Special Issue Article
Computer-aided diagnosis expert system for cerebrovascular diseases Xu Chen*1, Zhijun Wang*2, Chrisopher Sy3, Xiaokun Liu1, Jinwu Qian2, Jia Zheng1, Zhiqiang Dong1, Limei Cao1, Xiang Geng1, Shuye Xu2, Xueyuan Liu4 1
Department of Neurology, Shanghai Eighth People’s Hospital, Shanghai, China, 2School of Mechatronics Engineering and Automation, Shanghai University, Shanghai, China, 3Department of Neurological Surgery, Wayne State University School of Medicine, Detroit, MI, USA, 4Department of Neurology, Shanghai Tenth People’s Hospital, Shanghai, China Objective: To establish an expert diagnosis system for cerebrovascular diseases (CVDs) and assess accuracy of the diagnosis system. Methods: An expert diagnosis system for CVDs was established and evaluated using actual clinical cases. Results: An expert diagnosis system for CVDs was established and tested in 319 clinical patients. Diagnosis accordance was obtained in 307 patients (the diagnosis accordance rate was 96.2%). Involved were 223, 7, 23, 54 and 12 patients with cerebral thrombosis, cerebral embolism, transient ischemic attack, cerebral hemorrhage and subarachnoid hemorrhage, respectively; and diagnosis accordance was obtained in 219 (98.2%), 6 (85.7%), 23 (100%), 48 (88.9%) and 11 (91.7%), respectively. Conclusion: Overall, the case analysis results support and demonstrate the diagnostic reasoning accuracy of the expert diagnosis system for CVDs. With the expert diagnosis system, medical experts’ diagnosis of CVDs can be effectively mimicked and auxiliary diagnosis of CVDs has been preliminarily realized, laying a foundation for increasing the diagnostic accuracy of clinical diagnoses as it pertains to CVDs.
Keywords: Cerebrovascular diseases, Diagnosis, Computer, Expert system
Introduction An ‘expert system’ is an intelligent computer program, one which endows a computer with select knowledge of human experts in a specific field so that it can mimic experts’ thought processes, reasoning, and judgment – allowing problems to be solved at ‘expert’ levels. Namely, an expert system is an intelligent appliance which has a repertoire of expertise-level knowledge in related fields and can use artificial intelligent techniques to simulate a human expert’s approach to a particular problem.1 Cerebrovascular diseases (CVDs) are common and frequent in middle-aged and older individuals in China. Patient morbidity has increased in incidence with a trend towards lower age groups in recent years – notably also becoming more salient in rural areas. According to a publication of the Ministry of Health in 2008, CVDs have become the top cause of death in Chinese people. The high incidence rate, high death rate, and high mutilation rate as a result of CVDs not only seriously impairs patient health and quality of life but also results in heavy medical, *These authors contributed equally to this work. Correspondence to: Xu Chen, Department of Neurology, Shanghai Eighth People’s Hospital, Shanghai 200235, China. Email: [email protected]; Xueyuan Liu, Department of Neurology, Tenth People’s Hospital, Shanghai 200072, China. Email: [email protected]; Shuye Xu, School of Mechatronics Engineering and Automation, Shanghai University, Shanghai 200444, China. Email: [email protected]
468
ß W. S. Maney & Son Ltd 2014 DOI 10.1179/1743132814Y.0000000347
economic, and social burdens to the nation and patients’ families. At present, Chinese people usually do not have much knowledge on the prevention and treatment CVDs. Additionally, the diagnosis and treatment capacities of hospitals from different regions and of different grades vary widely – largely due to diagnostic and therapeutic methods which are not normalized, leading to diminished therapeutic responses and poorer outcomes. Conventional diagnostic methods in foreign and domestic medicine are greatly influenced by subjective factors, and the accuracy of clinical diagnoses correlates closely to a doctor’s skills, knowledge, and experience. Due to the diagnostic complexity of CVDs and the necessity of giving a definite diagnosis, the subjectivity of clinical judgements face pressures not only to be correct but to be relatively quick. For example, we had a patient who was suddenly afflicted with hemiparalysis, aphasia, headache, dizziness, etc. and we could not make a definite diagnosis the first time. The most pressing issue for us was to differentiate whether it was ischemic stroke or hemorrhagic stroke – an example of a time-sensitive situation where definitive treatment cannot be started without a sure diagnosis. In some remote areas of China, diagnosis of these diseases is severely limited by doctors’ experience, lack of medical equipment, and inadequate facilities resulting in a diagnostic accuracy rate that is very low. Clearly, in such
Neurological Research
2014
VOL .
36
NO .
5
Chen et al.
areas – where abundant experience and professional knowledge can have a large impact – the need to support expert system technology is magnified. A powerful tool for timely and accurate treatment of patients with CVDs can be implemented through a systematic, institutionalized, and normalized computer platform with sufficient expert knowledge for diagnosis of CVDs; especially one with a software database that can be further enriched with continuous addition of more practical clinical cases. Expert diagnosis systems have already been used in some medical fields including oncology,2 endocrinology,3 and medical imageology,4 to name a few and where all have yielded favorable results. In neurology, expert diagnosis systems for cerebral aneurysms5 and headache6 have also been reported. Regarding the expert diagnosis system for CVDs, Spitzer et al.7 from Germany developed an expert system named MICROSTROKE in as early as 1989 for diagnosing different apoplexy types; in the same period, Spitzer et al.7 developed the TOPSCOUT expert system for positioning apoplexy and identifying distribution of affected vessels depending on patients’ symptoms and signs. In 2005, Li8 from Guangzhou University initiated the design of a computer diagnosis system for CVDs using a Visual Basic 6.0 object-oriented highlevel programming language; in 2006, Li et al.9 designed an intelligent case-based reasoning system for auxiliary diagnosis of CVDs; in 2008, Liu et al.10 carried out a study on an expert diagnosis system for CVDs using BP neural networks; in 2008, Ziegler et al.11 conducted a study on the role of mobile computing systems in the preclinical care of stroke – one which can offer new and innovative approaches in improving intersectional acute stroke care; in 2012, Takao et al.12 developed a system named i-Stroke for rapidly exchanging diagnostic images, clinical details, and management information through mobile devices (smart phones) and a Twitter system (e.g. transferring clinical data, CT, MR, angiographic, intraoperative images, and expert opinions in real time). Despite substantial research within the field, in general, there are currently no reports available in China about development of related software in regards to the clinical diagnosis of CVDs. As such, this study was intended to establish an expert diagnosis system for CVDs and to test it with verified clinical cases to assess its diagnosis accuracy.
Materials and Methods Hardware environment for system development Personal computers with .20 GB of free hard drive space and .256 MB of random access memory were used; the CPUs were P 1.2 GHz above.
Software environment for system development The operating systems used were Windows98/2000/ XP; supporting software appliances were Excel 2000 and Access 2000 (or higher versions of the two).
Computer-aided diagnosis expert system for cerebrovascular diseases
Programming language and version number, program size The software was developed in Microsoft Visual Basic 6.0; the program size was more than 800 line codes.
Fundamental principles for system design 1. rough set theory;13,14 2. diagnosis and classification of CVDs.
Cerebrovascular diseases in the system include ischemic CVDs and hemorrhage CVDs. Ischemic CVDs include cerebral infarction (cerebral thrombosis and cerebral embolism) and transient ischemic attack. Hemorrhagic CVDs include subarachnoid hemorrhage and cerebral hemorrhage.
