An Intelligent Long-Term Care Information Platform using Machine Learning and Semantic Web

doi:10.21203/rs.3.rs-2986151/v1

Download PDF

Research Article

An Intelligent Long-Term Care Information Platform using Machine Learning and Semantic Web

https://doi.org/10.21203/rs.3.rs-2986151/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

The low use of Long-Term Care (LTC) services in Taiwan is mainly caused by the lack of convenient promotion tools. This study developed a general architecture integrating Clustering Algorithm into Domain Ontology based on Cloud Computing (CADOCC) to provide a cloud computing environment and support the big data required for machine learning. The proposed CADOCC comprised four integrated components: cloud computing, machine learning, data preparation, and semantic web modules. The proposed CADOCC was used to develop an LTC Application Platform (LTCAP) to assist users in automatically classifying and filtering articles related to LTC. Additionally, the LTCAP was used to produce an intelligent mobile app that continually learns from user question–answer interactions through machine learning to facilitate the use of LTC services for personal requirements. The performance and satisfaction scores of classification algorithms running in different spark cloud computing environments were tested and compared. The results revealed that LSI and K-means met requirements in a test scenario, and the total satisfaction score was 4.15, which confirmed the feasibility of CADOCC.

Long-Term Care

Machine Learning

Semantic Web

Cloud Computing

An aging society is a worldwide trend in advanced countries. Long-Term Care (LTC) is a crucial medical service that an aging society must provide to its residents. Taiwan provided LTC services in 2007, which was referred to as LTC 1.0. LTC 1.0 presented some limitations, such as an insufficient supply of service items, a narrow range of service targets, no support from medical institutions, and inconvenient application for funds. In recent years, the Taiwanese government has promoted LTC 2.0, which overcomes the aforementioned limitations of LTC 1.0. However, the rate of use of LTC 2.0 services is insufficient because people do not understand the complicated LTC 2.0 policy. A case manager (CM) of LTC is the most crucial staff member for LTC promotion. CMs use paper leaflets or tablets to introduce the LTC policy to people. Introducing the LTC 2.0 policy to people only through leaflets or tablets by the CM is inefficient. The CM passively introduced the LTC 2.0 policy to people instead of providing them information regarding LTC that can be easily and actively obtained. This study investigated an approach to make the LTC information available to people actively and easily.

An app is a computer program designed to run on a mobile phone. The use of apps to acquire information from the Internet has become popular globally. Some Taiwan-based apps introduce LTC policies and services developed by the Ministry of Health and Welfare[1]. However, the lack of artificial intelligence in these apps can be overcome by providing personal requirements and hyperlinks to the useful information for users to promote understanding of LTC policies. Moreover, this is the main reason for the low rate of use of LTC services in Taiwan. This study proposes a general architecture that integrates Clustering Algorithm into Domain Ontology based on Cloud Computing (CADOCC), which integrates statistical methods, vector space, machine learning and semantic web technologies in cloud computing ecosystem to facilitate performance and promote intelligence for various application domains. Based on the CADOCC, this study develops Long-Term Care Application Platform (LTCAP) to assist users automatically classify and filter out LTC related articles. Additionally, the LTCAP provides an intelligent mobile app continually learn from user question-answer interactions using machine learning technology to facilitate LTC services for personal needs.

The developed LTCAP adopts Spark cloud computing to promote computing efficiency and utilizes machine learning associated with open data approach to promote LTC 2.0 policies. Two requirements must be met when developing LTCAP. The first is the performance evaluation of statistical methods combined with machine learning through Spark cloud computing. It can propose statistical methods and combine clustering algorithms to form big data applications. The second is the correctness assessment of LTC information provided by the combination of open source and machine learning. It can facilitate to develop machine learning technologies and integrate into LTC policies to promote intelligent capabilities for various applications. This study makes three main contributions. First, an intelligent mobile app is developed to provide question-answer interactions to reduce waste on time for users when accessing LTC information and facilitate LTC services for personal needs. Second, the CADOCC is proposed to integrate statistical methods and vector space with machine learning into Spark cluster cloud computing. The proposed CADOCC is modular to make it easier to manage and expand the functions and various intelligent mobile applications .Third, the CADOCC is adopted to implement the LTCAP that helps government to promote LTC policies. The LTCAP carries out to integrate four emerging research areas: clustering algorithms, machine learning, open data, and cloud computing.

The remainder of paper is organized as follows. The next section presents some related studies. Section 3 presents an integrated Clustering Algorithm into Domain Ontology based on Cloud Computing (CADOCC). Section 4 developed a Long-Term Care Application Platform(LTCAP) based on the proposed CADOCC. In Section 5, this study presents performance evaluation and experimental results. Finally, summary and concluding remarks are included.

In an aging society, promoting Long-Term Care (LTC) to meet the care needs of the elderly is critically important. In order to respond to this demand, the government should provide accurate and sufficient LTC information to facilitate appropriate treatments. Unfortunately, LTC policies and information are usually so complicated and difficult to understand that applicants do not know how to apply for their LTC services[2]. The main focus of this study is to address how emerging information technologies, including machine learning, cloud computing, open data, mobile App, may be leveraged and integrated to overcome the aforementioned restrictions faced in LTC servicing.

