The process of early diagnosis in terms of determining the disease risks is considered. The relevance of the process of medical cards screening as one of the important task of maintaining health of the population of the country is substantiated. An analytical review of existing Data Mining methods for processing of medical data for solving the problem of disease risk assessment is conducted. The choice of Case Based Reasoning approach for searching the relevant medical cards is confirmed. The assessment based on the use of the k-Nearest Neighbors method is proposed. The functional model of the formation of the base of precedents in IDEF0 notation has been developed. A relevant case retrieving model based on the use of the Gower coefficient as a convolution criterion for mixed data has been proposed. Numerical studies of the task have been carried out. The process of early diagnosis using the developed methodology has been improved. Recommendations for assessing the effectiveness of using the proposed methodology in practice based on calculating the values of the confusion matrix have been developed. Screening of medical records, early diagnosis of diseases, Case Based Reasoning, k-Nearest Neighbors method, similarity measure for mixed data, Gower coefficient ORCID: 0000-0001-9642-5414 (K. Melnyk); 0000-0002-8834-2536 (N. Borysova); 0000-0003-3893-117X (S. Ershova)
Many countries have been seen a decline in economic growth recently. One of the reasons is the
labor force shrinking. Labor force or the labor market activity rate is the amount of active persons
(occupied labor force and the unemployed) as a percent of the total population [
2020 Copyright for this paper by its authors. the risk of wrong medical decisions. Officials need to develop mechanisms to monitor the organization of the health prevention and early diagnosis process.
In general, the procedure of early diagnosis or screening is a method of conducting simple and safe
researches of large groups of the population. It can be used in many fields of medicine to determine
and assess the likelihood of a disease. Therefore, it is a procedure for the disease risk assessment. To
date, in evidence-based medicine, there are proven tools for determining the risk of many diseases, for
example, the Framingham scale of 10-year risk can be used for assessing the risks of cardiovascular
disease (CVD) [
Thus, the purpose of this work is to develop the methodology of risk disease assessment based on data from medical records.
According to the current medical and technological documents and the WHO risk assessment
guidelines [
Let is a set of medical records from some clinic database. Let is a
set of risk assessment groups based on scale from [
To solve this task, it is necessary to: develop the methodology for risk assessment; choose a classification method for medical records; determine a class or appropriate risk group.
Let’s consider the most commonly used Data Mining methods of analysis of medical data for solving the problem of disease risk assessment in terms of the distribution of medical records into separate groups with different risk values.
1. Neural networks.
Using of neural networks in practice for solving the classification task of medical data for linearly
separable classes gives good results [
2. Bayesian networks.
The mathematical apparatus of Bayesian belief network allows to develop a model for assessing the disease risk for each group of diseases separately [16, 17]. The vertices in the directed graph can be the following: bad habits, bad heredity, diseases and their symptoms, test results. The edges in the probability-oriented graph reflect the conditional dependence of one vertex to another. That is, Bayesian networks allow to process mixed data, which is the advantage of this approach. With sufficient statistical information on patients, this approach gives good results in disease risk assessment. Disadvantages include the ambiguity of knowledge about the relationship between the selected input data and the complexity of network creation.
3. Theory of artificial intelligence, method of comparative identification.
The task of disease risk assessment can be solved with the help of the theory of intelligence proposed by school of Yu. P. Shabanov-Kushnarenko, namely the method of comparative identification [18-19]. This approach allows modeling the doctor’s reasoning based on the use of finite predicates algebra. In this case, the domain area should be represented as a logical network. The vertices in the network are predicate variables; the edges are the values of these variables. Predicate variables are information from medical records and other sources of medical data. This information can be in the form of qualitative, quantitative and categorical data. The disadvantage of this approach is the complex process of data preprocessing in compiling a risk assessment model, because the expert has to present all possible combinations of all values of predicate variables of the domain area in the form of system of predicate equations.
4. Decision trees.
The domain model is presented in the form of a hierarchical structure or a tree [
Case Based Reasoning (CBR) is an approach that help to make a decision based on the previous
cases and experience [
Comparing Data Mining approaches to determine the disease risk group, this study proposes to choose Case Based Reasoning as a method for classifying medical records.
Let’s consider the usage of Case-Based Reasoning for resolving the disease risk evaluation task. This approach allows making a decision for new cases using previously solved problems and reusing their solutions. In general, the CBR process for the task of disease risk assessment can be represented in Data Flow Diagram (DFD) notation (Fig. 1). Case base is a repository with medical data about patients and their anamnesis. New case is an external entity, which is considered as current state of some patient with his/her anamnesis. So, input data for CBR-method are appropriate set of features, symptoms and results of analysis. It can be identified according to the particular disease.
Features set of the case
Learned case
Set of retrieved cases Confirmed case
Solved case Retrieved case
Retrieve. It is a search of relevant precedents. According to chosen similarity measure or comparison method, it is necessary to find similar cases or case from case base datastore. Reuse. It is a process of using the retrieved precedent for solving the current problem. If proposed solution is suitable, than the problem is considered as solved. Revise. The suggested solution is evaluated for correctness and is adapted if necessary.
Result of this stage is the confirmed case.
Retain. The obtained case stored in the case base datastore as a newly learned case.
Let’s consider the methodology for determining the disease risk in more detailed way. In general, the methodology of any medical business process is the specific procedures or techniques used to identify, select, process, and analyze medical information. The disease risk assessment process is one of the business processes in a healthcare facility. Let us take a look at a functional model of this business processes in IDEF0 notation from the perspective of the CBR-approach (Fig. 2).
To eliminate the influence of the human factor, reduce the time spent for making medical decisions and decrease the errors, it is necessary to automate the process of risk assessment. It means the usage of informational technology to process medical data. Disease risk assessment with using CBRapproach consists of several stages. The outcomes of each stage are used as input data for the next one. Therefore, expert and medical informational system (IS) are mechanisms, which should be used together at each stage. Let’s take a closer look at the screening process for disease risk assessment. 1. Form a base of precedents. Each disease group is characterized by its own set of input data. Therefore, depending on the conditions of the assessment task, its own base of precedents is formed. It is the basis for medical decision for assessing the risks of a particular disease. At the same time, the formation of the base of precedents must comply with the current medical and technological documents. 2. Find relevant cases. This is the stage for retrieving the precedents. There are many methods for finding relevant cases. For example, Data Mining methods, genetic algorithms, decision trees, cluster analysis. One of the most common methods is the k-th Nearest Neighbors (kNN) method, which allows to calculate a similarity measure of cases [25]. Outcome of this method is a list of relevant -th precedents. The list is the basis for determination of characteristics of considered precedent. 3. Evaluate disease risks. At this stage, found precedent or list of precedents is reused and revised. The main question of this phase is whether there is a difference between the input dataset (current case) and the found cases. If the difference is insignificant, then the risk group is the same as the most frequently encountered group in the list of relevant cases. If the input data has many differences, then only a preliminary risk of the disease can be determined. Such risk should be refined. For instance, in work [26], it is proposed to use a model for choosing the optimal set of screening actions to form an individual survey plan. The resulting solution is saved to the base of precedents.
Let us consider the process of forming the base of precedents in the proposed methodology in more details. The model for identifying the set of precedents that is the basis for making a decision regarding to the risk group is presented on Fig. 3.
Analysis of the domain area allows forming the set of input data. It consists of symptoms of the disease or group of diseases, risk factors, concurrent diseases and the results of various examinations. The next step is to identify the selected medical data. For example, qualitative data should be formalized according to the selected scales; quantitative medical data should correspond to the accepted units of measurement according to WHO. According to the medical protocols and guidelines from WHO, a gradation of health groups is defined for the selected data. If the set of input information is too large, then the efficiency of the IS decreases. Therefore, it is necessary to evaluate and select informative features. This process is called dimensionality reduction [27, 28]. There are many different methods for feature selection or feature screening. The dimensionality reduction usually involves such approaches: filter, wrapper, embedded, and hybrid methods. The filter methods evaluate the relevance of features based on a given function. Spearman’s rank correlation coefficient or Information gain can be used as such function. Then, according to the given rule, expert defines which features can be left in the result set. The main idea of the wrapper methods is choosing the selection criterion or given classifier, which is a base for forming the feature sets. After that, the training quality of each set from the precedent base is assessed using given classifier. The next step is changing of the features set. It can be possible with the sequential removal or addition of features according to the convolution criterion changes. Another way is using the algorithm of full search (depth or breadth search), or a genetic algorithm. The embedded methods are similar to the wrapper methods, as the quality of training is also assessed using the selected classifier. The difference is that embedded methods allow the classifier to take into account some feature information during the training. This allows reducing the time of feature selection and reducing the risk of overfitting. The updated set of input data allows to select medical records from the database of records, which provide enough information for decision making.
