Hybrid clustering based health decision-making for improving dietary habits

2.1Healthcare for medical decision-making

The period and frequency of visits to medical institutions differ depending on individuals. People who suffer from similar diseases, such as flu, gastritis, enteritis, or chronic diseases can be guided in similar ways in terms of treatment, care, and prevention. For this reason, patients with similar disease can be prescription through classification.

Delias et al. [9] developed the medical care decision-making support model using log clustering. The developed model is able to cluster activity logs of patients, extract knowledge, and improve and monitoring process. Accordingly, the patients with similar activity are clustered and classified as related chronic diseases. In this way, it is possible to predict the workload per activity in association with an estimate of the statistical distribution of diseases. In addition, it is possible to give the medical staff the information to optimize the operation of medical guidelines. Based on a variety of information including the relations between activity and diseases, it is necessary to cluster them and extract useful information.

Figure 1.

Healthcare for medical decision-making.

Jenkins et al. [10] proposed the clustering method based on chronic obstructive pulmonary disease (COPD). The developed method clustered clinical test records of the adults with COPD. According to the clustering, the symptom of anxiety related symptom were found to be a common subject in a clusters of psychological symptoms. It is useful for managing the patients with different symptoms and making statistical approach. However, there were different results depending on the selection of analysis tools. Respiratory difficulty was found in the clusters with respiratory symptoms, but some had meaningless results from clustering. Accordingly, clustering the symptoms related to COPD and the factors influencing health is required to be consistent with a lapse of time.

For health decision-making, deep learning method, such as LSTM, CNN+RNN, Mask R-CNN, and GAN are used. In terms of the application of deep learning to decision-making, since the structure for optimal performance has yet to be established, there has been research and development of the optimized structure for learning. In addition, it takes a great deal of time to secure plenty of health data. For this reason, it is also necessary to build the deep learning structure that guarantees excellent performance with the initial small amount of data. Accordingly, Generative Adversarial Networks (GANs) is researched in order to increase the number of training data from existing data. In addition, in order for health decision-making with the uses of limited health information and the deep learning lightweight partial-learning structure, the establishment of knowledge base for performance deep learning, GAN model for generating new knowledge, and the high-performance deep learning structure optimized for the learning in combination of auto-encoder and CNN+RNN are developed. Figure 1 shows the healthcare for medical decision-making.

In Fig. 1, users’ lifelogs are collected for medical decision-making. Lifelogs are activity information, nutrition data, self-measured data, and social data. The medical platform for health decision-making analyzes personal medical records with which medical staff show users’ health conditions. As analysis methods, clustering and such deep learning as CNN, RNN, Mask R-CNN, GAN, and RNN+LSTM are used. As the result of the analysis, knowledge base is generated, and a variety of health information including exercise management, dietary management, potential risk of diseases, and disease association, are provided to each user through ambient IoT devices.

2.2Health service for lifecare

In modern society, the number of people with chronic diseases increases due to westernized dietary habits, stress, and irregular lifestyles in their life. Therefore, it is necessary to improve improper dietary habits and manage diet and nutrition for the quality of healthy life. In addition, it is required to provide services to obtain reliable healthcare information easily. Accordingly, a variety of healthcare services are developed at country and abroad.

Welltizen [11] developed the calorie manager mobile service. The calorie manager provides in real-time information on calories and nutrition of the food products to take in through mobile. In addition, the service can manage customized diets for weight control by individual users. Nutrients are different depending on a food recipe. For this reason, the calorie manager provides limited information. Therefore, it is necessary to find a plan for providing the diets and nutrition of diverse recipes.

Asan Medical Center [12] provides health contents that general users and medical staff can easily understand. Medication Application offers drug information to patients and general users. In the service, users are able to search for prescribed drugs at hospital by name, shape, and color. In addition, drug efficacy, side effects, and other information are provided. In this way, it is possible to increase patients and general users’ reliability of prescribed drugs. The personalized application helps patients receive their history of diseases, test results, drug injection management information, and personal healthcare information. By letting chronic disease patients regularly enter their lifelog for their continuous management of diabetes, blood pressure, and weight, it is possible to provide their body mass index and disease information and help them care for health and diseases. Such application has the limitations of service hospitals and users. Therefore, it is necessary to develop the application with which everyone can receive medical service.

