key: cord-0932058-jd0zc5ae authors: Lanza, Francesco; Seidita, Valeria; Chella, Antonio title: Agents and Robots for Collaborating and Supporting Physicians in Healthcare Scenarios date: 2020-06-27 journal: J Biomed Inform DOI: 10.1016/j.jbi.2020.103483 sha: 4326179cc11c9578381ba6977e33a20344aaed90 doc_id: 932058 cord_uid: jd0zc5ae Monitoring patients through robotics telehealth systems is an interesting scenario where patients’ conditions, and their environment, are dynamic and unknown variables. We propose to improve telehealth systems’ features to include the ability to serve patients with their needs, operating as human caregivers. The objective is to support the independent living of patients at home without losing the opportunity to monitor their health status. Application scenarios are several, and they spread from simple clinical assisting scenarios to an emergency one. For instance, in the case of a nursing home, the system would support in continuously monitoring the elderly patients. In contrast, in the case of an epidemic diffusion, such as COVID-19 pandemic, the system may help in all the early triage phases, significantly reducing the risk of contagion. However, the system has to let medical assistants perform actions remotely such as changing therapies or interacting with patients that need support. The paper proposes and describes a multi-agent architecture for intelligent medical care. We propose to use the beliefs-desires-intentions agent architecture, part of it is devised to be deployed in a robot. The result is an intelligent system that may allow robots the ability to select the most useful plan for unhandled situations and to communicate the choice to the physician for his validation and permission. The objective is to support the independent living of patients at home without losing the opportunity to monitor their health status. Application scenarios are several, and they spread from simple clinical assisting scenarios to an emergency one. For instance, in the case of a nursing home, the system would support in continuously monitoring the elderly patients. In contrast, in the case of an epidemic diffusion, such as COVID-19 pandemic, the system may help in all the early triage phases, significantly reducing the risk of contagion. However, the system has to let medical assistants perform actions remotely such as changing therapies or interacting with patients that need support. The paper proposes and describes a multi-agent architecture for intelligent medical care. We propose to use the beliefs-desires-intentions agent architecture, part of it is devised to be deployed in a robot. The result is an intelligent system that may allow robots the ability to select the most useful plan for unhandled situations and to Today there are many clinical scenarios in which a physician does not rely solely on himself, his knowledge or experience or his presence, to solve a patient's problems. Current scenarios are made more complicated by the increase in the average 5 life expectancy of citizens, especially in Western countries, which leads to an ever-increasing demand for healthcare systems. It is also remarkable today that societies are committed to ensuring access to care and well-being to all citizens [1] . There are also cases in which patients live in poor or not easily accessible places or in general cases in which the physician must provide, or 10 receive if he is not on-site, fast and reliable diagnoses to be able to establish a therapy or otherwise solve a problem. This last is the case of the emergency due to the COVID-19 pandemic. The problem often faced by emergency room physicians was not to have the means for early identification of infected cases. This fact caused a lot of people infected and then dead also among doctors and 15 nurses. Another case is the lack of professionals or, increasingly challenging, the presence of changing contexts. For instance, cases in which patients with the same disease but placed in a different family or social contexts have different characteristics and needs. Probably in these cases, a unique protocol cannot be applied, but doctors have to be able to decide on a case by case basis. 20 As evidenced by documents issued by the European Commission, the urgent need for intelligent systems for healthcare has not to be undervalued. Investigating the importance of AI & Robotics in healthcare is the current challenge for scientists and physicians, as illustrated by the Policy Department for Economic, Scientific and Quality of Life Policies in [2] . Here the main question 25 is "Robots in Healthcare: a solution or a problem?". Following these research lines, we are investigating "how can an intelligent system help a physician in making decisions, even in dynamic contexts?" Supporting physicians and patients in "complex" clinical contexts requires intelligent systems endowed with the ability to solve problems during interac- 30 tions. During the execution, the system interacts with users and the environment that often change continuously as a result of the interaction. These challenges may be solved by multi-agent systems able to self-adapt to changing situations and deliberate also in the total or partial absence of input data from physicians or patients. 