Supporting meningitis diagnosis amongst infants and children through the use of fuzzy cognitive mapping

Meningitis is defined as an inflammation of the membranes and cerebrospinal fluid that encases and bathes the brain and spinal cord. It is a serious disease which can be life-threatening and may result in permanent complications if not diagnosed and treated early. The pathogenic development of the disease suggests that meningitis can be broadly categorized into three main types [1]. Bacterial meningitis, which is rare, but more serious and can be life-threatening if not treated immediately. Fungal meningitis is typically diagnosed in patients with pre-existing conditions that have a weakened immune system, such as those living with lupus or HIV. Viral meningitis is caused by a virus (can be acute or chronic), is more common, but is far less serious and those who are diagnosed usually make a full recovery.

Symptoms of meningitis amongst children can appear very quickly or may take several days to make themselves known and include: fever; irritability; headache; photophobia (eye sensitivity to light); stiff neck; skin rashes; jaundice; inability to feed; high pitched cry; lethargy; seizures. Early diagnosis and timely interventions are the most effective ways for preventing negative outcomes associated with the disease.

Whilst meningitis cases affect all age demographics, the World Health Organisation has observed the highest rates of infection in young children [2]. For example, bacterial meningitis predominantly affects younger children and most cases of viral meningitis occur in children under the age of five years [3]. Epidemiological studies suggest rates of about two to ten cases per 10,000 live births with children particularly vulnerable to meningitis between the ages of 3 months and 3 years [4]. Fatality rates vary from as low as 2% for infants to 20 - 30% for neonates and adults. Since the mid-1980s, as a result of the protection offered by current vaccines and an increased understanding of the mechanisms of the disease [5], the median age at which bacterial meningitis is diagnosed has shifted from 15 months to 25 years. Geographically, meningitis epidemics have been experienced in various parts of the world, with research suggesting that climate might be a contributory risk factor in the spread of the disease [6].

In addition to the symptomatic development and epidemiological spread of the disease, there are other known risk factors associated with meningitis which include social, environmental and economic determinants. Although most cases are isolated, the disease can spread amongst people living in close social proximity, and outbreaks have occurred in those areas where there is a higher degree of social interaction or in areas experiencing overcrowding [7] which promotes exposure and transmission. The research indicates that meningitis in more prevalent in poorer areas then in affluent areas, suggesting that there is also a strong socio-economic component to the development of the disease [8]. Indeed the risk of invasive meningococcal disease (leading cause of bacterial meningitis) in children is strongly influenced by unfavorable socioeconomic conditions [9]. Increased levels of poverty are also linked to identified barriers in terms of geography, income, and socio-cultural differences. Research has found that presenting for treatment and early management of the disease is compounded by issues related to geography (access to medical facility), income (cost of healthcare), or cultural differences (attitudes towards illness and disease) which prevent lower socio-economic groups from receiving treatment, increasing the risk of adverse outcomes. Others have suggested that improvements in access to healthcare and earlier treatment are more likely to reduce the rate of mortality from meningitis [10].

Physicians are confronted with a broad range of symptoms and risk factors which they need to take into account when assessing a patient with possible meningitis, and when establishing the consequences of various treatment options. The ways in which these symptoms and risk factors inter-relate and how they are identified by healthcare professionals are integral to improving outcomes from the disease.

At a macro level, a number of studies have shown that the diagnosis and treatment management of meningitis is a complex and challenging problem for government and healthcare agencies requiring novel approaches to its management and intervention [11‐14]. This has involved the application of modelling approaches for diagnosis and treatment. Public health experts working at the health protection agencies have developed a model to determine if suspected meningitis is bacterial or viral in origin. Clinical prediction rules have also been used to develop bacterial meningitis scores that classify patients according to risk of contraction [15]. Some diagnostic decision rules for management of children with meningeal signs have also been proposed to assist in timely diagnosis and decision-making [13, 16]. Diagnostic scores have been constructed to predict disease outcomes and have been applied to successfully identify at-risk patients [17, 18]. Based on literature studies, the symptoms, clinical features and microbiological (lab) examinations are the principal factors contributing to the accurate diagnosis and risk assessment of meningitis.