Composition of the expert diagnosis system System structure The structure of the established expert diagnosis system for CVDs is presented in Fig. 1. It consists of six parts, i.e. a human-computer interface, an inference engine module, a knowledge base, a knowledge base management module, an interpretation module and a dynamic database. The knowledge base accommodates professional knowledge provided by medical experts in CVDs. The inference engine performs reasoning according to the knowledge and gives diagnostic conclusions; when a user selects and sets a patient’s specific symptoms, signs and diagnostic highlights and the levels of confidence, imprecise reasoning is performed following the controlling strategy of forward reasoning with knowledge in the knowledge base. The interpretation module interprets the reasoning process of the system. The dynamic database accommodates incipient facts provided by the user and such information as intermediate and final results obtained in operation of the system. The human/computer interface is an interface for interaction between the expert system and experts in the field, knowledge engineers and general users and also provides case maintenance and help. Users may input patients’ clinical symptoms, signs and past histories via the interface for user interaction.
Knowledge presentation, imprecise reasoning In the system, knowledge is presented with a classic production system and imprecise reasoning is performed based on the level of confidence so that each piece of evidence and each rule as well as diagnostic conclusions are presented with levels of confidence. Imprecise presentation of evidence and knowledge In confidence level method, the uncertainty of evidence E is expressed as CF (E). In this system, 0 # CF(E) # 1. The general form of a rule is: IF E THEN H(CF(H,E)), wherein CF(H,E) is the level of confidence of this rule and represents the degree how conclusion H is true is supported when evidence E is known. In this system, 0 # CF(H,E) # 1.
Neurological Research
2014
VOL .
36
NO .
5
469
Chen et al.
Computer-aided diagnosis expert system for cerebrovascular diseases
Figure 1 Structure of expert diagnosis system for CVDs.
Imprecise algorithm of combined evidence For the conjunctive level of confidence of multiple pieces of evidence, the algorithm is expressed as: CF(E1‘E2‘…‘En) 5 Min{CF(E1),CF(E2),…,CF(En)} For the disjunctive level of confidence of multiple pieces of evidence, the algorithm is expressed as: CF(E1~E2~…~En) 5 Max{CF(E1),CF(E2),…,CF(En)} Transfer algorithm of uncertainty The level of confidence of conclusions based on levels of confidence of evidence and rules is the transfer algorithm of uncertainty, expressed as CF(H) 5 CF(H,E)CF(E), so 0 # CF(H) # 1. Synthetic algorithm of the same conclusion reached based on multiple pieces of independent evidence If the same conclusion is reached based on two different rules with different levels of confidence, the synthetic level of confidence may be calculated by synthetic algorithm. Hypothesizing the following two rules, the level of confidence of the corresponding conclusion H may be separately calculated: IF E1 THEN H (CF(H,E1)) 5 . CF1(H) 5 CF(H,E1)CF(E1) IF E2 THEN H (CF(H,E2)) 5 . CF2(H) 5 CF(H,E2)CF(E2) As 0 # CF1(H) # 1 and 0 # CF2(H) # 1, the synthetic level of confidence of conclusion H, CF1,2(H), is: CF1,2(H) 5 CF1(H)zCF2(H)–CF1(H)CF2(H) For the same conclusion reached based on multiple rules, the synthetic level of confidence may be calculated by synthesis of rules in pairs.
Knowledge base/inference engine design based on relation database Design of knowledge base The knowledge base is organized following production rules. It consists of a fact base and a rule base. The fact base is categorized into the following five groups: 1. symptom facts: sudden monolateral limb numbness, sudden or regularly changed headache, vomiting, etc.;
470
Neurological Research
2014
VOL .
36
NO .
5
2. 3. 4. 5.
sign facts: neck rigidity, Brudzinski sign, Kernig sign; highlight facts: onset in resting state, subsidence of symptoms within 1 hour, hypertension, etc.; intermediate facts: suspected cerebral apoplexy, meningeal irritation sign, etc.; conclusion facts: transient ischemic attack (TIA), cerebral infarction (cerebral thrombosis), cerebral infarction (cerebral embolism), cerebral hemorrhage, subarachnoidal hemorrhage.
The rule base consists of multiple rules. By conjunction (‘) or disjunction (~) of multiple rule antecedents (symptom facts, sign facts, highlight facts, intermediate facts), conclusions (intermediate facts and conclusion facts) are reached based on rules with levels of confidence. The logical order of rules directly decides the order of rule matching; rules of intermediate conclusions as antecedents of subsequent rules should always be arranged in the front. In the system, the knowledge base is expressed as a relation database, which consists of four tablets, a fact table, a rule antecedent table, a rule conclusion table and a case table. Correlations among tables are presented in Fig. 2. By means of correlation of the above four tablets, knowledge on any conjunctive or disjunctive rule with corresponding level of confidence may be expressed. The field ‘rule number’ in the rule conclusion table reflects the physical order of a rule in the rule base; the reasoning order of a rule in the knowledge base is expressed with the field ‘reasoning order’. The case table records patients’ general information. Design of inference engine/interpreter The inference engine performs imprecise reasoning following the controlling strategies of forward reasoning with knowledge in the knowledge base. Each symptom fact, sign fact or highlight fact is provided with a default level of confidence of 0. In a rule, if the antecedent is an opposite fact, the default level of confidence is 1. In concrete case diagnosis, the value of
Chen et al.
Computer-aided diagnosis expert system for cerebrovascular diseases
algorithm of uncertainty and saved in the rule base. If the same conclusion of different levels of confidence is reached with multiple matched rules, the synthetic algorithm of the same conclusion is inferred based on multiple pieces of independence evidence to calculate the combined level of confidence of the conclusion. In this way, when rules are used up, the diagnostic conclusion with the corresponding level of confidence will be obtained. If no diagnostic conclusion is obtained, the expert system will give a warning of insufficient fact selection. All matched and calculated rules are saved in the temporary table; when the user needs to interpret the reasoning process, they will be displayed and output as an interpreter.
System Implementation Figure 2 Correlation design of knowledge base.
the level of confidence is set for the corresponding fact. Levels of confidence of rules are pre-established by medical experts in the knowledge base. The inference engine performs calculation for multiple rules one by one following the logical reasoning order with the uncertainty algorithm of combined evidence. If the combined level of confidence of multiple antecedents is §0.5, it will be matched, and then the level of confidence of conclusion is calculated with the transfer
The expert system for auxiliary diagnosis of CVDs consists of a knowledge base management module, a reasoning/interpretation module, a case maintenance module and a help module. The main function interface of the system is presented in Fig. 3.
Knowledge base management module Management and maintenance of the knowledge base involve addition, deletion and revision of facts and rules as well as setting of logical reasoning orders of rules. The operation may be carried out through menus: [knowledge base]–.[facts]; [knowledge base]
Figure 3 Main interface of expert system.
Neurological Research
2014
VOL .
36
NO .
5
471
Chen et al.
Computer-aided diagnosis expert system for cerebrovascular diseases
Figure 4 Management interface of knowledge base.