35 These aspects cannot be faced and solved at the design time. Developers cannot identify and implement all the possible situations where a high level of autonomy is required. At best, they can identify several conditional statements and allow for a set of possible alternatives in the system behavior. The more dynamism or uncertainty in clinical contexts exist, the more physi- 40 cians may need to be supported by an intelligent system. In this situation, developers have to implement mechanisms that let the system autonomously monitor patients, retrieve important information, reason and suggest actions to physicians if necessary. All this may be done employing robots. Thus, an efficient way to face the problem mentioned above is using the 45 agent-oriented paradigm [3] and robots. The use of agent-oriented paradigm in complex systems, such as robotic platforms or internet of things applications, was studied and treated by scientific community [4, 5, 6] during these years. For instance, the application of the agent-oriented computing, for the IoT domain and the cyber-physical systems, is increasing day-by-day. Weaknesses and 50 issues in these domains can be solved exploiting agent-oriented architectures and relative computing techniques as well discussed and highlighted in [7] . Another way to employ multi-agent systems for developing monitoring systems in the healthcare scenario may be seen in [8] . However, multi-agency has not been used for providing support in dynamic environments for healthcare do- 55 main. The contribution and the novelty of this paper lay in the creation of an agent-based architecture for healthcare systems. The architecture handles monitoring, knowledge management and deliberation module. The main novel idea we present is to employ the Belief-Desires-Intentions (BDI) paradigm and its reasoning cycle for implementing a multi-agent system able to deliberate 60 and plan teleoperation activities to help physicians in making decisions. Even when information about plans to execute and goals to pursue are lacking or incomplete. The multi-agent system resulting from the proposed architecture has been conceived for being deployed in a robot to support physicians or patients in their activities. The rest of the paper is structured as follows: in section 2 some related work are explored; section 3 briefly shows multi-agent systems topics, their features and how they handle the selection of plans during the execution phase; in section 4 we discuss the architecture we propose for assisting physicians and how plans are created is illustrated with an example; in section 5 we validate the 70 proposed architecture and finally in section 6 some discussions and conclusions are drawn. The need to build a robotic system able to satisfy patients brought scientists to design robotic platforms to face this problem, as in the case of Koceska et 75 al. [9]. Koceska et al. studied available platforms in the market of assistive robot systems. They designed and developed a low-cost assistive telepresence robot system for facilitating and improving the quality of life of elderly and people with disabilities. To improve the quality of life the robot system is able to interact with the environment (such as, moving small objects and measuring 80 vital parameters) and can be controlled by a remote assistant. The robot merely executes the commands of the remote assistant. An advantage is that this is a low-cost robot that can be easily used in everyday life by patients and professionals. The need for a telepresence robotic assistant has driven the interest of the EU 85 itself, which has financed FP7 and H2020 European Projects for this purpose. MARIO project 1 faces an important challenge into the field of telepresence robotic assistant. It aims to challenge loneliness, isolation and dementia in elderly people [10, 11, 12, 13] . The project connects elderly people with their needs. It aims to be independent of robotic platforms. Moreover, it supports 90 remote application installation and deployments onto the robotic platforms. The project uses data exchanged during personal or social interaction. Even if the project faced these challenges, no dynamic supports to patients or physicians were considered. It does not facilitate the interaction including new activities at runtime without the contribution of developers. 95 ENRICHME project 2 is another EU project financed through the H2020 Framework Program [14, 15] . The project aims to face the cognitive decline of cognitive capacities in elderly people. The solution consists of an integrated platform for ambient assisted living, integrated with a mobile service robot. The project aims to realize long-term human monitoring and interaction system for 100 letting elderly people stay independent in their home way. The strength of the project is to enable physicians and caregiver to analyze data for identifying changing in cognitive impairments and so early acting on them. This system does not take in consideration robots as a human supporter. MOVECARE project 3 aims to realize "an innovative multi-actor platform 105 that supports the independent living of the elder at home by monitoring, assisting and promoting activities to counteract physical and cognitive decline and social exclusion" [16, 17, 18] . The MOVECARE architecture monitors the frailty of the elderly specifically based on criteria identified and discussed by Fried et al. [19]. Like the previous approach, this framework provides a good means for 110 monitoring; it adds some kind of intelligence during the reasoning process. The reasoning system can alert caregivers about the status of the patients and pos-1 https://cordis.europa.eu/project/id/643808 2 https://cordis.europa.eu/project/id/643691 3 https://cordis.europa.eu/project/id/732158 sible motivations exploiting the Fried frailty criteria. Nonetheless, the robotic platform is not designed for acting as a true caregiver. Caresses project 4 , that stands for "Culture Aware Robots and Environmen-115 tal Sensor Systems for Elderly Support", is an H2020 Project financed by EU for building culturally competent care