Intelligent health monitoring system modeling based on machine learning and agent technology

Abstract

The use of machine learning approach in the field of health informatics improves the quality and effectiveness of decision-making process. Its particular integration into the Remote Health Monitoring Systems (RHMS) allows: collecting, analyzing and learning from real-time data, automatically gaining knowledge and making predictions on the patients state. The RHMS represent an effective solution to control and monitor a growing number of dependent or elderly patients. They showed impressive results in healthcare applications. These systems are built around the sensors installed on the body of the patient and others embedded in their environments. In this context, we are confronted with the challenge of temporal aspect of data. To meet this challenge, we propose that RHMS combine machine learning for generating intelligent valuable information, and visual analytics for gaining insight the collected real-time data. The success of these applications is based on the quality of their design and development. For this reason, we propose to design a RHMS using multi-agent technology. The developed prototype was evaluated to verify our proposal feasibility.

Keywords

Remote health monitoring system machine learning visualization data mining multi-agent system

1. Introduction

Nowadays remote healthcare monitoring is one of the most interesting research areas. Literature review has shown that this area is difficult to address due to its complexity [1]. Most traditional systems used to monitor patients in hospitals take time, such as the face-to-face consultation, the physicians’ availability, etc. These systems are, in general, based on classical computing paradigms, which are insufficient to model the interactivity and complexity of its environment [2, 3, 4]. With the rapid and successive progress in the technological tools, the networks of body sensors and the connected objects [5], Machine learning techniques must be integrated to provide and real-time remote surveillance of dependent or elderly patient [1].

Machine Learning approach consists of analyzing the real-time patient data for detecting risk situations and generating appropriate decisions. It proposes algorithms that learn from data and improve through experience [6]. Automatically generated knowledge can be integrated for decision-making tasks. In literature, automatic machine learning approaches presents interesting results in several Health Informatics studies [7]. It is a sub-field of data science. It aims to design Supervised Learning and Unsupervised Learning algorithms for making predictions. These algorithms allow discovering data sets properties, particular in real-time. Thus, Machine learning can be used for data mining. We talk about data mining based on Remote Healthcare Monitoring System (RHMS) [8].

Another interesting approach for making predictions is simply looking and interpreting the data sets using visualization techniques. It is an interactive visual analytics that is commonly named as visual data mining. Here interactive machine learning can be used. It consists of defining algorithms to interact with humans (decision-makers) and to improve their cognitive learning ability through Human-Computer Interactions [9].

Our context is then in the intersection of all these fields (i.e., RHMS data mining, visualization, automatic and interactive machine learning): it is the visual data mining based RHMS.

Another challenge of RHMS is to offer advanced machine learning services and to guarantee confidence, confidentiality and security of communications [10, 11]. The interdisciplinary and communication devices heterogeneity of such system leads to the complexity the decision-making tasks. To overcome such complexity, the multi-agent modeling is a useful solution.

We aim, by using the multi-agent technology to provide self-intelligent, organized, distributed and collaborative remote patient monitoring tasks [12, 13]. The Multi-agent system is used to model the dynamic and synchronous communication between different processes of decision making in the Health homecare envisaged architecture. To make our developed tool more intelligent, the data mining process is dividing into sub-tasks in our case. For each one we assigned a specific agent able to realize the appropriate treatments. The defined agents collaborate together ensure knowledge extraction from data. Different steps of data mining can be realized on the same time thinks to communication capacity of agents. Furthermore, knowledge is acquired, processed, learned and exchanged autonomously and synchronously. In fact, developing several agents might speed up the RHMS’s tasks by providing automatic and interactive machine learning methods for parallel computation. Such parallelism can aid to deal with the real-time reasoning requirements.

In this research work, we propose a multi-agent RHMS architecture that integrates machine learning by applying visual data mining techniques for combining artificial intelligence with human intelligence.

The remainder of this paper is organized as follows. Section 2 treats the theoretical background of our research: RHMS, visual data mining and multi-agent technology. Concerning Section 3, it describes our proposed architecture for developing RHMS based on vital and environments sensors. This architecture introduces the system into three layers. Concerning Sections 4 and 5, they respectively provide an implementation and experimental verification of the proposed architecture. Finally, Section 6 presents the conclusion and suggestions for future work.

2. Literature review

2.1 Remote healthcare monitoring system

Remote Healthcare Monitoring System (RHMS) is a decision support system that aims to collect medical real-time data to monitor patients outside of conventional clinical settings (e.g., in Home). It is a solution that allows decreasing the number of emergency services and hospitalizations. Even patients feel more comfortable knowing that they are being monitored and are supported in case of health problem [14, 15].

In recent years, a growing interest of the RHMS development for elderly and dependent patients is emerging. In fact, they are more vulnerable to diseases because of their physical conditions. So, they require continuous control and surveillance of their vital parameters by healthcare providers [16, 17].

The importance of developing a RHMS is twofold:

•
The constant remote monitoring can provide better data compared to the “snapshot” monitoring which runs at a clinical site. This allows a longer time scale and further granularity of health data monitoring,
•
The RHMS could assist to avoid a false or disrupted reading of the health data by the “white coat syndrome” in a clinical site [18].

The RHMS ability would assist health professionals (i.e., remote monitor and decision-makers) with diagnosis of the patient’s conditions, not simply a care of him/her. Such systems are built around vital signs and home environmental sensors to collect data: (i) vital signs sensors are fixed to the patient’s body to take measures (bio-data, temperature, heart rate, blood pressure) and data (ii) domestic environmental sensors are connected to a system bus from the house (business motion and transitions).

These collected data are stored in a temporal database to be analyzed using the data mining and visualization technologies. The integration of the data mining in the RHMS can considerably improve the patient remote monitoring by enabling the combination of human knowledge initialized by medical experts and Knowledge extracted from the collected data [19, 20] (cf. Fig. 1)

Figure 1.
Remote healthcare monitoring system architecture.