–.[rules]. Management and maintenance functions of rule bases are presented in Fig. 4. When the user has selected and set specific symptoms and signs of a patient, diagnosis highlights and the levels of confidence and clicks the button ‘diagnosis’ or carries out operation through menus: [auxiliary diagnosis]–.[reasoning diagnosis], the system will perform forward imprecise reasoning and output the diagnostic results with the corresponding levels of confidence. To obtain interpretation of the reasoning process, the user may click the button ‘open interpreter’ or carry out operation through menus: [auxiliary diagnosis]–.[reasoning interpretation].
hemorrhage and all presented as large-area cerebral infarction; cerebral thrombosis in one patient was misdiagnosed as transient ischemic attack, cerebral hemorrhage in three as cerebral thrombosis, subarachnoidal hemorrhage in one and cerebral hemorrhage in three as cerebral embolism. The process of the expert diagnosis system making a diagnosis is depicted as follows. When a patient attacked a sudden severe headache, nausea and vomiting, stiff neck, and positive meningeal irritation, we input the above symptoms and signs, our expert diagnosis system will draw a conclusion as subarachnoid hemorrhage (CF 90%). The process of the expert diagnosis system making a diagnosis is presented in Fig. 4.
Case maintenance module
Discussion
The case maintenance module implements saving and enquiry operations of case information in the expert system and consists of input of basic information, case resetting, case saving and case base.
The expert diagnosis system for CVDs simulates the thinking and reasoning processes of experts as it pertains to diagnosing and treating diseases. The program’s design integrates professional medical knowledge from books, literature, diagnostic records, or neurologists. It also takes into account mastered procedures, individual characteristics, and even technical diagnostic skills to then construct a reasoning model that can be formalized with computers to mimic real-world conditions. The twenty-first century is an era of information and knowledge; computers have been applied in nearly every corner of social progress and human life. This expert diagnosis system for CVDs is a computer-based auxiliary diagnosis tool and is also a knowledge base for diagnosis of different types of apoplexy. It is independent of laboratory biochemical tests, radiological examinations, and can assess patients’ disease conditions based on patients’ symptoms and signs. Additionally, due to its local software design, there is no dependence upon network connections when using the system, thus it is suitable for deployment in developing regions. The system is expected to be an assistant for neurologists, apoplexy units, emergency doctors, and rural/community doctors.
Reasoning/interpretation module
Help module The help module includes operation instructions of the expert system and also medical knowledge and techniques about diagnosis and treatment of various CVDs.
System Testing An expert diagnosis system for CVDs was established and tested in 319 clinical patients including 253 with ischemic CVDs and 66 with hemorrhagic CVDs. Diagnosis accordance was obtained in 307 patients, and the diagnosis accordance rate was 96.2%. Involved were 223, 7, 23, 54 and 12 patients with cerebral thrombosis, cerebral embolism, transient ischemic attack, cerebral hemorrhage and subarachnoid hemorrhage, respectively; and diagnosis accordance was obtained in 219 (98.2%), 6 (85.7%), 23 (100%), 48 (88.9%) and 11 (91.7%), respectively. Cerebral thrombosis in three patients and cerebral embolism in one were misdiagnosed as cerebral
472
Neurological Research
2014
VOL .
36
NO .
5
Chen et al.
As previously alluded to, it is not intended to replace physicians but is to act as an auxiliary tool for improving the working efficiency of physicians by increasing accuracy and reliability – therefore increasing the speed at which definite diagnoses can be made. Additionally, it may assist in training interns and other healthcare providers to diagnose patients who present with specific diagnostic problems. Overall, by implementing the computer-aided diagnosis tool, routine diagnoses may be rapidly and accurately carried out so that healthcare worker’s time and energy may be saved for patient treatment and education. The expert system’s high reliability is manifested as a ruling out of objective and subjective factors such as fatigue, negligence, nervousness, and other external pressures and in doing so is able to consistently provide a stable, rational, complete, and rapid response. Furthermore, the knowledge and experience of multiple experts may be integrated within a single unit to enhance problem solving capabilities. Unlike human experts that will rest, retire, and eventually die, an expert knowledge system is persistent and may last illimitably – never constrained by spatial or temporal considerations. Acute cerebral apoplexy occurs due to ischemic CVDs in most cases. In this study, patients with ischemic CVDs accounted for 79.3% of the study population, corresponding to large-scale epidemiological investigations.15,16 This system potentiates diagnosis and classification of apoplexy based on gathered clinical information. The diagnostic accuracy rate of the MICROSTROKE system in 250 patients was 72.8% according to the Hamburg apoplexy database (Germany). In the study of Allen,17 clinical physicians’ diagnostic accuracy rate of cerebral infarction was 89%, and the diagnosis accuracy rate of cerebral hemorrhage was 55%. In the community-based study on the diagnosis of acute apoplexy and TIA carried out by Morgenstern et al.,18 the diagnostic accuracy rate of apoplexy and TIA was as high as 92%. With the current study’s system, diagnosis accordance rates for apoplexy, cerebral infarction, and cerebral hemorrhage were as high as 96.2, 97.8, and 88.9%, respectively – higher or similar to the diagnostic accuracy rates of clinical physicians and the MICROSTROKE system. It is important to emphasize that expert knowledge is the cornerstone of the entire expert system for diagnosis of CVDs.6 It must also be kept in mind, however, that this is a long-term and challenging task that requires frequent correction of mistakes when extracting professional knowledge from experts and transforming it into a usable model.19 Limitations of the database result in a misdiagnosis rate of 3.8%, especially in diagnosing cerebral hemorrhage and cerebral thrombosis. For example, due to difficulties in interpretation of clinical data, the system cannot distinguish patients with cerebral hemorrhage from those with cerebral infarction
Computer-aided diagnosis expert system for cerebrovascular diseases
who also have a low level of consciousness. In patients with an incomplete medical history, those with earlystage apoplexy, or those with deep unconsciousness inevitable difficulties are present to make it difficult for both man and machine to make an accurate diagnosis. Along with further testing of the system, additional case data may and should be used to enrich the system’s database in order to improve performance. In the end, in order to upgrade the system into a high-performance diagnostic tool there will undoubtedly be many difficulties to be encountered that are not visible to us at the moment. On the one hand, clinical diagnosis itself is inherently an imprecise reasoning activity. Each disease may have specific clinical symptoms, but those same symptoms may originate from various diseases. Additionally, the complexity of some stubborn diseases makes it very difficult to describe them using rigid rules or to express them with programming languages due to limitations of symbolic representation and expression methods. On the other hand, the expert system is based on rules that are as straightforward as they are rigid; with continued expansion of the rule base, the search pool will sharply enlarge and allow for currently unfathomable reasoning combinations. Yet although this is possible, such reasoning cycles do involve large amounts of invalid matching attempts and may waste large amounts of time in the process – causing the reasoning efficiency to be very low if not fully developed and continually honed. Currently existing problems include:20 1. Judgment errors are frequent. For a given target, an expert may adopt different methods under different circumstances in utilizing his/her abundant knowledge, a consideration that the expert system is markedly vulnerable to. If knowledge presented in the form of rules within its knowledge base does not cover the current condition (e.g. the knowledge is incomplete), expert system performance decreases significantly. 2. In the intelligent diagnosis system, knowledge is acquired mainly through interaction between knowledge engineers and experts in certain fields. As it stands, however, data acquisition methods are imperfect. Most important to keep in mind is that experts are not accustomed to expressing their professional knowledge in the form of rules or other programming standards and discrepancies in this regard can result in incomplete knowledge (e.g. certain conditions are ‘forgotten’, over-detailed, or over-generalized; important rules are neglected; different narration methods of different experts may lead to different understandings; etc.). Additionally, most experts formalize their problem solving strategies in various unique ways. 3. Experts’ solution to problems is a complicated creative thinking process and not usually a simple formalized reasoning process – ultimately the process thrives on ‘intuition’, ‘insight’, and ‘inspiration’. To foster development in the realest sense within this field, breakthroughs must be made in
Neurological Research
2014
VOL .