robots [20, 21, 22, 23] . The project aims to build the first robot that assists the elderly, adapting itself with the culture of the user. At the best of our knowledge, the platform is not able to handle runtime planning for executing therapies. All previous projects consider several important aspects of telemedicine, assistive robotics and human-robot interaction for healthcare, but activities are not thought to support physicians by remote in a proactive manner. Another approach in the literature, to perform robotic teleoperations for ultrasonic medical imaging, is the OTELO System. OTELO performs tele-echography 125 [24] remotely. It supports the clinicians in making a diagnosis. This system even treats patients in their home, does not operate proactively and it needs for physicians to be used. Progress in networking lets computer scientists build applications using a network to exchange data and information. Systems can exchange informa-130 tion using the network and the spread of wireless. In [25] authors propose a mobile-care system integrated with a variety of vital-signs monitoring, where all involved devices are endowed with a wireless communication module for data exchanging. The mobile-care system uploads data into a care server via the internet. Data stored into the remote server are available for physicians 135 that need to check the health status of patients. The system serves as an alert system where interaction with patients is implemented. Evolution of telemedicine systems brought at the definition of novel systems able to handle more complex scenarios. Recent works introduce a new technological paradigm for elderly people that live alone. In [26] authors consider 140 IoT (Internet of Things) technology to connect devices in patient's home for data collecting and communication. Other studies involve the usage of robots, or better social robots, as telepresence systems [27] . In this work, the authors proposed a telepresence robot, built for a human-robot interactive experiment. Each of these works aims to realize complex systems with the ability to 145 supervise people at home letting them live without invasive instruments through the usage of robotics or sensors scattered into their homes. Other approaches are present in literature, most of them resolve problems related to telepresence, monitoring and teleoperation. We are now facing the following problem: each patient's environment is not 150 equal to another leading to difficulty in handling new occurring situations by clinicians. Moreover, every patient answers therapies in different ways and so, every care system has to be endowed with the ability to adapt its behaviors taking into account this weakness. For that, scientists and engineers need for methods to facilitate reasoning operation at runtime. Additionally, our approach employs a multi-agent architecture that gives intelligent support to the physician. The architecture, deployed in a robot, monitor and assist the patient as a caregiver, also in cooperation and collaboration with the doctor. 165 An agent is an autonomous entity able to act in response to stimuli coming from the environment and proactively act towards a specific goal [3, 28, 29] . The agent paradigm was born to better understand and model complexity in talking about agents let us refer to a functional program, a Java class, a compiler or something like that. All these kinds of software may be modeled by a function, they receive input and produce an output as the result of elaborating that input. Everything happens in this application is because we (the program-180 mers or the designers) want it to happen and to happen exactly in that way. An autonomous agent is conceived to be at the exact opposite of the applications above. Research in the field of agents and multi-agent systems is going towards means for building agents to which we can delegate tasks. It is up to the agent to decide how to reach the objectives, they act on the base of plans we give them. A plan defines the set of actions an agent may perform to pursue an objective. An agent is reactive in the sense that it is able to perceive the environment and act in response to changes coming from it, actions are directed towards the agent's objective. One of the key points related to agent autonomy is that an agent may put together plans to achieve our goals. So they are making able to 190 operate autonomously on the behalf of humans that delegate them their goals. Proactiveness is the ability "to exhibit a goal-directed behavior by taking the initiative". Social ability is the ability of an agent to interact with other agents and humans to purposefully reach its objective. This latest ability involves an important skill belonging to humans, i.e. communication abilities. To establish 195 a society, the agent cooperates and coordinates activities with other agents and therefore it is able to communicate to the other (also humans) its beliefs, its goals and its plans. Healthcare systems or other kinds of systems supporting physicians in the scenarios identified in the introduction are perfectly manageable with agents. This kind of support cannot be reached by using a simple reactive system. In the following section, we explain the agent architecture, involving BDI agents, 250 we propose for physicians support in a clinical scenario. The computational model implementing deliberation and means-ends reasoning contains four elements: B (beliefs), I (intentions), D (desire) and π (plan). At the beginning agent is endowed with a belief base, hence its knowledge about the environment, and a plan library, a set of possible plans for reaching objec- mentations of BDI agents. One of the most