2.2 Visual data mining

RHMS relies on collecting data using various sensors to detect and interpret changes in a patient’s state. The analysis process consists of reliable processing of real-time data, detection of damage and prediction of critical situations. This process requires observation of the data over a period of time to determine the caused damage [21]. Continuous patient monitoring provides the accumulation of a large amount of these data. We aim to treat this quantity of temporal data in order to improve the detection and the prediction ability of the RHMS [22, 23]. Such processing can be realized by developing an intelligent RHMS based on the data mining technology, which makes it possible to optimize, in real time and without false alarms, the prediction of the damage.

RHMS integrates data mining to analyze the collected data for real-time prediction [24]. Data mining is the process of discovering novel, valid, useful and potentially interesting patterns from data. It takes place into four phases: data pre-treatment, which consists of the selection, cleaning and transformation of data from its raw format into a format appropriate for mining, data mining, which consists of applying techniques on the pretreated data to discover interesting patterns, Model evaluation, which aims to interpret and validate the extracted patterns to generate knowledge from these patterns and, knowledge integration, which helps user to use of the generated knowledge to take decisions.

Data mining draws ideas from machine learning, pattern recognition, statics and database systems [25]. Among these fields, we are particularly interested by machine learning techniques for data mining (i.e., the second step of the data mining process). This automatic machine learning allows applying algorithms that learn from experience [26]. It makes the RHMS intelligent enough to learn by itself (cf. Fig. 2).

Figure 2.

Data mining process integrating automatic machine learning.

Data mining based on automatic machine learning algorithms improves through experience from large temporal data sets and presents impressive results [27]. However, currently, the amount of information is greater than ever, hence its analysis and exploration becomes more difficult. To overcome this difficulty, a new decision-making approach has emerged: it is the visual data mining [28, 29] (cf. Fig. 3). We talk about a decisional system that combines machine capabilities with human intelligence. This approach allows visualizing and interpreting raw data, pretreated date, extracted patterns, generated knowledge and integrated decision. It aims to integrate the user throughout the intelligent decision-making process. It is an interactive machine learning that helps to enhance the decision-maker cognitive learning ability through the interactions with the visualizations [30, 31].

Figure 3.

Data mining process integrating automatic and interactive machine learning.

To model and develop a visual data mining based RHMS, multi-agent approach is an interesting solution for two main reasons:

(i)

The RHMS (as a Decision Support System) based on the visual data mining (that applies automatic and interactive machine learning algorithms) is characterized by its interdisciplinary aspect, which makes its architecture more complex. The multi-agents approach can deal with this complexity by assigning for each step of data mining process an agent that will ensure the generation of a result according o its specified target. This strategy makes communication and data exchanges between different steps of data mining process faster. The quality of results supplied by each step can be improved by using agent technology given the capacity of this later to face the complexity of treatment that can be due to the large amount of data within a short time.

(ii)

RHMS involves several sensors, communication devices and actors (patients, physicians, health providers and medical centers) that are geographically dispersed. This system must enable these actors to visualize data and knowledge, check patient medical status, diagnose patient and take decisions. We are interested then in distributed artificial intelligence. The multi-agents approach can deal with this intelligence. The agent responsible for data, pattern and knowledge visualization generates a graphical representation according to; the user preference by giving him/her the ability to choose the visualization technique that he/she likes to be applied from the one hand and from the other hand, system agents must be able to differentiate the user profile to generate an appropriate representation of data according to the given permission. Different agents that collaborate together to provide a graphical representation should respect data integrity and confidentiality.

2.3 Multi-agent technology

In the past few years, the domain of remote monitoring based on the multi-agent approach [32] is one of the most studied health informatics methodologies and applications in Artificial Intelligence (AI). This approach can work in heterogeneous environments to improve the decision-making process.

An agent is a physical or virtual entity, which aims at satisfying its needs and objectives on the basis of all other elements (perceptions, representations, actions, communications and resources) at its disposal. The execution process of an agent consists of a set of modules (cf. Fig. 4): (a) environment perception, (b) event detecting, (c) applying machine learning algorithms for generating knowledge for decision-making, (d) executing the taken decision (as an action), (e) communicating the execution result to other agent(s) to a achieve a collective goal, and (f) acting on the external environment.

Figure 4.

The execution process of an agent.

Given the high capacity of agent technology in perception and communication with the different components of its environment as it is showed previously and by relying on the success of previous works dealing with agents based architectures that modeled distributed and complex environment. We have judge that multi-agent system can facilitate the communication and the collaboration between the heterogonous components of our envisaged system. The visual data mining tasks to be performed by a RHMS are complex and need to be decomposed into subtasks. Multi-agents modeling are considered as important for the following reasons [33, 34]: each agent performs a set of actions to modify its decision-making environment. It is autonomous and has the possibility of responding with requests or refusals to requests coming from other agents. The “cognitive” nature of agents helps them to perform their tasks and to manage the interactions with the other agents and with their environment. In this case, each agent has its knowledge base that includes all information and necessary “know-how” to accomplish its goals. However, the “reactive” nature of agents allows them to act only in response to signal, stimulation and perception. They have a knowledge base containing set of $<$ condition, action $>$ rules.

The multi-agent modeling of a RHMS aims to achieve the critical decision support tasks considered previously as based on automatic and interactive machine learning. It allows; (1) the automation of the repetitive tasks, (2) the extraction of full information from complex real-time data and (3) the making of the appropriate recommendations to the remote monitors concerning a specific action by exploiting certain prior knowledge of the user’s objectives [35]. The following section presents our proposal for visual data mining based RHMS modeling using multi-agent technology.