In this paper, we are proposing a modelling approach to understanding meningitis which focuses on capturing the various symptoms associated with the disease, incorporating specific risk factors such as socio-economic determinants as derived from expert knowledge provided by physicians. This work models the complex problem of meningitis diagnosis and severity assessment using Fuzzy cognitive mapping (FCM), which is an effective knowledge representation and modelling technique [19]. Through the proposed technique, the paper will develop and validate a simple tool to predict the likelihood of viral or bacterial meningitis in younger infants and children.

The main scope of this work is the construction of a knowledge based tool for modelling meningitis diagnosis for children living in semi-urban areas of India. The meningitis diagnostic procedure typically involves close interaction between the biologist, pathologist and the paediatrician and involves extracting and analyzing blood samples from the patient. The diagnosis of meningitis is more challenging within semi-urban areas of Indian cities given the lack of healthcare infrastructure, co-ordination between healthcare agencies and professionals and the shortage of qualified physicians which potentially delay identification of the disease. Moreover, the average costs of laboratory tests and potentially long hospital stays as a result, make treatment expensive and unaffordable for the majority of patients living within developing countries [20].

A decision-making tool to assist in the diagnosis of meningitis provides the potential for healthcare professionals to arrive at a decision sooner and alleviates the cost burden to the patient if laboratory tests and hospital stays are not required. No previous research has explored FCM methodology for assessing and diagnosing meningitis. The tool proposed in this research is designed to aid paediatricians who are responsible for clinical decision-making regarding the treatment of children with meningitis which involves: diagnosing the disease and its severity and making decisions regarding the most appropriate treatment.

This paper is structured into five sections. The section on Methods briefly describes the principal aspects of FCM formalization and describes the construction of a tool to support the diagnosis of meningitis. The Results describes the accuracy of the tool in predicting the diagnosis of meningitis. Finally the Discussion and Conclusions emerging from the study are presented.

After the construction of the FCM tool, the system was refined to perform with higher levels of accuracy. This has been achieved by dividing the dataset into two parts: one for training the system to imitate real world decision-making and other for testing the system against clinical decisions regarding the diagnosis of meningitis. Forty cases, including twenty patients diagnosed with meningitis and twenty undiagnosed with meningitis, were used to train the system, and the remaining sixteen cases were used to establish the accuracy of the system. During training as well as testing, the information of patient cases were used to formulate the initial activation vector A, discussed in subsection Main Aspects of Fuzzy Cognitive Maps, which is then further processed using Equation 1. In other words, each patient state is represented as a vector of 19 concepts, where each concept is a sign or symptom or risk factor. Each one concept takes an initial value (activated value); if for example the Brudzinski sign is on, the concept takes the value 1, or in the case of “OFF”, it takes the value 0. The paediatrician can provide any value between 0 and 1. Then a concept vector is produced with the concept values and it is used in the FCM simulation algorithm (presented in [49]). A new (final) concept vector is produced after the system convergence (actually a system’s equilibrium point). The final value of the decision node Mn is the value which is presented in Additional file 1 and assessed for the system decision. The system converges to a different final state, if the initial patient conditions are different. An analytical description of how the values of FCM status change in relation to patient status is presented in previous works of Papageorgiou 2011 [53].

The efficacy of FCM reasoning is determined by the ability of the system to accurately determine the opinions of the paediatricians. The proposed system is used to predict the possibility of meningitis diagnosis in all training cases with different (λ) values of the threshold function defined in Equation 2. Authors of [54] suggest that an appropriate value for λ, used in sigmoid transformation function, be determined whilst training the system. Thus, during training phase, with different λ values, the system can achieve the optimal value for an efficient sigmoid threshold function used in FCM reasoning. For the system described here, the accuracy obtained for λ = {0.9,0.8,0.7,0.6,0.5,0.45,0.4,0.3,0.2,0.1} is reported in Additional file 1. The accuracy has been calculated using Equation 3.

Five representative cases are reported in Table 3. Each row of the table includes the information for the patient, the decision taken by the pediatrician and the result produced by the system under various λ values. It was observed that the predictions made by the system were most accurate with the value of λ being 0.3.

Cases 1 and 5 in Table 3 are the cases where a value of λ above 0.5 suggests that meningitis is present but in practice a treatment for meningitis disease is not recommended. This is supported by values of λ at 0.45, 0.4, and 0.3. The decisions of the paediatrician coincide with that of the system at λ = 0.3 for most of the examined patient cases i.e., 38/40.