known and efficient ones is the Jason framework and its reasoning cycle [28, 33] . Jason is a powerful instrument for realizing planning in uncertain environments. In the next section, we detail the proposed agent architecture exploiting the said BDI features and we explain how we implement the architecture in a robot 280 supporting and cooperating the physician for reaching a common goal. This section starts with the description of a pilot scenario, useful for deeply understanding the domain context. To accomplish the previous scenario we need to develop a robot that might 320 serve as an intelligent object for monitoring and checking patients in their environment and also as a collector of data for enhancing diagnosis. The robot has also to be able to change its behaviors at runtime according to the physician's prescriptions. The idea is to develop such a kind of robot as a part of a cognitive system 325 realized employing BDI agents and on the base of the following architecture ( Figure 1 ). The architecture is composed of four modules: • Environment Management -it contains all useful elements for interfacing physicians, for monitoring the environment in which patients live 330 and for acquiring data; • Knowledge Management -it is the module for storing and managing data from the environment and patient; • Reasoning -this is the module devoted to compute data stored into the knowledge module and to produce a series of actions to accomplish a task; • Acting -given a set of actions, it extracts the proper action for a specific situation. The Knowledge Management module depends on the Environment Management module, indeed only data resulting from monitoring form the knowledge of the system. Data stored in the Knowledge Management module are used by the system for the reasoning process and, at the same time, all the results of the reasoning process update knowledge. The Reasoning module produces a set of actions, useful to reach one or more system goals. This set is the input for the Acting module where, it extracts the proper action sent to the system. The result of an action normally produces a change in the state of the environment. So, this module affects the monitoring one. The architecture in Figure 1 is implemented employing a BDI multi-agent system where a set of agents works to reach the system's goal. The efficiency of the multi-agent paradigm in modeling and handling complex systems has been widely discussed in the literature [3, 36] . Handling robots via a multi-agent system means creating a multi-agent architecture able to handle components and planning abilities. As said, our idea is to use the BDI agents paradigm for providing a intelligent support to the physician work, also allowing him to manage robot remotely. Beliefs, rules, desires and intentions for each agent are defined at design time by developers and users (physicians and patients in the clinical domain). Beliefs are generally of two types: (i) perceptions acquired through sensors 360 from the surrounding environment; (ii) information acquired through messages exchanged with the other agents into the platform. Beliefs are critical for the agent's planner. Beliefs and rules push a plan to be selected as the best set of activities to do. Selecting one plan, compared to another, means changing the behavior of the agent during its operating cycle. The agent's intelligence is defined with a descriptive file written in AgentSpeak(L) [30, 37] . AgentSpeak(L) is a language that uses logical formalism to define a set of plans for handling situations to reach the agent's goal. The agent programmer writes initial beliefs, rules, goals and plans. Each plan is composed of a head and a tail. The head is composed of a event 370 trigger and a circumstance. The former is the trigger condition that launches actions defined into the plan and the latter is a set of parameters used for validating the context in which some actions are allowed. More in detail, a plan is launched when the event trigger is scheduled and the set of beliefs contained in the circumstance are solved by a unification process where a literal is unified 375 with the data contained into the namesake belief and verified by a first-order logic process. The tail of a plan is composed of sub-plans or actions that can be internal actions or external actions. In the next paragraph, we show the architecture detailing how agents work, 380 communicate and cooperate to support physicians. Figure 2 represents the multi-agent system that works in a typical healthcare scenario. In this scenario, two environments are involved, the patient's residence and the physician's office. The two environments are connected through a multi-agent platform, each environment owns one or more agents, or better one or more agents are deployed in each environment. The multi-agent platform serves as a bridge among the patient and the physician. It is worth to note that the middle layer of this representation constitutes the intelligent part to be added to a simple teleop- Collected data are organized and stored to be saved in the remote server. The Knowledge Management module organizes data to be synced with the system and updates them with the last perceived one for the diagnostic purpose by physicians. The same data are handled from the agents into the deliberation process module. This module is composed of two sub-systems, the first for reasoning and the second for acting. Multi-agent systems are distributed systems over the network. Each agent could be relocated to other computers that host a node or a set of nodes of the system infrastructure. The multi-agent system can be distributed over remote sites connected via the internet. The