As we note, nowadays, there are a fast technological progress in the world that looks for making person live easier and facilitating distant monitoring and communication. As consequence, most of domains integrate new sensors technology to enable improvement of their owner information systems. Health homecare is one of those domains, so that, multiple works which are based on sensors and internet of things [36] dealing with homecare monitoring systems are proposed in literature. The main goal of those approaches is to make possible and easy patients’ supervising at home. Those works look for distant clinical data collection relying on sensors installed at patient’s home and/or wearable sensors. Among the relevant proposed related works, we can note, [37, 38, 39] that proposed different tools for health monitoring in order to control patient’s state by obtaining distant information about a set of parameters such as; body temperature, blood pressure and other clinical measures. As consequence, we concluded that most of sensors based homecare monitoring proposed works, focus on data collection but they neglect in most of time, the synchronization between data collection, data mining and knowledge extraction from data steps which are considered as principal points of decision making process. They failed also in some works through introduced tools, to make possible the visualization of data, information and knowledge according to users’ profiles in different steps of decision making process from the one hand and according to the user preference by giving him/her the ability to select the visualization technique that he/she likes to be applied by the other hand. Starting from those points of weakness that negatively influence on decision making process and from our previous hypothesis proposed by our teams in previous works [40, 41] which proved that intelligent Dynamic Decision Support System (iDSS) should be built on three main axis which are; first is the integration of intelligent data mining and second is the integration of visualization techniques, and third is the end-users integration by respecting a set of Humain Computer Interaction (HCI) requirement [42] that leads to obtain relevant proposed solutions given by the system, which are near from the Humain resonant. All those hypothesis are proved by the obtain of encouraging results in real hospital case study for nosocomial infection detection in CHU Habib Bourguiba hospital of Sfax, Tunisia [40, 41]. For all those reasons, we are thinking about enhancing our previous proposed iDSS architectures in our previous works [43], by extending and improving old suggested architectures in order to be applied in homecare monitoring case.

3. Multi-agent modeling of visual data mining based RHMS

This section is organized into two sub-sections. The first one introduces the proposed RHMS architecture. The second sub-section describes the different layers of this architecture.

3.1 System architecture

Figure 5 presents an overview of the visual data mining based RHMS architecture. Its components that must work together in multi-agents environment to provide healthcare services in hospital and homecare environments can be organized into three parts: real-time data acquisition, data processing and knowledge integration and communication. We propose then three layers respectively to the three identified parts: sensing monitoring layer, data analysis layer and decision-making layer (cf. Fig. 5).

The process of the RHMS begins with the acquisition and storage of data using biosensors and sends data to monitoring devices via communication protocols. Then the collected real-time data will be stored in a temporal database, to trigger subsequent processes. The process continues by involving interactive and machine learning algorithms for visualizing prepared data, understanding them, analyzing them, extracting interesting knowledge and progressively learning from them to generate assistance and guidance to the patient under supervision.

Modeling our architecture using multi-agent technology begins by identifying and assigning agents to each task of the visual data mining based RHMS. Figure 5 shows the identified cognitive and reactive agents integrated in each layer of our system. It models the statistical relations between different agents that belong to each layer.

Table 1
Agents of the sensing monitoring layer (data collection layer)

Agent	Nature	Description
Remote monitor	Cognitive	This agent is responsible for: (1) monitoring the specific parameters of the vital signs and the home environmental sensors and managing the different remote sensing agent, (2) executing a set of data signal processing tasks (demodulation, synchronization, decoding, etc.) to pull out the principal features from the sensing signal, and (3) estimation of the failure position by the precision of critical damage data (if exist).
Patient data base	Cognitive	It collected and treated data from remote monitoring agent different sensors and save them in the patient database. as consequence patients can view the recorded data from them Smartphone.
Network monitor	Cognitive	Responsible for network managing and data security to provide reliable means of confidentiality, security and privacy for the patient by verifying the access based on users accounts. This ensures also the most common network management such as interruption, injection, and interference by realising a set of tests management continuously through sending assessing request.

Figure 5.

Architecture of the machine learning based RHMS using multi-agent system.

3.2 RHMS layers

3.2.1 Sensing monitoring layer

To collect real-time medical data about a patient state, two types of sensors are used to enable simultaneous monitoring: (i) environmental sensors, which are located throughout the home to monitor the patient transitions movement, (ii) the vital signs sensors, which are attached to the patient’s body to collect real-time measuring parameters of our monitoring system (i.e., temperature, oxygen saturation, heart rate and blood pressure parameters). To take the appropriate decision in real time concerning patient’s timely diagnosis and treatment, this data is transformed in understandable characteristics by attributing specific agents (cf. Table 1).

As presented by Fig. 5, the three layers of our RHMS architecture involve two common agents: coordinator and visualization agents. Table 2 provides their description.

During the execution of the various tasks of our RHMS layers, several problems may occur with the sensors or the mobile devices, such as: loss of data, drainage of the battery, noise in data, etc. For this, we propose common failure management agents presented by Table 2.

Table 2
Common agents

Agent	Nature	Description
Visualization	Cognitive	It generates the visual representations of the real-time raw and prepared data as well as the extracted models (by the data mining agent). This agent applies one of a set of visualization techniques such as tree representation, spiral representation or life line time representation according to user choice. The graphical representations given by each technique introduce graphical representations that make interaction easier then textual representation cases. It makes also more quick and efficient decision making for experts.
Coordinator	Cognitive	It coordinates, negotiates and collaborates with other agents for resolving conflicts, time synchronization and distributing resources. For this raison we have specified a road map of interactions through a set of possible actions and scenarios that coordination agent should follow in order to ensure an intelligent coordination between different systems’ agents to ensure a permanent and efficient patient monitoring and to initiate the correct and specific actions. This agent has two main goals; ensures agents collaboration in order to satisfy a demanded service or to delegate actions in consequence of normal or abnormal environmental situation.

The interactions between the first layer agents are presented by the sequence diagram visible in Fig. 6.

Figure 6.

Sequence diagram for modeling the agents’ interactions.