It is imperative to understand the significance of selection of λ value. If λ value is close to 0, it converges very slowly but may produce results with high levels of accuracy; and if the value is near 1, it converges fast, but compromises accuracy. So, there is a trade-off between accuracy and the time of convergence which is dependent on the data set and the system under consideration. Our analysis shows that the system is performing accurately when λ = 0.3 and converges in approximately 20 iterations. The results of fine tuning are shown in Table 4.

This experimentation implies that the transformation function has to be adjusted to produce high levels of accuracy. Authors of [54] suggested that “..sigmoid function can be considered an excellent decision support tool within any scope” and we have been able to fine tune this function by adjusting the λ value as per our data set.

After fine tuning our system, we conducted the experiment on the remaining data set which consists of 16 patient cases. Out of these, 6 were diagnosed and treated for meningitis disease by the pediatricians. The statistical information on the performance of our system is provided below:

These results show that the system has a reasonably high level of accuracy and pediatricians can rely on the system in their clinical practices. The system is implemented in MATLAB R2012A, and provides a graphical user interface to the user which could be provided as a piece of fully functional software for pediatricians to use.

A key advantage of the FCM methodology used within this paper is that it can provide insight into the complexity of a specific problem through highlighting key concepts and feedbacks in the system which might otherwise remain unidentified. By underpinning the model development with the opinion of expert users, FCM is also capable of representing a system in a form that corresponds closely to the way humans perceive it. Therefore, the model is easily understandable, even by a non-technical audience, as each parameter can be easily interpreted within the context of the system as a whole. The FCM modeling technique also has an inherent flexibility and can be easily altered to incorporate new phenomenon, and if the behavior of the system operates differently than expected, FCM is more amenable to the modification of factors within the model.

The principal objective of the study described in this paper was to provide a decision making tool to the physicians working within infrastructural and economical constraints i.e., within a real world environment where solutions are sought to address external limitations that cause inefficiencies in the system. In this case the prohibitive costs to the patient of undergoing laboratory tests to detect meningitis and costly delays in the system which increase the likelihood of poorer outcomes. Even though the system has been able to achieve the desired functionality, there are a number of limitations of this modelling approach. Firstly, the experts who helped in designing the model assumed that the concept of high economic status is directly linked with achieving a good nutritional diet intake. This assumption is based on the prevalent socioeconomic conditions that the experts have witnessed within semi-urban India, which may vary over time and when applying the model across different geographical contexts [55]. Secondly, this study has incorporated 19 significant risk factors and signs/symptoms associated with meningitis associated in predicting and diagnosing meningitis, yet it is recognized that there are further factors which were not documented by experts but which are important in the diagnosis of the condition. Further work needs to be undertaken with experts-physicians in enhancing the proposed FCM model by adding more concepts and identifying the potential relationships among them, particularly across different socio-cultural contexts where risk factors and symptoms may vary. Thirdly, further work needs to be undertaken to clearly elucidate and comprehensively map the risk factors and symptoms associated with different forms of meningitis including baterial, viral and fungal strains of the disease. Lastly, the model described does not incorporate a temporal dimension, which is necessary to establish the urgency of treatment should the patient be diagnosed with meningitis i.e., urgent or non-urgent and should consider utilizing other modelling techniques.

This study presents the results from research which sought to model expert users knowledge within a FCM decision support system which accurately diagnoses meningitis amongst infants and children. More specifically, this work proposes the application of a decision support tool based on the soft methodology of FCM to diagnose meningitis within specific individual cases. The developed software tool makes a decision based upon the state and condition of the patient as observed by the physician and therefore does not need access to specific patient data to determine the possibility of meningitis. The model was tested against a number of real patient cases demonstrating its capability as a dynamic decision making tool which could be used within clinical practice. Moreover, whilst the tool is designed to address cost and efficiency savings, it has the potential to be used not only within the clinical settings but as a mobile tool for application within more rural areas where access to healthcare facilities are more problematic. Further work will be undertaken by the authors to extend the FCM model described here to more clearly elucidate the breadth and depth of risk factors and symptoms associated with the disease amongst infants and young children, whilst applying the model across different geographical areas and increased numbers of patient cases to further validate the model.

VKM and ELP developed the FCM model used in the diagnosis of meningitis among children and infants, and drafted the manuscript. RM provided the domain expertise and drafted the introductory section. RW assisted in manuscript development. All authors read and approved the final manuscript.