36
NO .
5
473
Chen et al.
4.
Computer-aided diagnosis expert system for cerebrovascular diseases
knowledge acquisition, knowledge presentation, and the capabilities of inference engines. The system cannot truly understand reasons and results contained in knowledge, and cannot determine problems that may be handled in an excessively biased and stereotyped way. Taken altogether, there is great potential for the expert system to neglect that medical diagnosis is a complicated process with particularities and extreme variability.
Measures to be taken for further improvement: 1. Diagnosis of CVDs involves huge amounts of information and data; in order to establish a relatively sound decision-making system, the knowledge base and inference engine needs to be further enriched and improved. The knowledge base should be continually renewed and enriched to avoid systematic errors, and expert physicians’ opinions should be continually matched. Studies should be carried out on how to optimize the inference engine to improve diagnostic efficiency and accuracy. 2. In using computer software for diagnosis of CVDs, standardization of disease symptoms is an important aspect, one which requires further scientific and clinical validation. 3. For more accurate diagnosis of CVDs, auxiliary examinations such as radiology and transcranial Doppler studies should be included into the diagnosis rules. A study carried out by Wang et al.21 showed that transcranial Doppler variables, when obtained from bilateral middle cerebral arteries, were independent predictors of 14-day intracerebral hemorrhage case fatality and would provide valuable prognostic information for future clinical decision-making. Another research study by Ma et al.22 presented that stroke magnetic resonance imaging provides relevant information that may allow the identification of optimal candidates for thrombolysis and also in recognizing patients who are not eligible. With the help of magnetic resonance angiograms, for example, patients who do not have initial vessel occlusion may not need thrombolysis in order to have a favorable clinical outcome. 4. It is necessary, in order to make the system as truly intelligent as possible, to decide how the system accomplishes automatic revision of diagnostic rules by analyzing feedback information after a definite consensus diagnosis is reached – including learning and monitoring of subsequent medical records; this will be a long-term research process. The system designed in this study is still a rudimentary computer-aided diagnosis system for CVDs and, accordingly, pertains to a simple medical diagnosis system. In regards to human diseases, however, diagnosis is a complicated process that cannot be optimally performed based solely on apparent symptoms; rather, other medical resources should be employed. Future pursuits will call for evaluation of the current system’s scientific rationality and its feasibility of its decision-making rules by experts across the spectrum of medical and computer science.
Disclaimer Statements Contributors Xu Chen is in charge of designing this project.
474
Neurological Research
2014
VOL .
36
NO .
5
Funding Shanghai Outstanding Discipline Open Fund. Conflicts of interest None. Ethics approval
References 1 Klar R, Zaiss A. Medical expert systems: design and applications in pulmonary medicine. Lung. 1990;168 suppl:1202–9. 2 Pereira MA, Schaefer MB, Marques JB. Remote expert system of support the prostate cancer diagnosis. Proceedings of the 26th Annual International Conference of the IEEE EMBS, San Francisco, CA, USA, 2004 September 1–5. pp. 3412–3415. 3 Beula Devamalar PM, Thulasi Bai V, Srivatsa SK. Design and architecture of real time web-centric tele health diabetes diagnosis expert system. Int J Med Eng Inform., 2009;1(3):307– 17. 4 Doi K. Current status and future potential of computer-aided diagnosis in medical imaging. Br J Radiol. 2005;78:S3–19. 5 Kakeda S, Korogi Y, Arimura H, Hirai T, Katsuragawa S, Aoki T, et al. Diagnostic accuracy and reading time to detect intracranial aneurysms on MR angiography using a computeraided diagnosis system. AJR Am J Roentgenol. 2008;190(2):459– 65. 6 Morris M, Wolfe WJ. An expert systerm for headache diagnosis: the computerized headache assessment tool. Headache. 2008;48(1):72–8. 7 Spitzer K, Thie A, Caplan LR, Kunze K. The TOPOSCOUT expert system for stroke localization. Stoke. 1989;20:1195–201. 8 Li L. Computer diagnostic system of cerebrovascular diseases. Shanxi Electr Technol. 2005;4:3–4. 9 Li FG, Ni ZW, Wang J, Gao L, Huang L. Design of an intelligent system for auxiliary diagnosis of cerebrovascular diseases with case-based reasoning. Chin Arch Tradit Chin Med. 2006;24(2):260–2. 10 Liu CZ, Song ZG. Study on expert diagnosis system for cerebrovascular diseases of Bp neural network. Comput Learn. 2008;4:13–4. 11 Ziegler V, Rashid A, Mu¨ller-Gorchs M, Kippnich U, Hiermann E, Ko¨gerl C, et al. Mobile computing systems in preclinical care of stroke. Results of the Stroke Angel initiative within the BMBF project PerCoMed. Anaesthesist. 2008;57(7):677–85. 12 Takao H, Murayama Y, Ishibashi T, Karagiozov KL, Abe T. A new support system using a mobile device (smartphone) for diagnostic image display and treatment of stroke. Stroke. 2012;43(1):236–9. 13 Pawlak Z. Rough set. Int J Comput Inform Sci. 1982;(11):341– 56. Spitzer K, Thie A, Caplan LR, Kunze K. The MICROSTROKE expert system for stroke type diagnosis. Stroke. 1989;20:1353–6. 14 Wu MF. Current situation and prospect of studies on rough set. J Wuyi Univ. 2002;16(2):16–21. 15 Besson G, Robert C, Hommel M, Perret J. Is it clinically possible to distinguish nonhemorrhagic infarct from Hemorrhagic stroke? Stoke. 1995;26(7):1205–9. 16 Kannel WB, Wolf PA. Epidemiology of cerebrovascular disease. In: Ross Russell RW, editor. Vascular disease of the central nervous system. London: Churchill Livingstone; 1983. 17 Allen CMC. Clinical diagnosis of the acute stroke syndrome. Q J Med. 1983;208:513–23. 18 Morgenstern LB, Lisabeth LD, Mecozzi AC, Smith MA, Longwell PJ, McFarling DA, et al. A population-based study of acute stroke and TIA diagnosis. Neurology. 2004;62(6):895– 900. 19 Torasso P. Medical diagnosis: Are expert system needed? Int J Biomed Comput. 1987;20:123–33. 20 Yang SZ, Ding H, Shi TL. Knowledge-based diagnostic reasoning. Beijing: Tsinghua University Press; 1993. pp. 11–12. 21 Wang W, Yang Z, Liu L, Dornbos D 3rd, Wang C, Song X, Gong X, et al. Relationship between transcranial Doppler variables in acute stage and outcome of intracerebral hemorrhage. Neurol Res. 2011;33(5):487–93. 22 Ma L, Gao PY, Lin Y, Xue J, Wang XC, Wang YJ, et al. Can baseline magnetic resonance angiography (MRA) status become a foremost factor in selecting optimal acute stroke patients for recombinant tissue plasminogen activator(rt-PA) thrombolysis beyond 3 hours? Neurol Res. 2009;31:355–61.