Occasionally some failures or errors can be occurred during an agent’s activity, the user will be informed by an alarm sent by the alarm monitor agent and may ask the system for a report on the problem that has occurred by the report monitor agent. Providing from the given report, different system agents’ learns from previous failures and wrong decision suggestions. The sequence diagram shown in Fig. 7 describes the interactions between involved agents and started actions in case of failure transmission and synchronization problems.

Figure 7.

Sequence diagram for modeling the agents’ interactions in failure transmission.

3.2.2 Data analysis layer

After performing the various tasks of the first layer correctly, comes the role of the data analysis layer. It begins with storing the collected real-time data in a database. It consists of assigning a set of tasks to intelligent agents in order to be able to suggest decisions for experts. The different tasks assigned for environment agents’ are those of Knowledge from Data Discovery (KDD) process. The data treatment and analysis tasks follow visual data mining steps presented in Fig. 3. The latter is based on four principal sub-tasks; data preparation, data mining, model evaluation and knowledge integration. In this interactive process, automatic machine learning methods can be applied to provide accurate predictions and decisions. Machine learning uses: (a) statistical automatic learning to identify patterns from data by applying decision tree algorithm and, (b) interactive learning using visualization, this step relies on one of visual techniques mentioned previously, it facilitates understanding and interpreting the output of automatic algorithms (classification, dimensionality reduction, etc.) generation implemented in our case. The different involved agents that belong to data analysis layer are described in Table 3.

Table 3
Agents of the data analysis layer (data collection layer)

Agent	Nature	Description
Database agent	Cognitive	It is responsible for grouping data values from sensors and placed them in the database. The stored data must have the same format. Given the diversity of data types, the agent must carefully make all data in the same format, so this agent transforms them in textual data. The fusion algorithm proposed by $\ldots$ is applied in our case. The missing data are treated by the database agent and the noise is cleaned. The main goal of the database agent is to save data of the vital signs and home environmental signs after treatment. It prepares data for the mining task.
Data mining	Cognitive	The main principal task of this agent is to generate patterns from transformed data. Those patterns are obtained by executing the data mining technique extended from machine learning algorithms. The technique applied in our case is the decision tree. During this process, we propose to use machine Learning algorithms. This agent has the ability to extract or retrieve the relevant data from the database and automatically alert the remote monitors to the presence of the selected information. Finally, this agent aims to evaluate the ability of the extracted models to predict the patient’s condition. It executes a set of quantitative and qualitative methods of evaluation proposed by the data-mining field. If these extracted models are validated, they can be displayed to give the remote monitor an idea of the result of the data analysis. If not, an alarm will be sent to the system to correct failure.

Different interactions that can occur between agents that belong to second layer that aim to generate patterns from temporal patient data are presented by the sequence diagram visible in Fig. 8.

Figure 8.

Interaction diagram for modeling the agents’ interactions in data analysis layer.

3.2.3 Decision-making layer

To take the appropriate decision, the RHMS must learn from the validated patterns received from the data analysis layer. These patterns must be stored in the database as a new knowledge related to a set of variables (defined by the machine learning algorithm). Such knowledge must be integrated to generate decision recommendations. These tasks are assigned to as set of corresponding agents presented in Table 4.

Table 4
Agents of the decision making layer

Agent	Nature	Description
Decision	Cognitive	Two principal tasks are given to this agent: (i) Using validated patterns for decision-making. It aims to extract knowledge issued from these patterns. The knowledge extraction task is based on knowledge base that contains over than one hundred rules specified on collaboration with physicians. As consequence of applying the knowledge rules, this agent produces decision recommendations. These recommendations can be interactively displayed to the remote monitor by applying the visualization agent. (ii) The second task of this agent is to learn from validate decision obtained after a set of negotiation and interaction between Decision agent and doctor. This communication between the two previous actors is done by the mean of remote monitor and to store it in the database to enrich the knowledge base of the RHMS. According to the taken decision, this agent sets and requests the visualization agent to display the actions to be done by the health providers.
Remote monitor interface	Cognitive	The role of this agent is to display the data required for each monitored patient according to his/her preference by providing a set of visual interface from what user can select the way of data representation. This agent helps to adapt application interface to the user desire in order to make easy the Human-Computer Interactions (HCI).

Agent

Nature

Description

Decision

Cognitive

Two principal tasks are given to this agent:

(i)

Using validated patterns for decision-making. It aims to extract knowledge issued from these patterns. The knowledge extraction task is based on knowledge base that contains over than one hundred rules specified on collaboration with physicians. As consequence of applying the knowledge rules, this agent produces decision recommendations. These recommendations can be interactively displayed to the remote monitor by applying the visualization agent.

(ii)

The second task of this agent is to learn from validate decision obtained after a set of negotiation and interaction between Decision agent and doctor. This communication between the two previous actors is done by the mean of remote monitor and to store it in the database to enrich the knowledge base of the RHMS. According to the taken decision, this agent sets and requests the visualization agent to display the actions to be done by the health providers.

Remote monitor interface

Cognitive

The role of this agent is to display the data required for each monitored patient according to his/her preference by providing a set of visual interface from what user can select the way of data representation. This agent helps to adapt application interface to the user desire in order to make easy the Human-Computer Interactions (HCI).

The sequence diagram visible in Fig. 9 shows the interactions between the different agents of decision making layer.

Figure 9.

Sequence diagram for modeling the agents’ interactions in data decision-making layer.

The development of all presented agents was done using the Jade platform (JADE). The Following section presents the implementation process and results in order to validate our proposed architecture and machine learning solutions (visual data mining approach).

4. Prototype implementation

4.1 Sensing monitoring layer

To monitor a patient, we should control vital signal of him/her continuously. The implemented visual data mining based RHMS based in architecture introduced in previous sections is based in real case on multiple sensors for a permanent monitoring of several patients on the same time. The use of several sensors at the same time to collect required data enhances the applicability of the developed system.

