical students, to deeply understand inference results. Explanations also make it easier for experienced doctors to revise related sample cases in training data. Currently RETS has general and detailed explanations about rheumatic states and prescriptions.

General explanation: after fuzzy inference process, from the knowledge base RETS obtains a general explanation about the most serious rheumatic state, then shows a fuzzy graph of all related states and fuzzy weights wR j of rules. Detailed explanation about similar cases: in training data RETS nds similar cases that have same

Vol.11 No.1, 2007

Journal of Advanced Computational Intelligence and Intelligent Informatics

Thang, C. et al.

Fig. 5. Interface of diagnosis.

Fig. 6. Interface of prescription results and explanations.

Journal of Advanced Computational Intelligence and Intelligent Informatics

Vol.11 No.1, 2007

Applying Fuzzy Logic and NN to Rheumatism Treatment in OM

7. Evaluations
Combining NN and fuzzy inferences, we can have a more powerful and effective DSS with learning, reasoning and explaining capabilities for evaluating and treating rheumatism. The inference rule given by (1) is equivalent to the following rule form: IF fuzzy severities of symptoms R R and S n j is S1 j is R j and THEN rheumatic state is H j with certainty factor wR j Using the rule form (8), a DSS may need thousands inference rules with many combinations of symptoms in premises. Not only do they take much time for developers to accomplish the rule acquisition but also much effort for domain experts to revise all of the rules. Using the rule form (1) with (2), RETS uses just 12 inference rules. Experienced doctors have conrmed that the inference results are quite good and it is easy to review the knowledge presented by the rules. In our experiment, we randomly split training data into two parts of 80% and 20% and used the former for training and the later for testing. All of nonlinear relations (real prescriptions and rules of prescribed herbs) as well as linear relations (ranges of herbal adjustments) were well learnt by NNs. Depending on the number of inputs and outputs, each NN can learn about 1000 prescriptions within an accuracy of 10 2 mean square error with both training and testing data (equivalent to error of 0.1 gram for each herb). In case of unknown inputs, RETS shows the fuzzy graph of infected rheumatic states, recommends the most proper state in which the patient seems to be infected and gives explanations by fuzzy inference, then shows the advised prescription with appropriate amounts of herbs by NN. Most of these prescriptions are completely compatible with the real prescriptions, prescribing rules and linear ranges of herbal adjustments in the training data. In an evaluation with doctors in Hanoi Oriental Medicine Institute, we asked the doctors for considering and then giving 50 rheumatic cases including real cases that they have treated, then compare RETS results with the doctors opinions. All prescriptions from RETS could be practically used said by doctors. About the herbal adjustments including additional herbs and amounts of herbs in the nal prescriptions, 94% prescriptions from RETS are totally agreed and 6% are fairly accepted. Experienced doctors have also used RETS to illustrate treatments of clinical rheumatism cases for medical students.
S1

Sn

Rj

. . . (8)

fuzzy inference and herbal prescribing system by NN. We conrmed that RETS has high performance and high applicability for diagnosing and prescribing rheumatism. Unfortunately, like other DSS and ES which are restricted to a narrow domain of expertise, RETS is developed for diagnosing and prescribing rheumatism only. It lacks much real knowledge of human philosophy [16]. If a patient has other diseases besides rheumatism, doctors cannot solely rely on this system since they do not have evidence to control potential effects of the herbal prescriptions on the other concurrent diseases. Hence, it is recommended that the system be used only for patients with rheumatism alone, not for those with other concurrent diseases.
References:
[1] L. A. Ba, Treating reality rheumatism, J. of Medicine and Pharmacy HCMC, June, 2001. [2] F. A. Maysam, D. G. von Keyserlingk, D. A. Linkens, and M. Mahfouf, Survey of utilization of fuzzy technology in Medicine and Healthcare, Fuzzy Sets and Systems, Vol.120, pp. 331-349, 2001. [3] M. B. Serrano, C. Sierra, and R. Lopez de Mantaras, RENOIR: an expert system using fuzzy logic for rheumatology diagnosis, Internat. J. of Intell. Systems, Vol.9, No.11, pp. 985-1000, 1994. [4] K. Boegl, F. Kainberger, K. P. Adlassnig, G. Kolousek, H. Leitich, G. Kolarz, and H. Imhof, New approaches to computer-assisted diagnosis of rheumatologic diseases, Radiologe, Vol.35, No.9, pp. 604-610, 1995. [5] E. H. ShortliEe, Computer-Based Medical Consultations, MYCIN, Elsevier, NY, 1976. [6] R. Montironi, P. H. Bartels, P. W. Hamilton, and D. Thompson, A typical adenomatous hyperplasia (adenosis) of the prostate: development of a bayesian belief network for its distinction from well differentiated adenocarcinoma, Hum. Pathol., Vol.27, No.4, pp. 396-407, 1996. [7] R. J. Guo, B. R. Ma, M. Xie, and W. L. Chen, Fuzzy set model and computerised diagnosis system in traditional Chinese medicine, Chinese J. of Comput., Vol.4, No.4, pp. 260-266, 1981. [8] H. P. Nguyen, Approximate reasoning for oriental traditional medical expert systems, IEEE Internat. Conf. on Systems, Man, and Cybernetics, Computational Cybernetics and Simulation, 4, NY, pp. 3084-3089, 1997. [9] R. Schmidt, B. Pollwein, L. Filipovici, and L. Gierl, Adaptation and abstraction as steps towards case-based reasoning in the real medical world: Case-based selection strategies for antibiotics therapy, Proc. of MEDINFO95, North-Holland, Amsterdam, pp. 947951, 1995. [10] H. P. Nguyen, B. T. Nguyen, and A. Ohsato, Develop Case-based Reasoning for Medical Consultation Using the Importance of Features, J. of Advanced Computational Intelligence, Vol.6, No.1, pp. 41-50, 2002. [11] H. P. Nguyen, Toward Intelligent Systems for Integrated Western and Eastern Medicines, TheGioi Publisher, Hanoi, 1997. [12] H. P. Nguyen, S. Pratit, and K. Hirota, Fuzzy Modeling for Modifying Standard Prescriptions of Oriental Traditional Medicine, J. of Advanced Computational Intelligence and Intelligent Informatics, Vol.7, No.3, pp. 339-347, 2003. [13] R. Dybowski and V. Gant, Clinical applications of articial neural networks, Cambridge University Press, 2001. [14] A. Bezerianos, S. Papadimitriou, and D. Alexopoulos, Radial basis function neural networks for the characterization of heart rate variability dynamics, J. of Articial Intelligence in Medicine, Vol.15, No.3, pp. 215-234, 1999. [15] T. Thuy, P. D. Nhac, and H. B. Chau, Lectures in Oriental Medicine, Medicine Pub. Hanoi, Vol.2, pp. 160-165, 2002. [16] M. Negnevitsky, Articial Itelligence A Guide to Intelligent Systems, Pearson Education Limited, 2002. [17] J. Durkin, Expert System, Design and Development, Prentice Hall Inc, NY, 1994.