Computer-aided diagnosis expert system for cerebrovascular diseases Xu Chen*1, Zhijun Wang*2, Chrisopher Sy3, Xiaokun Liu1, Jinwu Qian2, Jia Zheng1, Zhiqiang Dong1, Limei Cao1, Xiang Geng1, Shuye Xu2, Xueyuan Liu4 1
Department of Neurology, Shanghai Eighth People’s Hospital, Shanghai, China, 2School of Mechatronics Engineering and Automation, Shanghai University, Shanghai, China, 3Department of Neurological Surgery, Wayne State University School of Medicine, Detroit, MI, USA, 4Department of Neurology, Shanghai Tenth People’s Hospital, Shanghai, China Objective: To establish an expert diagnosis system for cerebrovascular diseases (CVDs) and assess accuracy of the diagnosis system. Methods: An expert diagnosis system for CVDs was established and evaluated using actual clinical cases. Results: An expert diagnosis system for CVDs was established and tested in 319 clinical patients. Diagnosis accordance was obtained in 307 patients (the diagnosis accordance rate was 96.2%). Involved were 223, 7, 23, 54 and 12 patients with cerebral thrombosis, cerebral embolism, transient ischemic attack, cerebral hemorrhage and subarachnoid hemorrhage, respectively; and diagnosis accordance was obtained in 219 (98.2%), 6 (85.7%), 23 (100%), 48 (88.9%) and 11 (91.7%), respectively. Conclusion: Overall, the case analysis results support and demonstrate the diagnostic reasoning accuracy of the expert diagnosis system for CVDs. With the expert diagnosis system, medical experts’ diagnosis of CVDs can be effectively mimicked and auxiliary diagnosis of CVDs has been preliminarily realized, laying a foundation for increasing the diagnostic accuracy of clinical diagnoses as it pertains to CVDs.
Keywords: Cerebrovascular diseases, Diagnosis, Computer, Expert system
Introduction An ‘expert system’ is an intelligent computer program, one which endows a computer with select knowledge of human experts in a specific field so that it can mimic experts’ thought processes, reasoning, and judgment – allowing problems to be solved at ‘expert’ levels. Namely, an expert system is an intelligent appliance which has a repertoire of expertise-level knowledge in related fields and can use artificial intelligent techniques to simulate a human expert’s approach to a particular problem.1 Cerebrovascular diseases (CVDs) are common and frequent in middle-aged and older individuals in China. Patient morbidity has increased in incidence with a trend towards lower age groups in recent years – notably also becoming more salient in rural areas. According to a publication of the Ministry of Health in 2008, CVDs have become the top cause of death in Chinese people. The high incidence rate, high death rate, and high mutilation rate as a result of CVDs not only seriously impairs patient health and quality of life but also results in heavy medical, *These authors contributed equally to this work. Correspondence to: Xu Chen, Department of Neurology, Shanghai Eighth People’s Hospital, Shanghai 200235, China. Email: [email protected]; Xueyuan Liu, Department of Neurology, Tenth People’s Hospital, Shanghai 200072, China. Email: [email protected]; Shuye Xu, School of Mechatronics Engineering and Automation, Shanghai University, Shanghai 200444, China. Email: [email protected]
468
ß W. S. Maney & Son Ltd 2014 DOI 10.1179/1743132814Y.0000000347
economic, and social burdens to the nation and patients’ families. At present, Chinese people usually do not have much knowledge on the prevention and treatment CVDs. Additionally, the diagnosis and treatment capacities of hospitals from different regions and of different grades vary widely – largely due to diagnostic and therapeutic methods which are not normalized, leading to diminished therapeutic responses and poorer outcomes. Conventional diagnostic methods in foreign and domestic medicine are greatly influenced by subjective factors, and the accuracy of clinical diagnoses correlates closely to a doctor’s skills, knowledge, and experience. Due to the diagnostic complexity of CVDs and the necessity of giving a definite diagnosis, the subjectivity of clinical judgements face pressures not only to be correct but to be relatively quick. For example, we had a patient who was suddenly afflicted with hemiparalysis, aphasia, headache, dizziness, etc. and we could not make a definite diagnosis the first time. The most pressing issue for us was to differentiate whether it was ischemic stroke or hemorrhagic stroke – an example of a time-sensitive situation where definitive treatment cannot be started without a sure diagnosis. In some remote areas of China, diagnosis of these diseases is severely limited by doctors’ experience, lack of medical equipment, and inadequate facilities resulting in a diagnostic accuracy rate that is very low. Clearly, in such
Neurological Research
2014
VOL .
36
NO .
5
Chen et al.
areas – where abundant experience and professional knowledge can have a large impact – the need to support expert system technology is magnified. A powerful tool for timely and accurate treatment of patients with CVDs can be implemented through a systematic, institutionalized, and normalized computer platform with sufficient expert knowledge for diagnosis of CVDs; especially one with a software database that can be further enriched with continuous addition of more practical clinical cases. Expert diagnosis systems have already been used in some medical fields including oncology,2 endocrinology,3 and medical imageology,4 to name a few and where all have yielded favorable results. In neurology, expert diagnosis systems for cerebral aneurysms5 and headache6 have also been reported. Regarding the expert diagnosis system for CVDs, Spitzer et al.7 from Germany developed an expert system named MICROSTROKE in as early as 1989 for diagnosing different apoplexy types; in the same period, Spitzer et al.7 developed the TOPSCOUT expert system for positioning apoplexy and identifying distribution of affected vessels depending on patients’ symptoms and signs. In 2005, Li8 from Guangzhou University initiated the design of a computer diagnosis system for CVDs using a Visual Basic 6.0 object-oriented highlevel programming language; in 2006, Li et al.9 designed an intelligent case-based reasoning system for auxiliary diagnosis of CVDs; in 2008, Liu et al.10 carried out a study on an expert diagnosis system for CVDs using BP neural networks; in 2008, Ziegler et al.11 conducted a study on the role of mobile computing systems in the preclinical care of stroke – one which can offer new and innovative approaches in improving intersectional acute stroke care; in 2012, Takao et al.12 developed a system named i-Stroke for rapidly exchanging diagnostic images, clinical details, and management information through mobile devices (smart phones) and a Twitter system (e.g. transferring clinical data, CT, MR, angiographic, intraoperative images, and expert opinions in real time). Despite substantial research within the field, in general, there are currently no reports available in China about development of related software in regards to the clinical diagnosis of CVDs. As such, this study was intended to establish an expert diagnosis system for CVDs and to test it with verified clinical cases to assess its diagnosis accuracy.
Materials and Methods Hardware environment for system development Personal computers with .20 GB of free hard drive space and .256 MB of random access memory were used; the CPUs were P 1.2 GHz above.
Software environment for system development The operating systems used were Windows98/2000/ XP; supporting software appliances were Excel 2000 and Access 2000 (or higher versions of the two).
Computer-aided diagnosis expert system for cerebrovascular diseases
Programming language and version number, program size The software was developed in Microsoft Visual Basic 6.0; the program size was more than 800 line codes.
Fundamental principles for system design 1. rough set theory;13,14 2. diagnosis and classification of CVDs.
Cerebrovascular diseases in the system include ischemic CVDs and hemorrhage CVDs. Ischemic CVDs include cerebral infarction (cerebral thrombosis and cerebral embolism) and transient ischemic attack. Hemorrhagic CVDs include subarachnoid hemorrhage and cerebral hemorrhage.