The main goal of our work is to develop and to setup a real-time remote RHMS for elderly and dependent patients distant monitoring. The proposed tool measures several vital signals such as heart rate, blood pressure, body temperature, oxygen saturation, motion, etc. using a set of diver sensors. Each sensor characterized by unique identification, collects patients’ data continuously. The different kind of sensors that we have used in the real case of patients’ distant monitoring are presented in Table 5.

Table 5
Various physiological wearable body sensors

Parameter	Description	Observation
Temperature	Thermometer that measures the body temperature	36.5 ${}^{\circ}$ C–37.7 ${}^{\circ}$ C
Blood pressure	Tensiometer for blood pressure monitoring	Systolic: 100–150 mmHg Diastolic: 60–90 mmHg
Heart rate	Heart device for monitoring Frequency of cardiac cycle	60–100 BPM
Oxygen saturation	Blood oxygen monitoring for monitoring a person’s oxygen saturation (SO2)	Above 95%

Our prototype aims to assist decision-makers (i.e., physicians) to visualize dynamic and real time patient data and to supervise his/her health status and to interact with the monitoring tool through simple interface easy to manipulate. The results delivered by each layer are displayed separately from others respecting the architecture proposed by Leite et al. [44]. One of positive point of our proposed tool is to present data in clear textual form instead of signal and symbols representation in order to facilitate it interpretation (C.f. Fig. 10).

Figure 10.

From the numeric representation to a visualization prototype.

The visual representation can be enhanced by dynamically applying the interaction mechanisms (for example zooming and filtering in Fig. 11). These interactions modes fall with the scope of interactive machine learning to better analyze and interpret the patient health situation.

Figure 11.

Filtering and zooming mechanisms.

Sometimes it is useless to display all the parameters. In some cases, physician can select only that he/she believes that there are useful for patient monitoring, as consequence, he/she may select only the needed parameter to visualize (cf. Fig. 12).

Figure 12.

Interface for visualizing only the value of temperature and heart rate in corresponding time choosing.

4.2 Data analysis layer implementation

After the previous step of collecting and storing data in the database, many treatments will be done by specific delegate agents (cf. Fig. 8) to prepare these data for further data mining applications. The process of data preparation technique that we have implemented performs the following tasks: (1) formatting the input raw data according to its type, (2) processing missing data values and (3) selection attributes and instances appropriate to execute our data mining algorithm as it was explained with details in previous sections. It is an interactive machine learning process that allows representing raw data in a clear dynamic graphical visualization to allow user to gain insight and dynamically interact with it, i.e., selecting the visual technique to apply, label and zoom, etc.).

The data-mining agent implemented in our RHMS prototype is the decision tree [45, 46]. This machine learning technique is chosen because it can be considered as:

(i)
Automatic machine learning technique that uses algorithm to predict the value of an outcome variable (represented by leafs) based on numerous input variables (represented by interior nodes) in order to detect possible patient state deterioration. The path from the root to the leaf throughout interior nodes describes the patient diagnosis (cf. Fig. 13).
(ii)
Interactive machine learning technique that can be used to visualize decisions. It is simple representation for classification. It describes classification approach applied and allows the remote monitor to interactively understand and interpret the diagnosis and cognitively learn and acquire knowledge.

Moreover, we have implemented the model evaluation agent. It is an algorithm that allows evaluating the performance and predictive ability of the knowledge extracted by the decision tree. This algorithm allows the remote-monitor to visually interact with the decision tree by changing values, filtering paths, etc. Such interaction helps remote monitor to estimate the importance of the extracted knowledge.

Figure 13.
Decision tree visualization.

4.3 Decision making layer implementation

Validated patterns, which are received from the data analysis layer, have to be integrated to produce the suitable recommendations for the decision-making activity. The knowledge integration agent generates two kinds of recommendations:

•
An immediate alert to the decision agent if a critical damage is detected with the decision tree. If the result is high then the decision agent proposes a treatment for a timely intervention.
•
A test and medication reminder alerts.

The decision can be interactively displayed to the remote monitor by sending a request to the visualization agent. The role of the remote monitor interface agent is to present these representations to the remote monitor and to receive the requested instructions from him/her. These representations offer doctors the possibility to directly interpret the patient’s condition by visualizing the received parameters with different colors: the normal values in green and those of alert cases in red (cf. Fig. 14).
5. Evaluation and discussion

Based on the learning ability of the machine learning techniques and the agent entity, the execution time of each layer increases in case of the current version. We have observed this gain by comparing it to the results of the old version. The Table 6 showed the obtained results in case of using 25MO of data.

Table 6
A comparative table of current and old systems execution time

	Execution duration
	Current version	Old version [47]
Sensing monitoring	5 s	10 s
Data analysis	14 s	15 s
Decision-making	2 s	6 s

Figure 14.

Remote monitor interface.

Thanks to our proposed approach (machine learning based RHMS), considerable saving of time has been achieved. This gain is over the 15 sec, which can be considered as interesting results, mainly in case of real time environment. The rapidity of treatment obtained in different modules; sensing monitoring, data analysis and decision-making and consequently the considerable gain of time respond are due to the intelligence and the high performance of integrated agents. The obtained results proved that developed agents in our case are able to negotiate and collaborate together in order to give relevant solutions in a quick time despite the complexity of data treatment due to heterogeneous sources.

However, it is difficult for health providers to deal with complex and uncertain problems and automatically make decisions in real time without continuous monitoring. The analysis of temporal data acquired without a history is insufficient [48]. In this study, we have proposed a multi-agent modeling and development of a real-time monitoring system based on interactive and automatic machine learning methods is developed for elderly and dependent patient. Several parameters such as oxygen saturation, blood pressure, temperature and heart rate are collected using wearable sensors. This patient’s medical data undergoes a machine-learning package before being visible to the physician to make a right decision.

The use of multi-agent approach presents interesting advantages and proves the feasibility and adequacy of using agent technology to solve our complex and distributed healthcare monitoring problem. This modeling aims to guarantee flexibility, adaptability to environment changes, gain in response time, reliability and management of the access to the patient data and knowledge: advantages of multi-agent approach.