system we propose uses the internet to share information, 410 data and alerts between the physician's office and the patient's environment. The physician's terminal is connected to the multi-agent system through the Virtual Assistant agent. It implements all necessary functionalities to be into agents' network and it receives data from the patient's environment via the internet. Data are stored using an OWL ontology, they are accessible from the physician by a terminal for diagnostic purpose. Data are also stored in remote servers for safety reasons. A memory manager holder organizes acquired perceptions and collected information by the system in an OWL ontology [38] . The knowledge is split into a long-term and working memory; in this way, we avoid prob-420 lems such as memory leak or memory explosion from agents when they have to use it. So the memory is handled as described in [39] and the agent which was delegated to manage knowledge is deployed into the system with the goal to keep all data synced and updated. The knowledge of the system is organized to be always synced, updated and shared on the entire system and to avoid As said in section 2, nowadays, there exist several teleoperated robots that let the physician operate into the patient's environment, no one of them works in autonomy during the interaction with the patient. Autonomy, in this case, means to be able to recognize or retrieve the best action to perform in a particular situation, propose the action to the physician and wait for his command. In this sense, the robot is autonomous and it can suggest some actions to the physician. Indeed, in our system physicians can teleoperate in the sense of changing previous therapies or adding new ones to enhance the quality of life of the patient. The robotic caregiver should not be seen as a professional robotic avatar but as a valid instrument useful for monitoring and assisting patients in 445 the place of clinicians. Virtual Assistant agent endows physician for checking, revising, updating a therapy or a set of therapies initially prescribed and provided to the robot for taking care of the health status of the patient. Therapies are plans in the sense of the agent's plan that the Planner agent selects when the context is verified. The Planner agent is responsible to realize the intelligence of the system. It communicates with the Knowledge Manager agent to obtain data synced, updated and stored into the ontology for selecting the best plan that fits the current situation. To satisfy the system goal, the Planner agent tries to find the right solution using the context of a logical formula; once the context is verified In Table 2 is summarized the description of the role of the agents that are 465 into the system. The robot situated in the patient's environment works as a robotic caregiver. The multi-agent architecture contains modules for handling knowledge, planning, motion, vision and sense for robots. Each agent implements its logical 470 model to accomplish established goals that represent the desires of the agent. It is a part of the multi-agent system used for managing the system's knowledge. It uses a double kind of knowledge model, an ontology for storing information and gathering them as concepts and relations and a knowledge base where beliefs are stored for acting in the working domain. It is a part of the multi-agent system used for planning operations. This agent is delegated for selecting which plan is more adapt to pursuing the goal. Generally, the system goal is defined at design time but it can be handled at runtime. This agent listens for information gathered and it deliberates the set of actions and plans that each agent has to execute using the agent's communication module. This agent communicates directly with the virtual assistant agent. This last is responsible for upgrading the planner's plan library to add plans and enabling it for revising plan at runtime; this means that the multi-agent systems can act dynamically. It is a part of the multi-agent system used for vision task operations. It is a part of the multi-agent system used for collecting data from sensors. These data contribute in decision-making phase, to select the plan that best fit with the current situation. It is a part of the multi-agent system used for handling robot motion. This agent implements modules for letting robot move and executing into the environment. It is a part of the multi-agent system used for letting physician be part of the system. This agent operates as a dashboard console on which the physicians retrieves data monitored by the robot and collected from the system. The agent lets physicians communicate new therapies to the robot that they will be translated into plans and added into the plan library of the planner agent. Therapy will be executed during the next agent's reasoning cycle if pre-conditions signed by the physicians will happen. If this does not happen, an alert will be sent to the dashboard console. Fig. 2 The robot uses algorithms for implementing obstacle avoidance and motion. The agent may navigate in unknown indoor places using an adapted version of a state-of-the-art algorithm for robot SLAM (Simultaneous Localization And Mapping) [40] . The Motion agent is strictly connected with the Vision agent. While the former deals with the motion of the robot, the latter deals with computer vision tasks. The Vision agent, continually, senses the environment using RGB camera to discover and to recognize objects located into the environment. The object recognition module works using the YOLO deep neural net-480 work 5 [41] . YOLO is a state-of-the-art system for implementing real-time object detection