Composition of the expert diagnosis system System structure The structure of the established expert diagnosis system for CVDs is presented in Fig. 1. It consists of six parts, i.e. a human-computer interface, an inference engine module, a knowledge base, a knowledge base management module, an interpretation module and a dynamic database. The knowledge base accommodates professional knowledge provided by medical experts in CVDs. The inference engine performs reasoning according to the knowledge and gives diagnostic conclusions; when a user selects and sets a patient’s specific symptoms, signs and diagnostic highlights and the levels of confidence, imprecise reasoning is performed following the controlling strategy of forward reasoning with knowledge in the knowledge base. The interpretation module interprets the reasoning process of the system. The dynamic database accommodates incipient facts provided by the user and such information as intermediate and final results obtained in operation of the system. The human/computer interface is an interface for interaction between the expert system and experts in the field, knowledge engineers and general users and also provides case maintenance and help. Users may input patients’ clinical symptoms, signs and past histories via the interface for user interaction.
Knowledge presentation, imprecise reasoning In the system, knowledge is presented with a classic production system and imprecise reasoning is performed based on the level of confidence so that each piece of evidence and each rule as well as diagnostic conclusions are presented with levels of confidence. Imprecise presentation of evidence and knowledge In confidence level method, the uncertainty of evidence E is expressed as CF (E). In this system, 0 # CF(E) # 1. The general form of a rule is: IF E THEN H(CF(H,E)), wherein CF(H,E) is the level of confidence of this rule and represents the degree how conclusion H is true is supported when evidence E is known. In this system, 0 # CF(H,E) # 1.
Neurological Research
2014
VOL .
36
NO .
5
469
Chen et al.
Computer-aided diagnosis expert system for cerebrovascular diseases
Figure 1 Structure of expert diagnosis system for CVDs.
Imprecise algorithm of combined evidence For the conjunctive level of confidence of multiple pieces of evidence, the algorithm is expressed as: CF(E1‘E2‘…‘En) 5 Min{CF(E1),CF(E2),…,CF(En)} For the disjunctive level of confidence of multiple pieces of evidence, the algorithm is expressed as: CF(E1~E2~…~En) 5 Max{CF(E1),CF(E2),…,CF(En)} Transfer algorithm of uncertainty The level of confidence of conclusions based on levels of confidence of evidence and rules is the transfer algorithm of uncertainty, expressed as CF(H) 5 CF(H,E)CF(E), so 0 # CF(H) # 1. Synthetic algorithm of the same conclusion reached based on multiple pieces of independent evidence If the same conclusion is reached based on two different rules with different levels of confidence, the synthetic level of confidence may be calculated by synthetic algorithm. Hypothesizing the following two rules, the level of confidence of the corresponding conclusion H may be separately calculated: IF E1 THEN H (CF(H,E1)) 5 . CF1(H) 5 CF(H,E1)CF(E1) IF E2 THEN H (CF(H,E2)) 5 . CF2(H) 5 CF(H,E2)CF(E2) As 0 # CF1(H) # 1 and 0 # CF2(H) # 1, the synthetic level of confidence of conclusion H, CF1,2(H), is: CF1,2(H) 5 CF1(H)zCF2(H)–CF1(H)CF2(H) For the same conclusion reached based on multiple rules, the synthetic level of confidence may be calculated by synthesis of rules in pairs.
Knowledge base/inference engine design based on relation database Design of knowledge base The knowledge base is organized following production rules. It consists of a fact base and a rule base. The fact base is categorized into the following five groups: 1. symptom facts: sudden monolateral limb numbness, sudden or regularly changed headache, vomiting, etc.;
470
Neurological Research
2014
VOL .
36
NO .
5
2. 3. 4. 5.
sign facts: neck rigidity, Brudzinski sign, Kernig sign; highlight facts: onset in resting state, subsidence of symptoms within 1 hour, hypertension, etc.; intermediate facts: suspected cerebral apoplexy, meningeal irritation sign, etc.; conclusion facts: transient ischemic attack (TIA), cerebral infarction (cerebral thrombosis), cerebral infarction (cerebral embolism), cerebral hemorrhage, subarachnoidal hemorrhage.
The rule base consists of multiple rules. By conjunction (‘) or disjunction (~) of multiple rule antecedents (symptom facts, sign facts, highlight facts, intermediate facts), conclusions (intermediate facts and conclusion facts) are reached based on rules with levels of confidence. The logical order of rules directly decides the order of rule matching; rules of intermediate conclusions as antecedents of subsequent rules should always be arranged in the front. In the system, the knowledge base is expressed as a relation database, which consists of four tablets, a fact table, a rule antecedent table, a rule conclusion table and a case table. Correlations among tables are presented in Fig. 2. By means of correlation of the above four tablets, knowledge on any conjunctive or disjunctive rule with corresponding level of confidence may be expressed. The field ‘rule number’ in the rule conclusion table reflects the physical order of a rule in the rule base; the reasoning order of a rule in the knowledge base is expressed with the field ‘reasoning order’. The case table records patients’ general information. Design of inference engine/interpreter The inference engine performs imprecise reasoning following the controlling strategies of forward reasoning with knowledge in the knowledge base. Each symptom fact, sign fact or highlight fact is provided with a default level of confidence of 0. In a rule, if the antecedent is an opposite fact, the default level of confidence is 1. In concrete case diagnosis, the value of
Chen et al.
Computer-aided diagnosis expert system for cerebrovascular diseases
algorithm of uncertainty and saved in the rule base. If the same conclusion of different levels of confidence is reached with multiple matched rules, the synthetic algorithm of the same conclusion is inferred based on multiple pieces of independence evidence to calculate the combined level of confidence of the conclusion. In this way, when rules are used up, the diagnostic conclusion with the corresponding level of confidence will be obtained. If no diagnostic conclusion is obtained, the expert system will give a warning of insufficient fact selection. All matched and calculated rules are saved in the temporary table; when the user needs to interpret the reasoning process, they will be displayed and output as an interpreter.
System Implementation Figure 2 Correlation design of knowledge base.
the level of confidence is set for the corresponding fact. Levels of confidence of rules are pre-established by medical experts in the knowledge base. The inference engine performs calculation for multiple rules one by one following the logical reasoning order with the uncertainty algorithm of combined evidence. If the combined level of confidence of multiple antecedents is §0.5, it will be matched, and then the level of confidence of conclusion is calculated with the transfer
The expert system for auxiliary diagnosis of CVDs consists of a knowledge base management module, a reasoning/interpretation module, a case maintenance module and a help module. The main function interface of the system is presented in Fig. 3.
Knowledge base management module Management and maintenance of the knowledge base involve addition, deletion and revision of facts and rules as well as setting of logical reasoning orders of rules. The operation may be carried out through menus: [knowledge base]–.[facts]; [knowledge base]
Figure 3 Main interface of expert system.
Neurological Research
2014
VOL .
36
NO .
5
471
Chen et al.
Computer-aided diagnosis expert system for cerebrovascular diseases
Figure 4 Management interface of knowledge base.