Despite the improvements made by our system, there is always more to be done and many enhancement to plan. Many points could be taken into account to gap between academic prototypes and systems used in real settings:

•

Health providers, the users of our RHMS, should be involved from the beginning in the development of the project. Particularly in the design of user interfaces with which feel comfortable, ensuring auto-adapting Human Computer-Interaction.

•

Integrating new users to our RHMS who can be the relatives of the patient to be monitored. Social networking can be used by these relatives for sending notifications, introducing parameters’ values, timely intervention, etc.

6. Conclusion

The lack of continuous monitoring of the elderly and dependent patients especially those residing in remote areas far from hospitals, the delay in diagnosis and the rapid spread of chronic diseases in the society become a paramount concern of health informatics. Timely diagnosis and treatment can solve these problems. With the advances in machine learning tools, communication technologies and wearable sensor technology, new horizons have been opened regarding these systems.

In this paper, a machine learning based RHMS has been designed and developed using the multi-agent paradigm to monitor elderly and dependent patients. The proposed system builds a solution to provide health care services, diagnose and counsel treatments from afar, monitor people in their own indoor and outdoor environments during everyday activities. It can potentially reduce physician burden, unnecessary hospitalization, improve treatment and reduce monitoring costs. It acts as a helping hand for caregivers and paramedics to improve the quality of patient care by accelerating medical decision-making and providing a reliable diagnosis.

This study shows that the multi-agent modeling of a RHMS, based on data mining as an automatic machine learning approach and visualization as interactive machine learning approach, can enhance the behavior of these systems and can produce predictive results for monitoring elderly and dependent people. We have proposed architecture composed of three layers: (1) sensing monitoring layer for collecting and visualizing real-time date, (2) data analysis layer based on the decision tree technique for analyzing and learning from data and patterns, and (2) decision-making layer for taken and executing decision. The tasks offered by the agents along the decision-making process like coordination, communication, sharing of resources and merging it in real-time, improves performance of our RHMS.

Footnotes

Acknowledgments

The authors would like to acknowledge the financial support of this research by grants from the ARUB program under the jurisdiction of the General Direction of Scientific Research (DGRST) (Tunisia).

Authors’ Bios

	Jihed Elouni is a PhD in Computer Science Engineering. He is member of the REGIM laboratory, ENIS, University of Sfax, Tunisia. He is an Assistant in computer Sciences. He focuses his research on Decision Support Systems, Knowledge Discovery from Data, Multi-agent System and Visualization. (e-mail: auni.jihed@gmail.com).
	Hamdi Ellouzi is a PhD in Computer Science Engineering. He is member of the REGIM laboratory, ENIS, University of Sfax, Tunisia. He is an Assistant in computer Sciences. He focuses his research on Decision Support Systems, Knowledge Discovery from Data, Multi-agent System and Visualization. (e-mail: hamdi.ellouzi@gmail.com). He is an author of several papers in international journal and conferences. He is member of organization committees and reviewer in 4th IEEE Middle East Conference on Biomedical Engineering (MECBME 2018).
	Hela Ltifi obtained her “University accreditation” (habilitation research) in Computer System Engineering from the Engineering School of Sfax (ENIS), Tunisia, in 2016. She has a PhD in Computer System Engineering from ENIS in 2011. She is member of the REGIM laboratory, ENIS, University of Sfax, Tunisia. She is an Assistant Professor in computer Sciences in the Faculty of Sciences and Techniques of Sidi Bouzid, Tunisia. She focuses her research on Decision Support Systems, Knowledge Discovery from Data, Human Computer Interaction and Visualization. She is an author of several papers in international journals and conferences. She is also a reviewer of the Decision Support Systems Journal and several conferences and the Decision Support Systems Journal. Hela Ltifi has been member of organization committees of several conferences such as ICALT-SIMTA’2013 and SoCPaR 2014. She is an IEEE member and IEEE-EMBS society member. (e-mail: hela.ltifi@ieee.org).
	Mounir Ben Ayed obtained his “University accreditation” (habilitation research) in computer system engineering from the engineering school of Sfax, Tunisia, in 2013. He has a PhD in biomedical engineering from Paris XII University, France. He also obtained a Post graduate Diploma in Biomedical Engineering from University of Technology of Compiègne, France. He is Associate Professor in the computer science department of the Faculty of Science, University of Sfax. He teaches mainly DBMS, data warehouse, data mining, software engineering and Human Computer Interaction. His research activities concern Decision Support System based on a Knowledge Discovery from Data process. Most of his research are designed and evaluated in the medical field. He is author of many papers in international journals and conferences. He also was chair, organization committee member, and technical committee member of many national and international conferences. Among the numerous conferences he organized the “Engineering Sciences for Biology and Medicine”, ESBM’2013 and ESBM’2015.Mounir Ben Ayed is IEEE member since 2001
and IEEE Senior member since 2013. He created the IEEE-EMBS Tunisia Chapter in 2009. He was Chair of this chapter from 2009 to 2012, Vice Chair from 2013 to 2014 and Chair for the second time for the period: 2015–2016. (e-mail: mounir.benayed@ieee.org).

References

Fox

Beveridge

and Glasspool

, Understanding intelligent agents: analysis and synthesis, AI Communications 16(3) (2003), 139–152.

Khanna

and Misra

, White Paper Life Sciences, The Internet of Things for Medical Devices-Prospects, Challenges and the Way Forward, Tata Consultancy Services, July 1, 2014.

Annicchiarico

Cortés

and Urdiales

, eds), Agent Technology and e-Health, Whitestein Series in Software Agent Technologies and Autonomic Computing, BirkhäuserVerlag, Basel, Switzerland, 2008.

Nealon

J.L.

and Moreno

, eds, Applications of Software Agent Technology in the Health Care Domain, Whitestein Series in Software Agent Technologies and Autonomic Computing, BirkhäuserVerlag, Basel, Switzerland, 2003.