and recognition. Recognizing objects is useful for motion's tasks, such as obstacle avoiding or to enhance the localization algorithm using some objects as landmarks or fixed-points. The Vision agent is also able to detect the status of the patient, such as un-485 derstanding when the patient is sitting or lying down or the patient is standing. Vision agent takes also care of other tasks; a deep description of this module is out of the scope of the paper. Other agents are deployed into the multi-agent platform to sense the environment and to enhance the quality of perceptions. An agent, delegated for Our multi-agent system is implemented using the Jason framework [33, 29] and the agent's reasoning cycle is well described in [29, 42, 43] . Briefly, each Jason agent reasoning cycle ( Figure 3 ) involves four phases: (i ) sensing phase (in yellow); (ii ) beliefs revision and updating phase (in pink); (iii ) deliberative phase (in blue); (iv ) acting phase (in green). Each cycle starts perceiving the environment and ends with acting operations in the surrounding context. Once the agent has perceived something from the environment, internal knowledge handling is executed. The core of the reasoning cycle is on how the agent selects the plan that best fits with the context. The reasoning cycle is responsible for the rational process for robot acting. As said before, plan is composed of three parts: triggering event : context ← sub-plans; actions; messages. and a unification process over beliefs and a first-order logic is applied from the 515 reasoner. The first-order logic evaluates logical conditions over boolean expressions. The formalism used in Jason is inherited from the reasoning system developed using AgentSpeak(L) [37] . A list of plans is written at design-time by agent programmer according to the goal that the agent has to pursue. Each plan 520 is evaluated in the order in which it is written into the agent description file, written in AgentSpeak(L) language. The order in which plans figure into the Plan Library is fundamental for the decision-making process. In fact, the Plan Library is handled as a queue, where the first plan deployed into the Plan Library is the first that is select to 525 be checked and executed if it is valid. The Plan Library contains all plans written at design-time and can be handled during agent life-cycle adding, revising or removing pre-defined plans. In this sense, the first plan in the queue that best fit with the condition is executed. Executing plans results in actions over the environment or exchanging infor- runtime. Ideally, this agent is divided into two parts: (i ) plans for robot acting, (ii ) plans for administering therapies. The former is configured for permitting to operate and this section of plans are written at design time by agent programmer, the latter is the section of plans that the system activates to administer therapies to the patient to serve the physician and taking care of the patient. In the next section, we describe how to revise therapies remotely starting from diagnosis and we introduce the mapping process between therapies and agents' plans. Diagnosis represents the status of the patient in a particular moment of his 545 life and the triggering condition for administering a therapy. Therapies or medical treatments are prescriptions aiming to heal a person with health problems. Each therapy contains a list of indications and contraindications for the patient and it can be effective or not. There are several types of therapies, in our approach, we take into consideration only the following type: Dr. Haus compiles the new therapy following some basic rules in order to be compliant with the system. Rules are necessary for translating therapies in the "AgentSpeak Language". This task is also in charge of theVirtual Assistant agent. The translation process follows some basic rules, resumed in Table 3 : • a therapy owns a triggering event, conditions for detecting it and a list of procedures to act when conditions appear and they are verified; • the triggering event has to catch the status that the system has to attention for handling the situation. It is defined with a label that represents the 590 triggering event name; • conditions have to be identified with in mind a list of symptoms or observations that the architecture can retrieve from the environment. Symptoms are perceptions or beliefs and they are used in the unification process and logical inference for validating a context; • the procedures list has to contain actions, behaviors or other therapies (plans) that try to resolve the issue acting or interacting with the agent; • a procedure can be a simple action or a more complex structure that involve the usage of other therapies; • an alternative therapy can be added into the plan library keeping the same .search_medicine(P); .prepare_medicine(P,Q); .take_drug_to(A); .check_therapy_taken_by(A); .signal_to_physician(D). The physician can select actions that have been previously programmed for its quality requirements and identifying risks [45, 46, 47] . So, what makes an architecture good is in how much it fits the goals and needs of the organization that is going to use it. A recent survey [48] presents a comparative analysis of software architectures evaluation methods, their result is a taxonomy of evaluation methods. From this taxonomy, it arises that comparing and validating software architectures is a hard task due to the different languages and notation used for defining architectures. The method we choose provides a notation for describing the architecture, mainly highlighting its structural perspective. Also, this method starts from considering the goals underpinning the creation of the software sys-660 tems and allows them to link them to some quality attributes. In so doing, we can focus on the quality concerns satisfying some software quality factors: maintainability, portability, modularity, reusability and robustness. The activities reported in SAAM [44] for evaluating an architecture are: 1. Characterize a canonical functional partitioning for the domain. 