8. Conclusions
We built RETS: Rheumatic Evaluation and Treatment System in OM, and showed the diagnosing system by Vol.11 No.1, 2007 Journal of Advanced Computational Intelligence and Intelligent Informatics 9

Thang, C. et al.

Name:
Cao Thang

Name:
Yukinobu Hoshino

Afliation:
Graduate School of Science and Engineering, Ritsumeikan University

Afliation:
Associate Professor, Department of Electronic and Photonic system Engineering, Kochi University of Technology

Address:
1-1-1 Noji-Higashi, Kusatsu, Shiga 525-8577, Japan

Address:
185 Miyanoguchi, Tosayamada, Kami, Kochi 782-8502, Japan

Brief Biographical History:

1994 Bachelor of Electronics, Hanoi University of Technology 2005 MS, Graduate School of Science and Engineering, Ritsumeikan University

Brief Biographical History:

2002 Ph.D., Ritsumeikan University 2002- Assistant Professor in Ritsumeikan University 2006- Associate Professor in Kochi University of Technology

Main Works:

A Decision Support System for Rheumatic Evaluation and Treatment in Oriental Medicine Using Fuzzy Logic and Neural Network, Lecture Notes in Articial Intelligence (LNAI), Vol.3558, pp. 399-409, Springer-Verlag, Berlin Heidelberg, 2005.

Main Works:

Membership in Learned Societies:

Vietnamese Fuzzy Society

An Application of FEERL (Fuzzy Environment Evaluation Reinforcement Learning) to Lights OutGame and Avoidance of Detour Actions in Search, Transactions of the Institute of Systems, Control and Information Engineers, Vol.14, No.8, pp. 395-401, 2001. A Proposal of Reinforcement Learning with Fuzzy Environment Evaluation Rules and Its Application to Chess, Journal of Japan Society for Fuzzy Theory and Systems, Vol.13, No.6, pp. 626-632, 2001.

Membership in Learned Societies:

Japan Society for Fuzzy Theory and Systems The Institute of Systems, Control and Information Engineers The Society of Instrument and Control Engineers Japan Society of Kansei Engineering

Name:
Eric W. Cooper

Name:
Katsuari Kamei

Afliation:
Associate Professor, Computer Science, College of Information Science and Engineering, Ritsumeikan University

Afliation:
Professor, Human and Computer Intelligence, College of Information Science and Engineering, Ritsumeikan University

Address:
1-1-1 Noji-Higashi, Kusatsu, Shiga 525-8577, Japan

Address:
1-1-1 Noji-higashi, Kusatsu, Shiga 525-8577, Japan

Brief Biographical History:

1998 MS, Ritsumeikan University 2002 Ph.D., Ritsumeikan University 2003 Joined COE program as post-doctoral researcher 2006- Associate Professor, Ritsumeikan University

Brief Biographical History:

1983- Assistant Professor, Ritsumeikan University 1993- Associate Professor, Ritsumeikan University 1998- Professor, Ritsumeikan University

Main Works:

Main Works:

Modeling Designers Color Decision Processes Through Emotive Choice Mapping, Lecture Notes in Computer Science, Springer-Verlag, Vol.3558, pp. 410-421, 2005.

Membership in Learned Societies:

Human Interface Society Institute of Systems, Control and Information Engineers Association for Computing Machinery

Townscape Colour Planning System Using an Evolutionary Algorithm and Kansei Evaluations, Proceedings of 2006 IEEE International Conference on Fuzzy Systems, pp. 4322-4329, 2006. A Fuzzy Clustering Based Selection Method to Maintain Diversity in Genetic Algorithms, Proceedings of 2006 IEEE Congress on Evolutionary Computation, pp. 10364-10369, 2006.

Membership in Learned Societies:

Japan Society for Fuzzy Theory and Intelligent Informatics The Institute of Systems, Control and Information Engineers Human Interface Society Japan Japan Society of Kansei Engineering IEEE Computational Intelligence Society

10

Journal of Advanced Computational Intelligence and Intelligent Informatics

Vol.11 No.1, 2007