Aniello

Massimo

and Giuseppe

D.P.

, Design and validation of a light-weight reasoning system to support remote health monitoring application, Engineering Applications of Artificial Intelligence 41 (2015), pp. 232–248, ISSN 0952-1976.

Cheng

Zhu

PossJohn

J.W.

Hirdes

Glenny

and Stolee

, Opinion versus practice regarding the use of rehabilitation services in home care: an investigation using machine learning algorithms, BMC Medical Informatics and Decision Making, Department of Statistics and Actuarial ScienceUniversity of Waterloo, Waterloo, Canada, 2015.

Revuelta-Zamorano

Sánchez

Luis Rojo-Álvarez

Álvarez-Rodríguez

Ramos-López

and Soguero-Ruiz

, Prediction of Healthcare Associated Infections in an Intensive Care Unit Using Machine Learning and Big Data Tools, Springer International Publishing Switzerland 2016 Kyriacou

et al., XIV Mediterranean Conference on Medical and Biological Engineering and Computing 2016, IFMBE Proceedings 57.

Sathiya Rani

Vimala

and Kalaivani

, A Remote Healthcare Monitoring System for Faster Identification of Cardiac Abnormalities from Compressed ECG Using Advanced Data Mining Approach, In: Kumar

S.M.

eds, Proceedings of the Fourth International Conference on Signal and Image Processing 2012 (ICSIP 2012). Lecture Notes in Electrical Engineering, vol 222. Springer, India.

Holzinger

, Human-Computer Interaction and Knowledge Discovery (HCI-KDD): What Is the Benefit of Bringing Those Two Fields to Work Together? International Conference on Availability, Reliability, and Security 2013, pp. 319–328.

10.

Atzori

Iera

and Morabito

, The Internet of things: A survey, Computer Networks 54(15) (2010), 2787–2805.

11.

Piette

and Caval

, A Multi-Agent System for Resource Privacy: Deployment of Ambient Applications in Smart Environments, Proceedings of the 15th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2016).

12.

Muhammad

S.R.

Aslam

and MartnezEnriquez

A.M.

, Remote Patient Monitoring and Healthcare Management using Multi-Agent based Architecture, Ninth Mexican International Conference on Artificial Intelligence, 978-0-7695-4284-3/10 2010 IEEE.

13.

Chan

Ray

and Parameswaran

, A Multi-Agent Collaborative Framework for Mobile E-HealthSystem Sciences, 2007. HICSS 2007, 40th Annual Hawaii Int. Conf.

14.

Lan

Alshurafa

Ghasemzadeh

and Sarrafzadeh

, WANDA: An End-to-End Remote Health Monitoring and Analytics System for Heart Failure Patients, Wireless Health Conf., 2012.

15.

Lee

S.I.

Ghasemzadeh

Mortazavi

B.J.

Lan

Alshurafa

Ong

and Sarrafzadeh

, Remote Health Monitoring Systems: What Impact Can Data Analytics Have on Cost? in Wireless Health, 2013.

16.

Teng

X.F.

Zhang

Y.T.

Poon

C.Y.

and Bonato

, Wearable Medical Systems for P-Health, IEEE Rev Biomed Eng 1 (2008), 62–74.

17.

Agoulmine

Ray

and Wu

T.H.

, Communications in Ubiquitous Healthcare, IEEECommun Mag 50 (2012), 16–18.

18.

Mladenic

Lavrac

Bohanec

and Moyle

, Data mining and Decision Support: integration and collaboration, Springer-Verlag New York Inc., (2012), p. 304.

19.

Iamsumang

Mosleh

and Modarres

, Hybrid DBN monitoring and anomaly detection algorithms for on-line SHM, 2015 Annual Reliability and Maintainability Symposium (RAMS), Palm Harbor, FL, IEEE, 2015, pp. 1–7.

20.

Ltifi

Kolski

and Ben Ayed

, Combination of cognitive and HCI modeling for the design of KDD-based DSS used in dynamicsituations, Decision SupportSystems 78 (2015), Elsevier, 51–64.

21.

Randev

, Data Mining for Structural Health Monitoring, IGI Global, distributing in print or electronic forms without written permission of IGI Global is prohibited, United Technologies Research Center, University of Minnesota USA, 2009, pp. 450–457.

22.

Yoo

Alafaireet

Marinov

Pena-Hernandez

Gopidi

Chang

J.F.

and Hua

, Data mining in healthcare and biomedicine: A survey of the literature, J Med Syst 36 (2012), 2431–2448.

23.

Sun

F.T.

Kuo

Cheng

H.T.

Buthpitiya

Collins

and Griss

, Activity-Aware Mental Stress Detection Using Physiological Sensors, In Mobile Computing, Applications, and Services; Gris

and Yang

, Eds, Springer: Berlin, Germany, Volume 76, 2012, pp. 211–230.

24.

Nangalia

Prytherch

and Smith

, Health technology assessment review: Remote monitoring of vital signs – current status and future challenges, Crit Care 14 (2010), 1–8.

25.

Lakshmanan

Gilleland

McGovern

and Tingley

, Machine Learning and Data Mining Approaches to Climate Science, Proceedings of the 4th International Workshop on Climate Informatics, ISBN: 978-3-319-17219-4, 2015.

26.

Mitchell

T.M.

, Machine Learning, McGraw-Hill Science/Engineering/Math, ISBN: 0070428077, 1997.

27.

Riccio

Cavuoti

Schisano

Brescia

Mercurio

Elia

Benedettini

Pezzuto

Molinari

and Maria Di Giorgio

, Machine Learning Based Data Mining for Milky Way Filamentary Structures Reconstruction, In: Bassis

Esposito

Morabito

and Pasero

, eds, Advances in Neural Networks. Smart Innovation, Systems and Technologies, vol 54. Springer, Cham, 2016.

28.