In recent years, some studies [3–13] have used information technology to promote LTC development and application. In [3], author proposed a new framework for long-term care using association rules to improve the problem of dynamically deploying caregivers. A kind of monitoring that collects patient activity data through wearable devices and games can be used in long-term care application. INTESA research project [4] provided monitoring and rehabilitation services based on mobile and wearable devices, which have been deployed in the nursing home. In [9], authors developed an elderly health monitoring model that adopts ambient technologies and serious games to facilitate the collection of medically relevant data. Mobile phones have been widely used in daily life today. A mobile App-based e-learning platform[6] is developed for has been developed as a long-term caregiver education and training. Open data (OD) [14] is available on the Internet with an open license. Anyone can use OD to facilitate the development of their information systems. Open government data (OGD) is a kind of OD and is high-quality data that is freely provided by governments for public use. In [8], authors used OGD to develop a visual analysis system for long-term care applications. Social robots are emerging information technologies in recent years. In [5], authors proposed a remote health care online system that employed a moving robot for the elderly at home. In [10], author proposed the Internet of Things (IoT) application in the field of LTC. The use of sensors to collect information for analysis has become an important issue in the safety of the elderly.

The main limitation of the aforementioned papers is that they have not developed a general framework to integrate various information technologies to promote LTC services. The general framework is necessary for developer to facilitate the implementation of LTC applications, which are critical for system modularity and reusability in heterogeneous development environment. In addition, none of these papers use machine learning to enhance the intelligence of the LTC applications. This study therefore defined a general framework, Clustering Algorithm into Domain Ontology based on Cloud Computing (CADOCC), to integrate machine learning, cloud computing, open data and semantic web technologies to improve above issue.

In recent years, cloud computing has become one of the most promising information solutions and business trends. There are many studies [15–18] to discuss the cloud computing technologies and applications in various domains to promote computing efficiency. Machine learning technologies are widely used in various domains to facilitate the intelligence of information systems [19–25]. Unverified LTC information and OD continue to increase on the Internet. Big data and OD are often discussed together because the heterogeneity, distribution, and processing of large amounts of such data play an important role in social web[26]. Machine learning and cloud computing can be used to analyze big data for intelligence and to improve the associated computing performance. In recent years, many studies [27–34] have used statistical methods, machine learning or semantic web technologies to improve the intelligence of information systems. These technologies may all need to process big data.

Cloud computing is used to process big data to improve performance [35, 36]. The most common example is Hadoop, where Dalia Sobhy and others[37] explore the impact of the cloud on healthcare. With its low cost, high scalability and usability, cloud computing can store and analyze medical records and provide an easy-to-use medical system. Prathyusha Rani Merla et al.[38] used the Hadoop MapReduce framework to analyze YouTube data to find hot topics, which could help invest in more trend areas. Akaash Vishal Hazarika et al.[39] discussed the performance comparison between Hadoop and Spark in their study, since Hadoop is mainly processed in hard disk, while Apache Spark is processed in memory to improve performance. The result is that Spark takes more time than Hadoop.

3.1 General Architecture

This paper defines an integrating Clustering Algorithm into Domain Ontology based on Cloud Computing (CADOCC) architecture for supporting various applications involved in dealing with intelligent mobile app development. This paper argues that CADOCC can be employed as a common framework to integrate clustering algorithm and domain ontology uniformly using a fundamental cloud computing environment. The proposed CADOCC consists of data preparation module, machine learning module, semantic web module and cloud computing module. Figure 1 displays the CADOCC, which implies the application framework at a high level of abstraction.

3.2 Data Preparation Module

In the data processing stage, the data preparation module performs data acquisition, filtering, and formatting operations to generate available training data to the machine learning module. In data acquisition operation, data source can be web-based data acquired from different web sources. The collected data in this study included OGD, LTC related articles and online news. In data filtering operation, Jieba [40]is employed for automatic segmentation of Chinese word. In the data formatting operation, Term Frequency–Inverse Document Frequency (TF-IDF) [41] is used to evaluate the weight of words from a article. A practical case description of the data preparation module is in Section 4.1 and 4.2.

Machine Learning Module

The machine learning module adopts mathematical and statistical techniques, for example, K-means clustering, K-NN(K Nearest Neighbor)[42] classifier, fuzzy c-means, bayesian network, neural network, Latent Semantic Indexing (LSI)[43], etc. Using these algorithms, programs can learn the rules in data for use in decision-making and prediction. This study employed K-NN classifier to categorize each class on domain ontology and K-means clustering to extract hidden topics from article posted on social web. In the training process, the machine learning module is used to design a predictive model that can be reused by the semantic web module once it is established. Therefore, the training process is not required every time the App interacts with a user.

LSI associates words with similar semantics in documents. Generally, the correlation between individual words is ignored when searching for keywords. However, in practical applications, individual words may actually be related to each other. LSI uses singular value decomposition (SVD)[44], dimension reduction (DR) and vector space model to consider these correlations. SVD is highly useful in machine learning to gather intelligence for various applications [45]. The association between an article and a word is used in SVD to determine the category of an article. In this study, we used clustering and feature extraction to identify LTC information and LSI to determine the similarities among OGD, news, and LTC information.