–.[rules]. Management and maintenance functions of rule bases are presented in Fig. 4. When the user has selected and set specific symptoms and signs of a patient, diagnosis highlights and the levels of confidence and clicks the button ‘diagnosis’ or carries out operation through menus: [auxiliary diagnosis]–.[reasoning diagnosis], the system will perform forward imprecise reasoning and output the diagnostic results with the corresponding levels of confidence. To obtain interpretation of the reasoning process, the user may click the button ‘open interpreter’ or carry out operation through menus: [auxiliary diagnosis]–.[reasoning interpretation].
hemorrhage and all presented as large-area cerebral infarction; cerebral thrombosis in one patient was misdiagnosed as transient ischemic attack, cerebral hemorrhage in three as cerebral thrombosis, subarachnoidal hemorrhage in one and cerebral hemorrhage in three as cerebral embolism. The process of the expert diagnosis system making a diagnosis is depicted as follows. When a patient attacked a sudden severe headache, nausea and vomiting, stiff neck, and positive meningeal irritation, we input the above symptoms and signs, our expert diagnosis system will draw a conclusion as subarachnoid hemorrhage (CF 90%). The process of the expert diagnosis system making a diagnosis is presented in Fig. 4.
Case maintenance module
Discussion
The case maintenance module implements saving and enquiry operations of case information in the expert system and consists of input of basic information, case resetting, case saving and case base.
The expert diagnosis system for CVDs simulates the thinking and reasoning processes of experts as it pertains to diagnosing and treating diseases. The program’s design integrates professional medical knowledge from books, literature, diagnostic records, or neurologists. It also takes into account mastered procedures, individual characteristics, and even technical diagnostic skills to then construct a reasoning model that can be formalized with computers to mimic real-world conditions. The twenty-first century is an era of information and knowledge; computers have been applied in nearly every corner of social progress and human life. This expert diagnosis system for CVDs is a computer-based auxiliary diagnosis tool and is also a knowledge base for diagnosis of different types of apoplexy. It is independent of laboratory biochemical tests, radiological examinations, and can assess patients’ disease conditions based on patients’ symptoms and signs. Additionally, due to its local software design, there is no dependence upon network connections when using the system, thus it is suitable for deployment in developing regions. The system is expected to be an assistant for neurologists, apoplexy units, emergency doctors, and rural/community doctors.
Reasoning/interpretation module
Help module The help module includes operation instructions of the expert system and also medical knowledge and techniques about diagnosis and treatment of various CVDs.
System Testing An expert diagnosis system for CVDs was established and tested in 319 clinical patients including 253 with ischemic CVDs and 66 with hemorrhagic CVDs. Diagnosis accordance was obtained in 307 patients, and the diagnosis accordance rate was 96.2%. Involved were 223, 7, 23, 54 and 12 patients with cerebral thrombosis, cerebral embolism, transient ischemic attack, cerebral hemorrhage and subarachnoid hemorrhage, respectively; and diagnosis accordance was obtained in 219 (98.2%), 6 (85.7%), 23 (100%), 48 (88.9%) and 11 (91.7%), respectively. Cerebral thrombosis in three patients and cerebral embolism in one were misdiagnosed as cerebral
472
Neurological Research
2014
VOL .
36
NO .
5
Chen et al.
As previously alluded to, it is not intended to replace physicians but is to act as an auxiliary tool for improving the working efficiency of physicians by increasing accuracy and reliability – therefore increasing the speed at which definite diagnoses can be made. Additionally, it may assist in training interns and other healthcare providers to diagnose patients who present with specific diagnostic problems. Overall, by implementing the computer-aided diagnosis tool, routine diagnoses may be rapidly and accurately carried out so that healthcare worker’s time and energy may be saved for patient treatment and education. The expert system’s high reliability is manifested as a ruling out of objective and subjective factors such as fatigue, negligence, nervousness, and other external pressures and in doing so is able to consistently provide a stable, rational, complete, and rapid response. Furthermore, the knowledge and experience of multiple experts may be integrated within a single unit to enhance problem solving capabilities. Unlike human experts that will rest, retire, and eventually die, an expert knowledge system is persistent and may last illimitably – never constrained by spatial or temporal considerations. Acute cerebral apoplexy occurs due to ischemic CVDs in most cases. In this study, patients with ischemic CVDs accounted for 79.3% of the study population, corresponding to large-scale epidemiological investigations.15,16 This system potentiates diagnosis and classification of apoplexy based on gathered clinical information. The diagnostic accuracy rate of the MICROSTROKE system in 250 patients was 72.8% according to the Hamburg apoplexy database (Germany). In the study of Allen,17 clinical physicians’ diagnostic accuracy rate of cerebral infarction was 89%, and the diagnosis accuracy rate of cerebral hemorrhage was 55%. In the community-based study on the diagnosis of acute apoplexy and TIA carried out by Morgenstern et al.,18 the diagnostic accuracy rate of apoplexy and TIA was as high as 92%. With the current study’s system, diagnosis accordance rates for apoplexy, cerebral infarction, and cerebral hemorrhage were as high as 96.2, 97.8, and 88.9%, respectively – higher or similar to the diagnostic accuracy rates of clinical physicians and the MICROSTROKE system. It is important to emphasize that expert knowledge is the cornerstone of the entire expert system for diagnosis of CVDs.6 It must also be kept in mind, however, that this is a long-term and challenging task that requires frequent correction of mistakes when extracting professional knowledge from experts and transforming it into a usable model.19 Limitations of the database result in a misdiagnosis rate of 3.8%, especially in diagnosing cerebral hemorrhage and cerebral thrombosis. For example, due to difficulties in interpretation of clinical data, the system cannot distinguish patients with cerebral hemorrhage from those with cerebral infarction
Computer-aided diagnosis expert system for cerebrovascular diseases
who also have a low level of consciousness. In patients with an incomplete medical history, those with earlystage apoplexy, or those with deep unconsciousness inevitable difficulties are present to make it difficult for both man and machine to make an accurate diagnosis. Along with further testing of the system, additional case data may and should be used to enrich the system’s database in order to improve performance. In the end, in order to upgrade the system into a high-performance diagnostic tool there will undoubtedly be many difficulties to be encountered that are not visible to us at the moment. On the one hand, clinical diagnosis itself is inherently an imprecise reasoning activity. Each disease may have specific clinical symptoms, but those same symptoms may originate from various diseases. Additionally, the complexity of some stubborn diseases makes it very difficult to describe them using rigid rules or to express them with programming languages due to limitations of symbolic representation and expression methods. On the other hand, the expert system is based on rules that are as straightforward as they are rigid; with continued expansion of the rule base, the search pool will sharply enlarge and allow for currently unfathomable reasoning combinations. Yet although this is possible, such reasoning cycles do involve large amounts of invalid matching attempts and may waste large amounts of time in the process – causing the reasoning efficiency to be very low if not fully developed and continually honed. Currently existing problems include:20 1. Judgment errors are frequent. For a given target, an expert may adopt different methods under different circumstances in utilizing his/her abundant knowledge, a consideration that the expert system is markedly vulnerable to. If knowledge presented in the form of rules within its knowledge base does not cover the current condition (e.g. the knowledge is incomplete), expert system performance decreases significantly. 2. In the intelligent diagnosis system, knowledge is acquired mainly through interaction between knowledge engineers and experts in certain fields. As it stands, however, data acquisition methods are imperfect. Most important to keep in mind is that experts are not accustomed to expressing their professional knowledge in the form of rules or other programming standards and discrepancies in this regard can result in incomplete knowledge (e.g. certain conditions are ‘forgotten’, over-detailed, or over-generalized; important rules are neglected; different narration methods of different experts may lead to different understandings; etc.). Additionally, most experts formalize their problem solving strategies in various unique ways. 3. Experts’ solution to problems is a complicated creative thinking process and not usually a simple formalized reasoning process – ultimately the process thrives on ‘intuition’, ‘insight’, and ‘inspiration’. To foster development in the realest sense within this field, breakthroughs must be made in
Neurological Research
2014
VOL .