3. Choose a set of quality attributes with which to assess the architecture. Choose a set of concrete tasks which test the desired quality attributes. Evaluate the degree to which each architecture provides support for each 670 task. The first activity implicitly led us to identify roles and responsibilities to assign to agents in the architecture in Figure 1 . Functional partitioning consists of the separation of concerns between managing the environment and managing knowledge for the decision process. In Figure 5 , we represent the result of 675 converting our architecture using SAAM notation. In so doing, we highlight the computational entities and, data and control connections among them. SAAM then prescribes to choose some quality attributes. In our case, maintainability, portability, modularity and reusability are intrinsic in the nature of an agent architecture. The agent paradigm allows designing a software system with a 680 high degree of modularity and a low level of coupling among components that guarantee these quality factors. Going into details of this topic is out of the scope of this paper. We selected robustness for validating the architecture and we also based on FURPS [49, 50] For validating the architecture in supporting these tasks, we have to refer 695 to its representation in the SAAM notation ( Figure 5 ). The SAAM notation has been created for separating the control flow from the data flow. As can be seen, we identified three basic processes whose thread of control is assured by the control flow and the data flow among the computational components, the active and passive data. are two data and one control flow exchange. The process involved in the thread of control is of two different levels where one comprises the second. Dependencies between processes could create an architectural coupling that may slow the thread of control for supporting that task. For instance, the scheduled plan might depend on other processes' results and the overall outcome is afflicted by 735 the computational time needed. In so doing the reaction to the change may not be responsive and other sudden events may invalidate the chosen plan, affecting the system's robustness. Hence, this situation may make all the system asynchronous towards changes. We are investigating an intelligent support system based on scene understanding for forecasting events to face the problem. However, the task is supported, and adaptation to changes is guaranteed. We evaluated some other tasks that we do not report here because they are minor tasks, and the conclusion we drew was in the same line as the previous ones. From the validation process, we realized that the proposed agent architecture 745 fits all the needs related to the functional requirements in section 3. As said, architecture is not good or bad in general but in concerning some specific goals. In this case, SAAM revealed that the architecture is modular and decoupled enough for supporting in reacting to changing at runtime. In the future, we will complement this validation with metrics and analysis 750 results. We will inspect the flow of events and how they will be supported by the robotic platform. Now, this cannot be evaluated, because we tested modules independently from the operational scenario. Modules were tested onto the robot Pepper 6 , simulating hypothetical input perceived from the environment. From a state-of-the-art review, robots used in healthcare act mainly as teleoperators. We propose the use of robots and agent technology to provide the doctor with slightly more intelligent support than a simple teleoperator system. The robot does not make decisions but, programmed based on the agent architecture we propose, can interact with patients and doctors in a changing 760 environment and alert the doctor or suggest strategies when necessary. The robot is endowed with the ability to handle unforeseen situations and to communicate and collaborate with the physician, thus providing him with intelligent support. Using a robot produces several advantages, mainly having an avatar of the 765 physician that acts on his behalf. Employing robots increases healthcare facilities' efficiency when the number of physicians is low, reduces the risk of infections due to pandemic such COVID-19 [52] , because a robot is immune, speeds up the detection of new occurring situations. The architecture we propose (shown in Figure 1 ) resembles a cognitive model 770 and it is implemented using the multi-agent approach. It has been inspired by our previous works in the Human-Robot Interaction (HRI) area [53, 54] . The main contribution of this work is the creation of an intelligent bridge 6 https://www.softbankrobotics.com/emea/en/pepper between the physician and the patient. It has been realized by fully exploiting the strength of BDI multi-agent systems. The proposed architecture offers the advantage to split in the space all the system's functionalities. In this way, the overall system may be easily scalable and adaptable to any context. The agents, to be implemented based on the architecture, have been conceived at a higher level than the implementation one. Hence, thinking to their macro-level functionalities. It is for this reason that the 780 combination of architecture plus agent system can be