Semantic Web Module

The Semantic Web Module is composed of related web technologies, including XML-based markup language[15], open data, ontology, and logic rule, etc. The XML-based markup languages contain Extensible Markup Language (XML), Resource Description Framework (RDF), RDF Schema (RDFS), Web Ontology Language (OWL), etc. A web-based open data that can be accessed using URL. An ontology is developed with an RDF-based markup language to define a shared conceptualization of a special domain. There is a detailed definition of the ontology in our previous paper[46]. The complex semantics can be presented with logic rules, for example, Semantic Web Rule Language (SWRL), Jena-based rule, etc.

Cloud Computing Module

The cloud computing model can be represented as a three-tier architecture, including Infrastructure as a Service (IaaS), Platform as a Service (PaaS), and Software as a Service (SaaS). The developers would typically deploy their applications on the infrastructure. Cloud computing module is used to improve the performance of machine learning module and semantic web module. Excessive data can lead to slow computing or even system failure. Cloud computing technology can solve these problems.

Hadoop is a cloud computing platform that contains of MapReduce and HDFS (Hadoop Distributed File System) [47]. The MapReduce enables distributed computing, and the HDFS achieves a distributed file system with the advantages of high fault tolerance. The YARN [48] is a kind of cluster manager that is proposed in Hadoop 2 version. Apache Spark [49] is also an open-source platform that adopts in-memory technique and Resilient Distributed Dataset (RDD). The data can be calculated in the memory and finally written into the hard disk to be accessed. In this study, the cloud computing technology used Spark for the computation of LSI and K-means clustering.

3.2 Long-Term Care Application Platform

In order to verify the feasibility of CADOCC, a Long-Term Care Application Platform (LTCAP) is developed based on CADOCC, as shown in Fig. 2.

1. User

1.1 User operates App to access LTC policies and information in the LTCAP.

1.2 The LTCAP presents the corresponding information and replies as required.

2. Developer

2.1 System manager imports the LTC related OGD and news to the database.

3. Computing

3.1 User access the information to the database through LTCAP for subsequent recommendation.

3.2 Google Maps API is invoked to retrieve OGD information and process the LTC article information into the format required by the machine learning.

3.3 The user profile is converted to an RDF file which excludes personal security data. The Jena inference engine uses the RDF file to infer the LTC services and organizations for user needs. The detailed flow chart of Jena inference engine in introduced in Section 4.5 and 4.6.

3.4 Machine learning technology, including LSI, K-means and K-NN, is used to design a predictive model that can be reused once it is established. An example of machine learning processing is introduced in the Section 4.3.

3.5 LSI, K-means and K-NN algorithms are developed on the Spark cloud environment to improve machine learning performance.

4.1 Data Collection and Preprocessing

This study collected 8089 LTC related articles from internet. This study used Python Scrapy for extracting these LTC related articles from websites, and then imported these LTC data into database. These extracted LTC data is shown as Fig. 3.

The collected LTC data must be processed before further use. In this research, the three steps of data preprocessing are Chinese word segmentation, punctuation removal and stop word removal. As shown in Fig. 4, in Chinese word segmentation, Chinese content is divided into multiple words by Jieba. Because punctuation does not help extract semantics, it is required to delete punctuation. The stop word is not necessary for content processing, so it has been deleted.

4.2 Term Frequency–Inverse Document Frequency

Term Frequency–Inverse Document Frequency(TF-IDF) is adopted to calculate the weighted value of each topic in each of the 8089 preprocessed articles. For example, suppose calculations must be performed for 100 articles, and the TF-IDF value of a single topic “health” in an article d must be calculated. As expressed in formula (1), tf(t,d) denotes the frequency at which a single topic appears in an article. Assuming that the single topic “health” appears four times in article d, the tf(t,d) of the single topic “health” is 0.25.

$$tf\left(t,d\right)={f}_{t,d}$$

Assume that the single topic “health” appears in five of the 100 articles. According to formula (2), the idf(t, D) of "health" is about 1.22184.

$$\text{i}\text{d}\text{f}\left(\text{t},\text{D}\right)=\text{log}\frac{N}{1+\left|\right\{d\in D:t\in d\left\}\right|}$$

The TF-IDF of “health” in article d is calculated using tf(t,d) and idf(t,D). According to formula (3), the tfidf(t,d,D) = 0.25×1.22184 = 0.30546 for “health” in article d.

$$\text{t}\text{f}\text{i}\text{d}\text{f}\left(\text{t},\text{d},\text{D}\right)=tf\left(t,d\right)\times \text{i}\text{d}\text{f}\left(\text{t},\text{D}\right)$$

The TF-IDF for all other topic–articles can be calculated by repeating these steps. The larger the TF-IDF of the topic in the article, the higher the relevance to the article.

4.3 Machine Learning

This section describes the process of machine learning using K-means and K-Nearest Neighbors(K-NN) algorithms. Machine learning requires feature vectors of data. The feature vector is composed of numerous numbers. Articles must be converted into vectors to be learned by K-means and K-NN algorithms. In this paper, the Latent Semantic Indexing (LSI) was adopted to convert the articles to vectors.