Weng

Jiang

Guo

and Wang

, Applied research on visual mining technology in medical data, 2016 4th International Conference on Cloud Computing and Intelligence Systems (CCIS).

29.

Elouni

Ltifi

and BenAyed

, Knowledge visualization model for intelligent dynamic decision-making, The 15th International Conference on Hybrid Intelligent Systems (HIS 2015) in Seoul, South Korea, November 2015, pp. 223–235.

30.

Elouni

Ltifi

Ben Ayed

a nd Masmoudi

, Visual knowledge generation from data mining patterns for decision-making, International Journal of Advanced Computer Science and Applications (IJACSA). – Volume 7 Issue 7 July 2016.

31.

Ltifi

Ben Mohamed

and Ben Ayed

, Interactive visual knowledge discovery from data-based temporal decision support system, Information Visualization 15(1) (2016), 31–50.

32.

Foster

McGregor

and El-Masri

, A survey of agent-based intelligent decision support systems to support clinical management and research, in: Armano

Merelli

Denzinger

Martin

Miles

Tianfield

and Unland

Eds, Proceedings of MAS-BIOMED’05, Utretch, Netherlands, 2005.

33.

Nealon

J.L.

and Moreno

, Eds, Applications of Software Agent Technology in the Health Care Domain, Whitestein Series in Software Agent Technologies and Autonomic Computing, BirkhäuserVerlag, Basel, Switzerland, 2003.

34.

Annicchiarico

Cortés

and Urdiales

, eds, Agent Technology and e-Health, Whitestein Series in Software Agent Technologies and Autonomic Computing, BirkhäuserVerlag, Basel, Switzerland, 2008.

35.

Benjemmaa

Ltifi

and Ben Ayed

, Multi-Agent architecture for Visual Intelligent Remote Healthcare Monitoring System, The 15th International Conference on Hybrid Intelligent Systems (HIS 2015) in Seoul, South Korea, November 2015, pp. 223–235.

36.

Musiat

and Hoffmann

, Personalised computerised feedback in E-mental health, Journal of Mental Health 21(4) (2012), 346–354.

37.

Filipe

Fdez-Riverola

Costa

and Pereira

, Wireless Body Area Networks for Healthcare Applications: Protocol Stack Review, International Journal of Distributed Sensor Networks 2015 (2015), 1–23.

38.

Khan

Hussain

M.A.

and Kwak

K.S.

, Medical applications of wireless body area networks, International Journal of Digital Content Technology and its Applications 3(3) (2009).

39.

Moosavi

Gia

Nigussie

Rahmani

Virtanena

Tenhunena

and Isoahoa

, End-to-end security scheme for mobility enabled healthcare Internet of Things, Future Generation Computer Systems 64 (2016), 108–124.

40.

Ellouzi

Ltifi

and Ben Ayed

, New Multi-agent Architecture of Visual Intelligent Decision Support Systems Application in the medical field, 13th ACS/IEEE International Conference on Computer Systems and Applications, AICCSA, Marrakch, Marroc, Novembre 2015.

41.

Ellouzi

Ltifi

and Ben Ayed

, Modeling of distributed visual Knowledge discovery from data process, International Conference on intelligent systems and knowledge engeneering, ISKE, Nanjing, China, Novembre 2017.

42.

Ltifi

Ben Ayed

Kolski

and Alimi

, HCI-enriched approach for DSS development: the UP/U approach, IEEE Symposium on Computers and Communications (ISCC’09, July 5–8), IEEE Press, Sousse, Tunisia, 2009, pp. 895–900.

43.

Ellouzi

Ltifi

and Ben Ayed

, Multi-agents modelling of Decision Support Systems based on visual data mining, Multi Agent and Grid Systems, Multi Agent and Grid Systems – An International Journal 13(1) (2017), 31–45.

44.

Leite

and Girardi

, A hybrid and learning agent architecture for network intrusion detection, Journal of Systems and Software 2017, pp. 59–80.

45.

Soumyajit

Shreeasish

and Monideepa

, Dynamic Service Deployment Using JSON Based Decision Trees in Remote Healthcare, Intelligent Systems and Control (ISCO), 2016 10

{}^{\rm th}

International Conference, School of Computer Engineering, KIIT University, Bhubaneswar, Odisha, India.

46.

El Haddioui

El Mamoune

and Loubna

, New system forecast e-health by using hybrid of association rules and decision tree, 2014 5th Workshop on Codes, Cryptography and Communication Systems (WCCCS) university, Faculty of Sciences, Tetuan, Morocco.

47.

Manogaran

Varatharajan

Lopez

Kumar

Sundarasekar

and Thota

, A new architecture of Internet of Things and big data ecosystem for secured smart healthcare monitoring and alerting, Future Generation Computer Systems 82 (2018), 375–387.

48.

Dotoli

Epicoco

Falagario

Seatzu

and Turchiano

, A decision support system for optimizing operations at intermodal railroad terminals, IEEE Transactions on Systems, Man, and Cybernetics: Systems 47(3), 2017.

Intelligent health monitoring system modeling based on machine learning and agent technology

Abstract

Keywords

1. Introduction

2. Literature review

2.1 Remote healthcare monitoring system

3.1 System architecture

Table 1 Agents of the sensing monitoring layer (data collection layer)

3.2.1 Sensing monitoring layer

Table 2 Common agents

Table 3 Agents of the data analysis layer (data collection layer)

Table 4 Agents of the decision making layer

4.1 Sensing monitoring layer

Table 5 Various physiological wearable body sensors

Table 6 A comparative table of current and old systems execution time

Footnotes

Acknowledgments

Authors’ Bios

References

Table 1
Agents of the sensing monitoring layer (data collection layer)

Table 2
Common agents

Table 3
Agents of the data analysis layer (data collection layer)

Table 4
Agents of the decision making layer

Table 5
Various physiological wearable body sensors

Table 6
A comparative table of current and old systems execution time