In Japan, SUZUKEN [13] provides a health creation service in which general users are able to receive dietary management support, health information, and hospital search results on their mobile for a healthy life. In addition, it offers such services as walking, personal health data collection, diet simulation, diary, body check, health questions, hospital search, and dietary information. It had the advantages of convenience and continuance, since it supports healthcare at any times and any places. Nevertheless, it still provides healthcare contents which are similar to those of other products so that differentiated services need to be offered. Figure 2 shows the configuration of context knowledge mining analysis for lifecare.

Figure 2.

Configuration of context knowledge mining analysis for lifecare.

In Fig. 2, the lifecare service is had composition to three type: dietary management, exercise management, and daily management. It is through mobile applications and IoT based wearable devices, it is possible to collect and monitor lifelogs generated in everyday life. With the use of the collected data, pattern analysis, correlation analysis, and potential health risk analysis are conducted. As a result, a personalized service for dietary habit management and exercise management is offered depending on a user’s health condition. Users with high potential health risks can have medical examination and treatment at any times and any places through telemedicine. In addition, in terms of dietary habit management, a variety of information on food products, such as calories, nutrients, and recipes, can be offered differently depending on chronic diseases. In terms of exercise management, exercise effects of cycling, jogging, weighting, yoga, stretching, soccer, and others, and exercise guidelines are offered. In this case, weight, blood pressure, blood sugar, and sleep hours are monitored personalized to manage lifecare in a customized way. Given all, many different social issues, including ageing acceleration, rapidly rising medical cost, and poor medical infrastructure, are solved by telemedicine platform and healthcare solution.

3.1Chronic disease clustering based dietary nutrition ontology

Appropriate dietary habits can be determined differently depending on personal health conditions, and it is necessary to take into account user’s chronic diseases. In order to analyze the correlations between dietary habits and chronic diseases, context knowledge mining is used so as to find the factors that positively influence health conditions. As for the health data of chronic diseases, the raw data [15] of National Health and Nutrition Examination Survey [14] are used. They include a variety of data about chronic diseases, including obesity, diabetes, dyslipidemia, and high blood pressure. In addition, with the use of information on weight, height, waist line, body mass index, fasting blood sugar, glycated hemoglobin cholesterol, regularity of heart rate variability, and pulse rate, it is possible to predict whether to have a chronic disease [15, 16]. Clustering can occur by extracting the characteristics and representative values though the context knowledge mining. Raw data has variety attributes of context information. In this study, the health data about typical chronic diseases, including obesity, diabetes, dyslipidemia, and high blood pressure, are preprocessed. Table 1 shows the properties of chronic disease data. It is possible to predict obesity with 5 attributes, diabetes with 3 attributes, dyslipidemia with 6 attributes, and high blood pressure with 13 attributes.

In this study, we used k-means algorithm for clustering the chronic disease data in healthcare area [17]. In the algorithm, the number of clusters is determined by the k value. In case of 11 number of clusters, data distribution is the most even through repeated experiments. Table 2 shows the characteristics of each cluster of the data about chronic diseases. A representative value means the value representing the overall characteristic of data, and the mean is used as a representative value [16]. With the use of a representative value, it is possible to predict where each cluster data is mainly distributed in obesity, diabetes, dyslipidemia, and high blood pressure. Table 2 shows the chronic disease characteristics of each cluster according to properties. If a user belongs to C1, the person is highly likely to have low weight in terms of obesity and normal levels in terms of diabetes, dyslipidemia, and high blood pressure.

With the uses of the characteristics of clusters and the diet and nutrition service ontology [18], food products are recommended to each cluster. The diet and nutrition service ontology has the relations between chronic diseases and food products. With the service ontology, it is possible to recommend the food products for nutrition therapy, nutrition treatment, and dietary therapy which are helpful to chronic disease patients. It expresses that the food products analyzed by persons with socially experimental knowledge or researchers positively or negatively influence chronic diseases. With the use of the rules inferred from the diet and nutrition service ontology, it is possible to recommend diet and nutrition and guidelines which are expected to give positive influence to the health of each cluster. Figure 3 shows the diet and nutrition service ontology [18, 19].

Figure 3.

Diet and nutrition service ontology.

In Fig. 3, an objective is classified into Class and Property. With the use of the relations between class and class, and between class and property, a variety of information on food products, diet and nutrition, and health condition is presented. The diet and nutrition service ontology expresses the complicated relation between class and property with the uses of health context and diet and nutrition information.

3.2Classification of chronic disease patients according to health status

Based on the clusters of chronic disease data, it is possible to classify the clusters which have the most similarity with active users. In other words, with the use of similarity scale, the cluster that has the closest with an active user’s property value becomes the user’s cluster. As similarity scale, euclidean distance is used. An euclidean distance [20, 21] represents the distance between two properties, which can easily be measured two-dimensionally or three-dimensionally. The smaller the value is, the closer two properties are and the more they have similarity. Active users are classified with the cluster that has the nearest euclidean distance value. Table 3 shows cluster classification of users on the basis of euclidean distance. In the table, User1 has the nearest distance with the cluster C3. User1 is classification into the cluster C3. As shown in Table 2, the person is highly likely to have normal weight in terms of obesity, fasting glucose insufficiency of diabetes, abnormal dyslipidemia, and normal blood pressure. It means that the classified users have the prevalence of chronic diseases mainly found in clusters or have the high possibility of prevalence. Accordingly, with the use of diet and nutrition service ontology [18], it is possible to recommend users to the food products customize the chronic diseases mainly found in their cluster. In this way, it is possible to drawn the dietary habit that positively influences a particular user’s health condition. In fact, this is the stereotypical food recommendation method that fails to consider a user’s preference. Therefore, an active user has difficulty maintaining the recommended diet or food intake, often. Accordingly, a user’s preference is additionally taken into account through chronic diseases clustering, the diet and nutrition service ontology, and food clustering based hybrid clustering.

3.3Food cluster according to dietary composition

Food intakes are directly related to health in everyday life [1]. Appropriate food intakes help to supply essential nutrients for physical activity. Food products are classified into general food, health food, and health functional food by manufacturing type, main ingredients, and use. General food means the food products taken basically in everyday life. It has the highest intake frequency in human life so that managing general food leads to healthcare. The components of food include calorie, carbohydrates, proteins, fats, sugars, and sodium. Although food products have different ratios of components depending on ingredients, they share common factors. Chronic diseases need to be cared for continuously, and accordingly, appropriate nutrient intakes are of importance [1, 2, 3, 4, 5, 6]. It is hard or easy to take food products helpful for health promotion depending on users’ preference. Therefore, it is necessary to recommend alternative food whose nutrition structure is similar to that of the food for user’s health promotion [8]. In conventional data, food is classified into rice, soup, grilled food, braised food, and other types of food by main ingredients and recipes. A food product has a calorie and such contents as carbohydrates and proteins fats differently depending on the ratio of main ingredients and sub ingredients. Therefore, the classification method does not include diet and nutrition. To analyze the diet and nutrition components of food products, groups of similar food products are clustered. For the data for clustering, the food and nutrient database [23] provided by the portal site of Food Safety Korea [22] is used. It consists of 4,577 food data, and particular food products have representative values of regions, which can be considered to be data duplication. In this study, such duplication is unnecessary for clustering food products and, therefore, 3,153 food data are clustered without duplication. A food product can have different diet and nutrition depending on its one-time offering amount. To analyze how much components of food products are similar, calories (kcal) are applied to the normalization of 1,000, equally, and then cluster analysis is conducted with the use of k-means algorithm [17]. The proposed method needs additional operations to recommend both the food products helpful for users’ health and the food products whose diet and nutrition is similar to that of the healthy food. To minimize the worsened turnaround time by the additional operations, the algorithm is repeated to find the k value in which the size of each cluster is 300 or less. In k-means algorithm, clustering precision is different depending on initialization and cluster count (k) values. For optimal clustering, cluster analysis is conducted repeatedly. As the result of the repetition, when the k value is 25, the sizes of all clusters are 300 or less. The 25 clustering results made at that time are saved and used for a model. Table 4 shows the food product clusters on the basis of the components of diet and nutrition. With the use of them, it is possible to recommend the food products customized each user’s health condition and their similar food products.

3.4Knowledge base-based healthcare for improving dietary habits

The diet and nutrition knowledge base includes weights for recommending the food products customized each user’s health condition and their similar food products. It has four filters. Through each filtering, the final food products are recommended to users. Figure 4 shows the structure of the diet and nutrition knowledge base. As filters, there are food cluster filter, food similarity filter, universal preference filter, and user feedback filter.

Figure 4.

Structure of diet and nutrition knowledge base.

The food cluster filer means the representative value of the 25 food product clusters classified by the components of the diet and nutrition (Table 4). The food similarity filter means the filter of a similarity coefficient between food products in each one of 25 clusters. As a similarity coefficient, Pearson correlation coefficient (also known as Pearson’s r) is used [24]. A Pearson’s r ranges from -1 to +1. If the value is closer to -1, it means negative correlation; if it is closer to +1, it means positive correlation. In the Eq. (1), f⁢s⁢(a,b) represents the Pearson’s r of food products a and b represents the ith diet and nutrition value for the food product a. i represents carbohydrates (1), proteins (2), fats (3), sugars (4), sodium (5), cholesterol (6), fatty acids (7), and trans-fatty acids (8), sequentially. a¯ is the mean of a values. Accordingly, a1 represents the carbohydrate content of the food product a.

The universal preference filter is the preference weight for the food products that Korean people mainly have in their everyday life according to a questionnaire survey. The preference data for the universal preference filter is based on the preference of Korean people’s representative food products collected in the survey [25, 26]. The data are collected from 197 persons who gave answers with sincerity among 250 questionnaire participants in the online and offline environments. Of the 3,940 preference data of 20 food products evaluated by 197 participants, 2,758 (70%) are used in order to make a preference filter, and the rest (30%) is used for performance evaluation. With the use of the preference for food products in the collected data, it is possible to predict users’ preference on the basis of food components. A Pearson’s r is applied in order to analyze the relation between food components and preference and is saved as a weight of universal preference [27, 28]. The weight presents the diet and nutrition component preferred by people universally. In the Eq. (2), u⁢w⁢(n,s) represents the Pearson’s r of diet and nutrition = {carbohydrates, proteins, fats, sugars, sodium, cholesterol, fatty acids, trans-fatty acids} and preference = {1, 2, 3, 4, 5}. ni represents the value of the diet and nutrition n for the ith food product. i means the sequential input value of 20 food products. si represents the value of the preference s for ith food product. With the use of the equation, it is possible to present the relation between food and diet and nutrition numerically.

In the Eq. (2), the relation between diet and nutrition and preference is calculated. And then, collaborative filtering is applied to the calculated result in order to predict user’s universal preference for new food products. The Eq. (3) presents the collaborative filtering for predicting food preference. In the equation, P⁢(p⁢s,k) represents the universal user preference predictive value for the recommended food product k. nk means the value of diet and nutrition n of the recommended food product k.

The user feedback filter means the preference weight filter continuously updated through active user’s recommendation and feedback. It presents the preference weight of the food product evaluated by an active user, which is also found in the universal preference filter [29, 30, 31]. The user feedback filter displays the goodness-of-fit of each recommended alternative food, and lets the model adjust their weight in order for the next recommendation. With the use of the Eq. (2), the initial weight is calculated in the same way as the universal preference. An active user and the weight of the diet and nutrition property are presented with w⁢(a⁢u,n). For the prediction of user preference, w⁢(a⁢u,n) is substituted for u⁢w⁢(s,n) in the Eq. (3). As the result of four filters, both the predicted universal preference and user preference are offered to a user so as to support the user’s decision-making of diet and nutrition. The Eq. (4) presents the correction of the preference weight of an active user through back propagation. In the equation, e represents an error which is a difference between the predicated value through the diet and nutrition knowledge base and the value selected by a user. normal(n) represents the value of the normalized diet and nutrition n from 0 to 1. ‘0.1’ is the weight update rate.