LSI

A topic–article can be categorized into K topic categories according to certain requirements. LSI model requires TF-IDF vectors, pre-processed articles and K topic categories. The K value is defined as K topic weights for each article, and the K value is set to 300 in this study. The 8089 LTC related articles transform to an 8089$\times$300-dimensional matrix using the TF-IDF and LSI, as shown in Fig. 5.

K-means

k-means clustering directs to partition n objects into k clusters in which each object belongs to the cluster with the nearest cluster centroid. This study uses K-means clustering to classify 8089 articles into 300 groups. Loading 8089 × 300-dimensional article vectors from LSI (refer to Fig. 5) as feature fields into the K-means algorithm will predict different clustering results for 8089 articles. Figure 6 shows the K-means clustering results.

K-NN

k-Nearest Neighbors(K-NN) is a kind of instance-based learning that all computation is deferred until function evaluation. This study uses K-NN to classify user’s search question into 300 groups, and then compare which one is closer to the search question. K-NN is assumed to be 30, which means that the vectors of the search question and the 30 nearby vectors of the nearest articles should be used to determine the classification results. First, the user's search question uses LSI transformation to generate a 1 × 300-dimensional matrix vector. Second, the 8089 × 300-dimensional article vectors of K-means is used as the characteristic vector associated with 1 × 300-dimensional search question vectors imported to the K-NN algorithm. The K-NN predicts the classification result of the user search question, as shown in Fig. 7.

4.4 Cosine Similarity

This study tries to recognize the relationship between user’s search question and LTC article by calculating the cosine similarity between vectors, as expressed in (4). A larger cosine value implies that there is a high similarity. Since both the LTC articles and user’s search question in our system are represented as vectors, the customized retrieval result can be directly calculated based on their cosine similarity. Figure 8 shows the top five articles that are most similar to user’s search question.

$$\text{cos}\left({\theta }\right)=\frac{\sum _{i=1}^{n}\left({A}_{i}\times {B}_{i}\right)}{\sqrt{\sum _{i=1}^{n}{\left(Ai\right)}^{2}}\times \sqrt{\sum _{i=1}^{n}{\left({B}_{i}\right)}^{2}}}=\frac{A\cdot B}{\left|A\right|\times \left|B\right|}$$

4.5 Semantic Web

Ontology and rule are the core technologies of semantic web. Ontology is composed of classes and relationships between classes. Rule is developed based on ontology classes to support inference for intelligent mobile application. In this study, Domain Ontology is coded in OWL and rule is coded in Jena-based rule.

Domain Ontology

As shown in Fig. 9, Domain Ontology is developed based on LTC domain that consists of seven classes: long-term care cases, articles, keywords, services, regions, service centers and comments, which will be explained as follows.

long-term care cases class: the users of long-term care.
articles class: LTC related articles of interest to users
keywords class: topics that users are interested in.
Services class: services provided by long-term care agencies.
regions class: regions where users and service centers are recorded.
service centers class: LTC organization.
comments class: capturing comments from the service centers through the Google Maps API.

There are 14 and 3 subclasses in long-term care cases class and service centers class, respectively. The instances are the fact information in the rule inference processing. There are 553, 8089, 301, 153, 1217, 216 and 6518 instances in long-term care cases class, articles class, keywords class, services class, regions class, service center class and comments class, respectively.

Rule

Figure 10 presents Jena-based Rule 1, where an LTC case (user) lives in an area (area), and the LTC case requires a service (service). Furthermore, the service center is located in the same area. The service center class is denoted as a class, and the class provides the service item that the LTC user requires. The service center is recommended to the LTC case.

Figure 11 presents Jena-based Rule 2 with the current status of the LTC case and indicates that when a service item (service) fits the current status of the LTC user, the service item is recommended to the LTC user.

Figure 12 presents Jena-based Rule 3, which denotes the article in which the LTC case is interested. The keywords included in the article are the same as the service item. The service item is recommended to the LTC case.

4.6 Integrating Machine Learning and Semantic Web into Cloud Computing

The developed LTCAP integrates machine learning and semantic web into the Spark cloud computing environment, as shown in Fig. 13. The detailed process steps of machine learning and semantic web are shown in Fig. 13(a) and Fig. 13(b), respectively.

a. Integrating machine learning into the Spark cloud environment

Retrieves and preprocesses the LTC related information.
The preprocessed LTC data are converted into a vector matrix.
Use TF-IDF to get the weights of the words in the LTC related article.
Set K value 300 to establish the LSI model.
8089$\times$300-dimensional vector matrix are generated by step 4.
The 8089$\times$300-dimensional vector matrix is grouped by K-means algorithm and the K value is 300. The results will be the grouping to which each article belongs.
User input a search question through LTCAP.
The question statement is processed by Jieba and TF-IDF to establish LSI module.
1$\times$300-dimensional vector matrix is generated through Step 8.
The clustering results generated by K-means are loaded into K-NN algorithm as training data, and the 1$\times$ 300-dimensional vector matrix is loaded into K-NN algorithm to predict the classification of the search question.
Through K-means clustering and K-NN classification results, the similarity between the same group of articles and the user's search question is compared by using cosine similarity.
Reply the top five most similar LTC related articles to users.

b. Integrating semantic web into Spark cloud environment

Convert user profile into an RDF file.
Convert LTC related OD into an RDF file.
Import RDF, OWL and Rule files to start Jena inference engine.
The inference results contain LTC related resources is displayed in a mobile App.

4.7 App Demonstration

The LTC 2.0 app is a client of the developed LTC application platform (LTCAP). Figure 14(a) illustrates the login screen of the LTC 2.0 app. Its main menu contains basic information, organization recommendation, organization recommendation results, and question search tabs [Fig. 14(b)].

A user profile contains details such as name, account, password, phone number, gender, birthday, address, household income, health status, and LTC requirements [Fig. 15(a) and (b)]. After confirmation that the user has entered the correct details of the user profile, the “Send” button is to be pressed. The LTCAP converts the user profile into an RDF file required by the Jena inference engine and performs Jena inference functions with RDF, OWL, and Jena-based rules. Subsequently, after completion, an organization recommendation is displayed as a success window [Fig. 15(c)].

As a result of the recommendation of Rule 1 [Fig. 16(a)], an LTC organization is provided based on the preference of the user. As a result of the recommendation of Rule 2 [Fig. 16(b)], an LTC service is provided based on the preference of the user. As a result of the recommendation of Rule 3 [Fig. 16(c)], based on analysis of the article with keywords through the question, an LTC service is recommended based on the preference if the user.

The user enters a question as an input to interact with the LTCAP (Fig. 17). The user enters a question regarding an LTC problem [Fig. 17(a)]. The user then presses the “send” button to activate machine learning and start the Jena inference engine [Fig. 17(b)]. Consequently, the top five LTC related articles similar to the question are displayed on the screen [Fig. 17(c)].

5.1 Experimental Environment

In this study, Spark 2.2.0 and Hadoop 2.7.4 were built on seven servers respectively. One server was employed as the Mater, and the other six servers employed as the slaves. Master server is responsible for assigning all the tasks to the six slave servers. The specification of each server is given in Table 1, for example, CPU i7-2600, 3.4 GHz, 16 GB memory, 1 TB hard disk, Ubuntu16.04 LTS edition. The architecture is shown in Fig. 18. The study performs implementation and testing in two different Spark computing environments, including Spark Standalone and Spark YARN.

5.2 Cloud Computing Cluster Architecture

The cloud computing clustering architectures used in this study are Spark Standalone mode and Spark YARN mode.

Spark Standalone mode

The Spark Standard mode is composed of four components: Machine Learning, Spark Core, Standard Cluster Manager and File System, as shown in Fig. 19. Three machine learning algorithms used in Spark, namely LSI, K-means and K-NN. The Spark Standalone mode takes the build-in Standalone as the cluster manager and HDFS as the file system. Hadoop and Spark must be installed on all nodes for use, and the Spark environment must be started with instructions.

Spark YARN Mode

The Spark YARN mode is composed of four components: Machine Learning, Spark Core, YARN Cluster Manager and File System, as shown in Fig. 20. The architecture of the Spark YARN mode uses the YARN provided by Hadoop as the cluster manager. Hadoop and Spark must be installed on all nodes for use, and starting Spark YARN and HDFS file system via commands.

5.3 Spark Cluster Performance Testing

Performance testing of machine learning in LTCAP contains the LSI and K-means algorithms. The same experiment will be executed on the Spark Standalone and Spark YARN, respectively.

Test Data

A performance test was used to compare the time spent by the Spark Standalone and Spark YARN cluster managers with the machine learning algorithms for dataset sizes. Table 2 shows the data size of the corresponding data records. 1MB of data contains about 175 LTC related articles in our training dataset.

Test 1 : Spark Standalone cluster mode

K value is the main factor affecting the dimension of matrix in LSI and K-mean algorithms. Test 1 was performed in the Spark standalone cluster mode and was based on two parameters, data size and K value. The horizontal axis indicates the data size in MB, and the vertical axis indicates the testing time in minutes. Performance was evaluated based on different K values, including 100, 200, 300, 400, 500, and 1000. The test results of different K values exhibited a linear increase. The larger was the K value, the larger was the dimension of matrix generated using the LSI algorithm and the higher was the number of groups searched using the K-means algorithm, which affected work performance; thus, more time was spent. In Fig. 21, K = 100.T1 indicates that K was 100 in test 1, K = 100.T2 denotes that K was 100 in test 2, and so on.

Test 2 : Spark YARN cluster mode

Test 2 was executed in the Spark YARN mode, and the data size and K value used in this test were the same as those used in test 1. Figure 21 presents similar curve trends for different data sizes and K values. When the data size was increased, the execution time exhibited a linear increase.

Test 1 vs. Test 2

The developed Long-Term Care Application Platform (LTCAP) is tested in two different Spark cluster computing environment, including Spark Standalone (Test 1) and Spark YARN (Test 2). Based on the test results presented in Fig. 21, observations were made as follows:

Experimental results showed that for each data size and K value, Spark standalone (test 1) was faster than Spark YARN (test 2); however, the results of the two tests did not considerably differ. YARN was speculated to provide various schedulers with highly complex resource management capabilities; however, because the data complexity was not high, the allocation of resources and tasks became highly time consuming. Thus, the speed of the YARN mode was slower than that of the standalone mode.
Moreover, the experimental results indicated a notable linear relationship between the computing time and K value. For smaller k values, the fewer were the topics or clusters the LSI and K-means algorithms were divided into, the more were the data strokes that represented a certain topic or cluster; thus, different data may be divided into the same topic or cluster, which causes incomplete classification. For larger K values, the more were the topics or clusters the LSI and K-means algorithms were divided into, the fewer were the data strokes that represented a topic or cluster; thus, the data of the same topic or cluster may be disassembled. Therefore, selection of K value affected the result accuracy; thus, the K value must be a highly appropriate value.

5.4 System Evaluation

The feasibility of this study was verified by evaluating user satisfaction with a questionnaire. The questionnaire comprised 15 questions. To facilitate and simplify the process of answering the questionnaire for respondents, the answers to a question were only provided in three sets, namely K = 100, K = 300, and K = 500. For each answer set, the respondent must select one of the following options: strongly agree, agree, neutral, disagree, and strongly disagree. Figure 22 illustrates question 1, three answer sets, and five options. A total of 50 respondents completed the questionnaire. Respondents must select the K value option first and then the satisfaction option. The questionnaire statistics focused on the K value (Fig. 23). Respondents were most satisfied with the answer for K = 300, followed by K = 500 and K = 100. With question 1 an example, the percentage of respondents who believe that the answers provided by K = 100, K = 300, and K = 500 as the most satisfactory was approximately 25%, 64%, and 11%, respectively. This finding indicated that K = 300 is the most suitable for the proposed approach and is accepted by most users.

Table 3 and Fig. 24 present the satisfaction of respondents with the answers provided for each question for K = 300. The average satisfaction was approximately 4.15. The score of the tenth question was 3.8, which was slightly low; this may be because the question was considerably short to generate a useful LSI matrix for machine learning prediction. Therefore, users should provide substantial details when entering a question to prevent the LTCAP from being unable to retrieve useful keywords for generating the LSI matrix.

The Taiwanese government is promoting LTC 2.0 to improve health care for elderly people. Because of its complexity, the LTC 2.0 policy is difficult to understand and be used by the people. This study proposed to integrate a clustering algorithm into domain ontology based on cloud computing (CADOCC), which integrated statistical methods, machine learning, and semantic web technologies into a cloud-computing environment to promote intelligence in various application domains. Based on CADOCC, the study developed an LTCAP to facilitate LTC policy delivery and to recommend suitable services present in LTC to elderly people.

The provision of different questions and answers to different users is the most frequently encountered problem in LTC promotion. The proposed LTCAP provides customized questions and answers to meet various user requirements. The vector matrix of the question and answer is generated by the established LSI model and input to the K-NN algorithm to classify the user's questions. Finally, K-means clustering results and K-NN classification results are classified as the same group of articles, and the similarity is judged through Cosine Similarity. The top five articles are offered to the user.

An intelligent chatbot is a program that simulates the human dialogues for interaction with users. Chatbot has been widely used on major Instant Messenger platforms (IM platform), and these IM platforms provide easy-to-integrate webhooks allowing developers to easily develop an intelligent chatbot. At present, chatbot has found its applications in various domains[50, 51], for example health care, e-commerce, and finance. All of them have launched their chatbots to promote services. Future research will integrate the chatbot in the developed LTC 2.0 app to enhance the interaction between the LTCAP and user.

Ethics approval and consent to participate

Not Applicable.

Consent for publication

Not Applicable.

Availability of data and material

The datasets used in this paper are not publicly available due to personal information involved. However, the exclusion of personal privacy data can provide part of the data through the corresponding author.

Competing interests

The authors declare no conflict of interest.

Funding

Not applicable

Author contributions

I-Ching Hsu wrote the main manuscript text and system modeling. Shih-Hsuan Chen wrote program and test system. All authors reviewed the manuscript.

Acknowledgements

Not applicable

W. Zhu, W. Zhang, G.-Z. Li, C. He, and L. Zhang, "A study of damp-heat syndrome classification using Word2vec and TF-IDF," in Proceedings - 2016 IEEE International Conference on Bioinformatics and Biomedicine, Shenzhen, China, 2016, pp. 1415-1420.
G. L. Alexander et al., "Advancing health information technology roadmaps in long term care," International Journal of Medical Informatics, vol. 136, 2020.
H.-M. Chou, "A Collaborative Framework With Artificial Intelligence for Long-Term Care " IEEE Access, vol. 8, pp. 43657 - 43664, 2020.
F. Delmastro et al., "Long-term care: How to improve the quality of life with mobile and e-health services," in International Conference on Wireless and Mobile Computing, Networking and Communications, Pisa, Italy, 2018, pp. 12-19: IEEE Computer Society.
B. Zhou et al., "A New Remote Health-Care System Based on Moving Robot Intended for the Elderly at Home," Journal of Healthcare Engineering 2018.
A. Y. H. Liao and S.-P. Huang, "App-based Long-Term Care E-Learning Platform," in 5th International Conference on Educational Innovation through Technology, Tainan, Taiwan 2016, pp. 26-31: Institute of Electrical and Electronics Engineers Inc.
S.-C. Chen, C. Jones, and W. Moyle, "Health Professional and Workers Attitudes Towards the Use of Social Robots for Older Adults in Long-Term Care," International Journal of Social Robotics, 2019.
K.-C. Chu and P.-H. Jiang, "Visualization analysis of medical resource utilization for long-term care services " in IEEE International Conference on Information Reuse and Integration, San Diego, CA, United States, 2017, pp. 376-381: Institute of Electrical and Electronics Engineers Inc.
A. Wilkinson, T. Tong, A. Zare, M. Kanik, and M. Chignell, "Monitoring Health Status in Long Term Care Through the Use of Ambient Technologies and Serious Games," IEEE Journal of Biomedical and Health Informatics, vol. 22, no. 6, pp. 1807-1813, 2018.
H. Koezuka, "Utilization of IoT in the long-term care field in Japan," in 2nd International Conference on Cloud Computing and Internet of Things, Dalian, China, 2016, pp. 166-169: Institute of Electrical and Electronics Engineers Inc.
A. Cid, R. Sotelo, M. Leguisamo, and M. Ramírez-Michelena, "Tablets for deeply disadvantaged older adults: Challenges in long-term care facilities," International Journal of Human Computer Studies, vol. 144, 2020.
W. Li et al., "A Comprehensive Survey on Machine Learning-Based Big Data Analytics for IoT-Enabled Smart Healthcare System," Mobile Networks and Applications, vol. 26, no. 1, pp. 234-252, 2021.
K. J. Moon, C.-S. Son, J.-H. Lee, and M. Park, "The development of a web-based app employing machine learning for delirium prevention in long-term care facilities in South Korea," BMC Medical Informatics and Decision Making, no. 22:220, 2022.
A. Abella, M. Ortiz-de-Urbina-Criado, and C. De-Pablos-Heredero, "The process of open data publication and reuse," Journal of the Association for Information Science and Technology vol. 70, no. 3, pp. 296-300, 2019.
I.-C. Hsu, "XML-Based Information Fusion Architecture Based on Cloud Computing Ecosystem," Computers, Materials & Continua, vol. 61, no. 3, pp. 929-950, 2019.
R. Zhang, "The impacts of cloud computing architecture on cloud service performance," Journal of Computer Information Systems, vol. 60, no. 2, pp. 166-174, 2020.
W. Qiang, "Performance and security in cloud computing," Journal of Supercomputing vol. 75, no. 1, 2019.
Y. Zhao, M. Ren, S. Jiang, G. Zhu, and H. Xiong, "An efficient and revocable storage CP-ABE scheme in the cloud computing," Computing, vol. 101, no. 8, pp. 1041-1065, 2019.
H. Zhao, J. Zheng, W. Deng, and Y. Song, "Semi-Supervised Broad Learning System Based on Manifold Regularization and Broad Network," IEEE Transactions on Circuits and Systems I: Regular Papers vol. 67, no. 3, pp. 983-994, 2020.
W. Deng, H. Liu, J. Xu, H. Zhao, and Y. Song, "An Improved Quantum-Inspired Differential Evolution Algorithm for Deep Belief Network," IEEE Transactions on Instrumentation and Measurement, 2020.
G. V. Demarie and D. Sabia, "A machine learning approach for the automatic long-term structural health monitoring," Structural Health Monitoring vol. 18, no. 3, pp. 819-837, 2019.
S. Mukhopadhyay, "InfoGram and admissible machine learning," Machine Learning, vol. 111, no. 1, pp. 205-242, 2022.
L. Miranda, J. Viterbo, and F. Bernardini, "A survey on the use of machine learning methods in context-aware middlewares for human activity recognition," Artificial Intelligence Review, vol. 55, no. 4, p. Artificial Intelligence Review, 2022.
S. Saeedi, S. Rezayi, H. Keshavarz, and S. R. N. Kalhori, "MRI-based brain tumor detection using convolutional deep learning methods and chosen machine learning techniques," BMC Medical Informatics and Decision Making, no. 23:16, 2023.
N. Liang, C. Wang, S. Li, X. Xie, J. Lin, and W. Zhong, "The classification of flash visual evoked potential based on deep learning," BMC Medical Informatics and Decision Making, no. 23:13, 2023.
M. Lnenicka⁎ and J. Komarkova, "Big and open linked data analytics ecosystem: Theoretical background and essential elements," Government Information Quarterly, vol. 36, pp. 129-144, 2019.