Thus, the base of precedents is formed, which is then used for searching of relevant cases. 4.2.
In this study, it is proposed to conduct the process of retrieving the relevant cases using the kNNmethod. The main idea of this method is to select a set of precedents that are similar to the new given precedent. Firstly, the expert should choose the similarity measure. If type of input data is quantitative type, then the similarity measure can be calculated using the following metrics: Euclidean measure, Manhattan distance, the Mahalanobis distance, Chebishev measure and others. For qualitative and categorical information, it is possible to use Hamming distance, Rogers-Tanimoto dissimilarity index, Rao distance etc. But the medical information usually has mixed type, so it is recommended to use Zhuravlyov metric, Gower coefficient, Voronin measure, Mirkin metric [29, 30]. Let’s consider the algorithm of the kNN-method for screening the medical records for a selected disease (Fig. 4).
Identify input data
Normalize input data Calculate similarity measure for all medical
records Sort medical records according to obtained
similarity measure Input k or threshold value for similarity
measure Choose k precedents from sorted list of
medical records Are there equal values of similarity measure? no yes
Determine weighted vote for every class
Determine class with max number of votes
Consider the process of identifying the input data. Let’s denote as a set of different records of patient information in a medical card, for example, risk factors, diagnoses, symptoms, medical research results. Then is the set of records of the -th type. The example of features from risk factors group is a presence of bad habits, obesity, tobacco smoking, etc.
Let’s assume is a value of -th record of -th type from -th medical card. The range of changing of quantitative data in medical records can be very wide and differ from each other by several orders of magnitude. The kNN-method is sensitive to the range of changes of input variables, since the imbalance between the quantitative values of the input medical data set can cause the instability of the algorithm and can worsen the learning outcomes. Therefore, for quantitative data from medical cards, it is desirable to perform normalization or scaling. It means finding as the normalized value of -th record of -th type from -th medical card. One of the common variants of normalization is used the standard deviation according to the formula (1): where – is an average of -th record of -th type of all -th given medical cards from database, calculated by formula (2); is the standard deviation of -th record of -th type from the average , calculated by the formula (3): ,
The minimax normalization is also often used in the process of preparing medical data for the kNN- method:
The next step is to calculate the similarity measure between the medical cards. Let us denote by the similarity measure of -th and -th medical cards for -th record of -th type. The similarity measure for dichotomous and ordinal data is calculated by formula (5), and for quantitative data by formula (6): (6) It is proposed to use the Gower coefficient as a convolution criterion for mixed data. Denote as a weighted coefficient of the importance of the -th record of -th type. The coefficient , if -th record of -th type is absent in the -th or/and -th medical cards, in other case it equals to 1. (1) (2) (3) (4) (5) (7)
The obtained set of values of the convolution criterion is sorted in descending order. Next step
is determination of the number of the nearest medical cards, which are base for defining the health
risk group. Alternatively, expert can specify a convolution threshold value instead of the number of
the nearest neighbors. The top part of the sorted list of medical cards is the set , while
, that is, these are -th precedents, which are the basis for determining the class of the given
card. The set consists of 4-th subsets of medical cards according to the number of health risk
groups under guide [
If the set contains the different values of the Gower coefficient for medical records, then the given case is assigned the class with the highest number of votes, since the similarity measure of each record no longer plays a role in voting. If is the medical card from medical database and is -th risk group of -th medical card, where , then the risk group for the current precedent is:
If the set contains the same values of the mixed convolution criterion, then it is necessary to carry out a weighted vote of the relevant cases: (8) (9)
Thus, the model of retrieving the relevant medical records has been proposed for screening and identifying disease risk groups.
4.3.
Consider the use of the proposed methodology for determining the disease risk assessment for
cardiovascular disease. According to medical protocols and clinical guidelines [
According to medical documents [
If the data is mixed data with different dimensions, or an indicator is needed that is based on several symptoms or test results, than the input data must be normalized and identified. The following set of informative features with their possible values regarding the presence of cardiovascular disease was proposed in [19]: – sex: – woman, – man; – age: – less, than 40 years, – 40-50 years, – more, than 50 years; – diabetes: – presence, – absence, relevant data, – absence, irrelevant data, – no record; – hypertension: – presence, – absence, relevant data, – absence, irrelevant data, – no record; – renal dysfunction: – presence, – absence, – no record; – tachycardia: – presence, relevant data, – presence, irrelevant data, – absence, relevant data, – absence, irrelevant data, – no record; – inherited heart conditions: – presence, – absence, – no record; – smoking: – presence, – absence, – no record; – alcohol abuse: – presence, – absence, – no record; – hypodynamia: – presence, – absence, – no record.
Let’s consider the medical records of several patients with the unknown risk of cardiovascular disease. Information from medical records is presented as a set of features in Table 1.
A base of precedents has been formed for determining the disease risk assessment for cardiovascular disease. The base consist of 197 medical records with confirmed health groups. According to the proposed methodology, the measures of similarity were calculated between medical records from the base of precedents and new ones using formula (7). After sorting the obtained values, medical experts proposed to take 14 relevant medical records to determine the risk group. The type of input data is ordinal data (Table 1). Therefore, a large number of obtained data from calculating of the Gower coefficient has the same results. The results are presented in Table 2. In this case, to determine the class, it is necessary to use formula (9) to conduct a weighted vote of the relevant cases.
Analysis of the input and obtained data allows seeing that third card contains many gaps. Nevertheless, the known age and the presence of hypertension are the basis for identifying this card in a group with a high level of cardiovascular disease risk. At the same time, first and second card are filled almost completely, which allows concluding that they belong to the second health group. The obtained results provide information for medical decision-making by expert doctors in the field of cardiology regarding the early diagnosis of patients.
In order to use the proposed methodology for early diagnosis of patients based on data from medical records in practice, it is necessary to prepare data and check the adequacy of the medical information system. For instance, good way to evaluate the performance of a classifier is to look at the confusion matrix [32]. There are several terms, which are the base for creating of confusion matrix: Precision, Recall, F-measure, Accuracy [32, 33]. The matrix compares the actual target values with those predicted by the kNN-model. If is the number of true positive results; is the number of true negative results; is the number of false positive results; is the number of false negative results, then we have the following. Precision is the proportion of objects classified as that really belong to the class :
(10) (11) Accuracy is the proportion of right classified objects in the all classified objects:
It is difficult to compare two sets of input data with low precision and high recall or vice versa. So to make them comparable, it is proposed to use F-measure. It helps to assess Recall and Precision at the same time. It uses Harmonic Mean in place of Arithmetic Mean.
A series of experiments were carried out with different conditions. During every stage, part of medical cards with known the disease risks was considered as training pattern. The proposed methodology allows determining particular group for records from the training pattern. The results of the experiments are shown in Table 3.
Obtained data from above table proves the medical information system efficiency for determining the disease risks of cardiovascular diseases. Results of the confusion matrix can be the base for changing of the input data. For instance, if F-measure has low value, the number of input features is large and/or the processing time of the data by the informational system is very long, then we can use not only proposed way of screening and selection of informative features, but also the following method. Input features can be grouped by their meaning. For example, the concurrent diseases can be combined when diagnosing a specific disease. Then the numerical value of selected convolution criterion is determined for each group, as proposed in [19]. Such grouping of criteria will allow searching faster for the necessary information in the database with precedents and allow getting the best F-measure values. When training template is being used for checking the correctness of the kNNalgorithm, it is necessary to change the value of as well, because it determines the cardinality of the set of relevant precedents. If the value of the external proximity criterion is reliably known from the expert, then the number of closest cases can be replaced by the threshold value of the similarity measure between medical records.
Therefore, aforementioned recommendations of changing the input data can help in improving the performance of medical IS. The conducted studies and the obtained results show the feasibility of using the proposed methodology in real conditions.
In the course of this study, a methodology for risks disease assessment based on the medical records has been proposed. The following results have been presented: the analytical review of early diagnosis methods or methods of disease risk assessment has been conducted, which allowed to choose the CBR-approach to the implementation of the task; the formalization of the screening process of medical cards using IDEF0-notation was presented, which allowed to choose several stages of medical data processing and combine them into a common methodology for disease risk assessment; the algorithm of kNN-method has been developed for identification of the groups of risks diseases; the model of retrieving of relevant precedents has been proposed, which allows to solve the problem; the numerical studies have been conducted, which shows the possibility of usage of the proposed methodology in real world settings.
The scientific novelty of the obtained results is the improvement of the process of early diagnosis with the help of the proposed methodology, which allows to process quantitative, qualitative and scaled data simultaneously from medical records for more effective medical decision-making.