36
NO .
5
473
Chen et al.
4.
Computer-aided diagnosis expert system for cerebrovascular diseases
knowledge acquisition, knowledge presentation, and the capabilities of inference engines. The system cannot truly understand reasons and results contained in knowledge, and cannot determine problems that may be handled in an excessively biased and stereotyped way. Taken altogether, there is great potential for the expert system to neglect that medical diagnosis is a complicated process with particularities and extreme variability.
Measures to be taken for further improvement: 1. Diagnosis of CVDs involves huge amounts of information and data; in order to establish a relatively sound decision-making system, the knowledge base and inference engine needs to be further enriched and improved. The knowledge base should be continually renewed and enriched to avoid systematic errors, and expert physicians’ opinions should be continually matched. Studies should be carried out on how to optimize the inference engine to improve diagnostic efficiency and accuracy. 2. In using computer software for diagnosis of CVDs, standardization of disease symptoms is an important aspect, one which requires further scientific and clinical validation. 3. For more accurate diagnosis of CVDs, auxiliary examinations such as radiology and transcranial Doppler studies should be included into the diagnosis rules. A study carried out by Wang et al.21 showed that transcranial Doppler variables, when obtained from bilateral middle cerebral arteries, were independent predictors of 14-day intracerebral hemorrhage case fatality and would provide valuable prognostic information for future clinical decision-making. Another research study by Ma et al.22 presented that stroke magnetic resonance imaging provides relevant information that may allow the identification of optimal candidates for thrombolysis and also in recognizing patients who are not eligible. With the help of magnetic resonance angiograms, for example, patients who do not have initial vessel occlusion may not need thrombolysis in order to have a favorable clinical outcome. 4. It is necessary, in order to make the system as truly intelligent as possible, to decide how the system accomplishes automatic revision of diagnostic rules by analyzing feedback information after a definite consensus diagnosis is reached – including learning and monitoring of subsequent medical records; this will be a long-term research process. The system designed in this study is still a rudimentary computer-aided diagnosis system for CVDs and, accordingly, pertains to a simple medical diagnosis system. In regards to human diseases, however, diagnosis is a complicated process that cannot be optimally performed based solely on apparent symptoms; rather, other medical resources should be employed. Future pursuits will call for evaluation of the current system’s scientific rationality and its feasibility of its decision-making rules by experts across the spectrum of medical and computer science.
Disclaimer Statements Contributors Xu Chen is in charge of designing this project.
474
Neurological Research
2014
VOL .
36
NO .
5
Funding Shanghai Outstanding Discipline Open Fund. Conflicts of interest None. Ethics approval
References 1 Klar R, Zaiss A. Medical expert systems: design and applications in pulmonary medicine. Lung. 1990;168 suppl:1202–9. 2 Pereira MA, Schaefer MB, Marques JB. Remote expert system of support the prostate cancer diagnosis. Proceedings of the 26th Annual International Conference of the IEEE EMBS, San Francisco, CA, USA, 2004 September 1–5. pp. 3412–3415. 3 Beula Devamalar PM, Thulasi Bai V, Srivatsa SK. Design and architecture of real time web-centric tele health diabetes diagnosis expert system. Int J Med Eng Inform., 2009;1(3):307– 17. 4 Doi K. Current status and future potential of computer-aided diagnosis in medical imaging. Br J Radiol. 2005;78:S3–19. 5 Kakeda S, Korogi Y, Arimura H, Hirai T, Katsuragawa S, Aoki T, et al. Diagnostic accuracy and reading time to detect intracranial aneurysms on MR angiography using a computeraided diagnosis system. AJR Am J Roentgenol. 2008;190(2):459– 65. 6 Morris M, Wolfe WJ. An expert systerm for headache diagnosis: the computerized headache assessment tool. Headache. 2008;48(1):72–8. 7 Spitzer K, Thie A, Caplan LR, Kunze K. The TOPOSCOUT expert system for stroke localization. Stoke. 1989;20:1195–201. 8 Li L. Computer diagnostic system of cerebrovascular diseases. Shanxi Electr Technol. 2005;4:3–4. 9 Li FG, Ni ZW, Wang J, Gao L, Huang L. Design of an intelligent system for auxiliary diagnosis of cerebrovascular diseases with case-based reasoning. Chin Arch Tradit Chin Med. 2006;24(2):260–2. 10 Liu CZ, Song ZG. Study on expert diagnosis system for cerebrovascular diseases of Bp neural network. Comput Learn. 2008;4:13–4. 11 Ziegler V, Rashid A, Mu¨ller-Gorchs M, Kippnich U, Hiermann E, Ko¨gerl C, et al. Mobile computing systems in preclinical care of stroke. Results of the Stroke Angel initiative within the BMBF project PerCoMed. Anaesthesist. 2008;57(7):677–85. 12 Takao H, Murayama Y, Ishibashi T, Karagiozov KL, Abe T. A new support system using a mobile device (smartphone) for diagnostic image display and treatment of stroke. Stroke. 2012;43(1):236–9. 13 Pawlak Z. Rough set. Int J Comput Inform Sci. 1982;(11):341– 56. Spitzer K, Thie A, Caplan LR, Kunze K. The MICROSTROKE expert system for stroke type diagnosis. Stroke. 1989;20:1353–6. 14 Wu MF. Current situation and prospect of studies on rough set. J Wuyi Univ. 2002;16(2):16–21. 15 Besson G, Robert C, Hommel M, Perret J. Is it clinically possible to distinguish nonhemorrhagic infarct from Hemorrhagic stroke? Stoke. 1995;26(7):1205–9. 16 Kannel WB, Wolf PA. Epidemiology of cerebrovascular disease. In: Ross Russell RW, editor. Vascular disease of the central nervous system. London: Churchill Livingstone; 1983. 17 Allen CMC. Clinical diagnosis of the acute stroke syndrome. Q J Med. 1983;208:513–23. 18 Morgenstern LB, Lisabeth LD, Mecozzi AC, Smith MA, Longwell PJ, McFarling DA, et al. A population-based study of acute stroke and TIA diagnosis. Neurology. 2004;62(6):895– 900. 19 Torasso P. Medical diagnosis: Are expert system needed? Int J Biomed Comput. 1987;20:123–33. 20 Yang SZ, Ding H, Shi TL. Knowledge-based diagnostic reasoning. Beijing: Tsinghua University Press; 1993. pp. 11–12. 21 Wang W, Yang Z, Liu L, Dornbos D 3rd, Wang C, Song X, Gong X, et al. Relationship between transcranial Doppler variables in acute stage and outcome of intracerebral hemorrhage. Neurol Res. 2011;33(5):487–93. 22 Ma L, Gao PY, Lin Y, Xue J, Wang XC, Wang YJ, et al. Can baseline magnetic resonance angiography (MRA) status become a foremost factor in selecting optimal acute stroke patients for recombinant tissue plasminogen activator(rt-PA) thrombolysis beyond 3 hours? Neurol Res. 2009;31:355–61.