K. Mansouri, A. Alti, P. Roose, and S. Laborie, "Dynamic semantic-based green bio-inspired approach for optimizing energy and cloud services qualities " Transactions on Emerging Telecommunications Technologies, vol. 29, no. 5, 2018.
F. Orciuoli and M. Parente, "An ontology-driven context-aware recommender system for indoor shopping based on cellular automata," Journal of Ambient Intelligence and Humanized Computing, vol. 8, no. 6, pp. 937-955, 2017.
M. U. Bokhari, Q. M. Shallal, and Y. K. Tamandani, "Reducing the Required Time and Power for Data Encryption and Decryption Using K-NN Machine Learning," IETE Journal of Research, no. January 1, 2018 pp. 1-9, 2018.
I. Portugal, P. Alencar, and D. Cowan, "The use of machine learning algorithms in recommender systems: A systematic review," Expert Systems with Applications, vol. 97, pp. 205-227, 2018.
I.-C. Chen and I.-C. Hsu, "Open Taiwan Government data recommendation platform using DBpedia and Semantic Web based on cloud computing," International Journal of Web Information Systems, vol. 15, no. 2, pp. 236-254, 2018.
Y. Bao, Z. Tang, and H. Li, "Compressive-sensing data reconstruction for structural health monitoring: a machine-learning approach," Structural Health Monitoring, vol. 19, no. 1, pp. 293-304, 2020.
R. Kebede, A. Moscati, H. Tan, and P. Johansson, "Integration of manufacturers' product data in BIM platforms using semantic web technologies," Automation in Construction, vol. 144, 2022.
C. Peng, X. Zhou, and S. Liu, "An Introduction to Artificial Intelligence and Machine Learning for Online Education," Mobile Networks and Applications, vol. 27, no. 3, pp. 1147-1150, 2022.
A. Verma, A. H. Mansuri, and N. Jain, "Big Data Management Processing with Hadoop MapReduce and Spark Technology: A Comparison," presented at the Colossal Data Analysis and Networking (CDAN), Indore, India 18-19 March 2016 2016.
S. Ramirez-Gallego et al., "An Information Theory-Based Feature Selection Framework for Big Data under Apache Spark," IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 48, no. 9, pp. 1441-1453, 2018.
D. Sobhy, Y. El-Sonbaty, and M. A. Elnasr, "MedCloud: healthcare cloud computing system," in 2012 International Conference for Internet Technology and Secured Transactions, 2012, pp. 161-166: IEEE.
P. Merla and Y. Liang, "Data analysis using hadoop MapReduce environment," in 2017 IEEE International Conference on Big Data (Big Data), 2017, pp. 4783-4785: IEEE.
A. V. Hazarika, G. J. S. R. Ram, and E. Jain, "Performance comparision of Hadoop and spark engine," in 2017 International Conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud)(I-SMAC), 2017, pp. 671-674: IEEE.
Jieba. (2020). Traditional Chinese text segmentation. Available: https://github.com/ldkrsi/jieba-zh_TW
Wikipedia. (2019). TF-IDF. Available: https://en.wikipedia.org/wiki/Tf%E2%80%93idf
N. García-Pedrajas, J. A. R. del Castillo, G. J. I. t. o. n. n. Cerruela-García, and l. systems, "A Proposal for Local $ k $ Values for $ k $-Nearest Neighbor Rule," vol. 28, no. 2, pp. 470-475, 2017.
Wikipeda. (2020, May 20, 2020). Latent Semantic Indexing. Available: https://en.wikipedia.org/wiki/Latent_semantic_analysis
N. Akter, A. S. Hoque, R. Mustafa, and M. S. Chowdhury, "Accuracy analysis of recommendation system using singular value decomposition," in 2016 19th International Conference on Computer and Information Technology (ICCIT), 2016, pp. 405-408: IEEE.
Y. Wang and L. Zhu, "Research and implementation of SVD in machine learning," in 16th IEEE/ACIS International Conference on Computer and Information Science, ICIS 2017 Wuhan, China 2017, pp. 471-475: Institute of Electrical and Electronics Engineers Inc. .
I.-C. Hsu, "Integrating Ontology Technology with Folksonomies for Personalized Social Tag Recommendation," Applied Soft Computing, vol. 13, no. 8, pp. 3745-3750, 2013.
J. J. P. c. s. Anuradha, "A brief introduction on Big Data 5Vs characteristics and Hadoop technology," vol. 48, pp. 319-324, 2015.
Apache Hadoop YARN. Available: https://hadoop.apache.org/docs/current/hadoop-yarn/hadoop-yarn-site/YARN.html
Apache. (2017, Sep. 14, 2018). Apache Spark. Available: http://spark.apache.org/
T. Okuda and S. Shoda, "AI-based chatbot service for financial industry," Fujitsu Scientific and Technical Journal vol. 54, no. 2, pp. 4-8, 2019.
M. Bates, "Health Care Chatbots Are Here to Help," IEEE Pulse, vol. 10, no. 3, pp. 12-14, 2019.

Tables 1 to 3 are available in the Supplementary Files section.

No competing interests reported.

floatimage18.png
Table 1. Specifications of Spark cluster
floatimage22.png
Table 2. Test data size
floatimage27.png
Table 3. Satisfaction statistics based on K=300

Download PDF

Version 1

posted

You are reading this latest preprint version

An Intelligent Long-Term Care Information Platform using Machine Learning and Semantic Web

Status:

Version 1

Abstract

Figures

1. INTRODUCTION

2. Related Work

3. System Architecture and Design

4. System Implementation

4.1 Data Collection and Preprocessing

4.2 Term Frequency–Inverse Document Frequency

4.3 Machine Learning

4.4 Cosine Similarity

4.5 Semantic Web

4.6 Integrating Machine Learning and Semantic Web into Cloud Computing

4.7 App Demonstration

5. System Experiment and Evaluation

6. CONCLUSION

Declarations

References

Tables

Additional Declarations

Supplementary Files

Status:

Version 1