adapted and extended to other and more complicated clinical scenarios. For instance, specialized agents could support specialized physicians through intelligent software that uses a data-driven approach for building interfaces as in [55] or including techniques for data manipulation as in [56, 57] . Moreover, everything related to monitoring can 785 be deployed in various sensors to adopt IoT techniques for patient monitoring. In the next future, we are planning to realize a physical deployment on nursing homes to validate our system. We will evaluate trials using the experience of domain experts and metrics for software analysis on collected system results. Given our experience in the development and use of robotic systems, we claim 790 that some disadvantages and limitations we might found so far are that the potential efficiency of the whole system clashes with technological problems related for example to the robot's abilities or the interaction of the robot with the human. In the future, we will fine-tune these last two aspects, especially the one related to interaction, including in this context elements of Human- Robot Teaming Interaction and the results that the Robotics Lab of Palermo has achieved so far in this field, such as the integration of the method proposed for learning plans [58] autonomously. [ Agent-Oriented Software Engineering: The State of the Art, Agent-Oriented Software Engineering: First International Workshop Multiagent systems: A survey from a machine learning 810 perspective Multi-agent robot systems as distributed autonomous systems Multiagent robotic systems Agent-based internet of things: State-of-the-art and research challenges Jade implemented mobile multi-agent based, distributed information platform for pervasive health care monitoring A telemedicine robot system for assisted and independent living So what have the roman's researchers ever done for us-mario Managing active and healthy aging with use of caring service robots Mario: Managing active and healthy aging with use of caring service robots What people with dementia want: designing mario 835 an acceptable robot companion A robotic solution for assisting people with mci at home: Preliminary tests of the enrichme system The enrichme project The MOVECARE Project: Home-based Monitoring of Frailty A multi-actor framework centered around an assistive mobile robot for elderly people living alone Evaluating the acceptability of assistive robots for early detection of mild cognitive impairment Frailty in older adults: evidence for a phenotype Caring robots are here to 865 help Views of nurses and other health and social care workers on the use of assistive humanoid and animal-like robots in health and social care: a scoping review The influence of culture on attitudes towards humanoid and animal-like robots: an integrative review An experimental study on culturally competent robot for smart home environ-875 ment Advances in robotic teleechography services-the otelo system Design and implementation of a mobile-care system over wireless sensor network for home healthcare applications Intelligence iot (internal of things) telemedicine health care space system for the elderly living alone Remote healthcare for the elderly, patients by tele-presence robot Reasoning about rational agents Systems in AgentSpeak Using Jason (Wiley Series in Agent Technology BDI agents: from theory to practice., in: ICMAS Intention, plans, and practical reason What is intention International Workshop on Computational Logic in Multi-Agent 905 Seeking prevention of cognitive decline in elders via activity suggestion by a virtual caregiver Digitalized cognitive assessment mediated by a virtual caregiver Multi-agent systems: an introduction to distributed artificial intelligence AgentSpeak (l): BDI agents speak out in a logical computable language Web ontology language: OWL, in: Handbook on ontologies Representing and developing knowledge using JASON 3-d mapping with an rgb-d camera Yolov3: An incremental improvement Human-agent interaction, the system level using JASON Decision process in human-agent interaction: Extending JASON reasoning cycle puter Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 11375 LNAI Saam: A method for analyzing the properties of software architectures The goal of software architecture analysis: Confidence building or risk assessment A survey on software architecture analysis meth-945 ods A framework for classifying and comparing software architecture evaluation methods International Conference on Computing for Sustainable Global Development (INDIACom) Capturing architectural requirements, IBM Rational developer works Evaluation of software quality Object-oriented software engineering: Using UML, Patterns and Java Combating covid-19-the role 960 of robotics in managing public health and infectious diseases Human-robot teaming: Per-965 spective on analysis and implementation issues A cognitive architecture for human-robot teaming interaction A framework 970 for data-driven adaptive gui generation based on dicom Pervasive access to mri bias artifact suppression service on a grid An automatic method for metabolic evaluation of gamma knife treatments Agents in dynamic contexts, a system for learning plans • Multi-Agent system architectures for collaborative work in medical environments • Monitoring patients through robotics telehealth system • Assisting physicians in complex and dynamic clinical scenarios • Human-Robot teaming interaction in clinical scenarios • Intelligent agents deployed in the environment • Runtime changing therapies ☒ The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.☒The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: