The aging population brings many challenges surrounding the quality of life for older people and their carers, as well as impacts on the healthcare market. Several initiatives all over the world have focused on the problem of helping the aging population with Arti cial Intelligence (AI) technology, aiming at promoting a healthier society, which constitutes a main social and economic challenge. In this paper, we focus on an Ambient Assisted Living scenario in which a Smart Home Environment is carried out to assist elders at home, performing trustworthy automated complex decisions by means of IoT sensors, smart healthcare devices, and edge nodes. The core idea is to exploit the proximity between computing and information-generation sources. Taking automated complex decisions with the help AI-based techniques directly on the Edge enables a faster, more private, and context-aware Edge Computing empowering, called Edge Intelligence.
One of the main issues that healthcare is facing is the aging population, which will lead to an ever-increasing rise in the costs associated with prevention, diagnosis, and treatments. In recent years there has been an increasing attention on Ambient Assisted Livng (AAL) topics such as \aging well" or \domiciliary hospitalization". In particular, the latter deals with the situation in which a person is considered or treated as hospitalized even when he/she is at home. In this scenario, Arti cial Intelligence (AI) techniques can play a crucial role. The development of new AI-based techniques, supporting older adults and helping Copyright c 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). them to cope with the changes of aging and healthcare assistance, represents one of the most advanced Information & Communication Technology (ICT) areas.
Thanks to the ever more availability of the Web resources, the Internet of Things (IoT) and mobile technologies, the healthcare system has now the possibility of moving a step forward the path of prevention, diagnosis and care to the patient's home by delegating the use of healthcare facilities, personnel, and machinery only in cases of urgency or specialized expertise.
Due to various challenging issues such as computational complexity and more delay in Cloud Computing, Edge Computing is an emerging paradigm and a promising solution that pushes computing tasks and services from the network core to the network edge. Recently, Edge Computing has overtaken the conventional process by e ciently and fairly allocating the resources i.e., power and battery lifetime in IoT-based industrial applications. In the meantime, considering that AI is functionally necessary for quickly analyzing huge volumes of data and extracting insights, there exists a strong demand to integrate Edge Computing and AI, which gives the birth of Edge Intelligence. Moreover, Big Data has recently gone through a radical shift of data source from the mega-scale cloud datacenters to the increasingly widespread end-devices, e.g., mobile devices and IoT devices. We are then facing an urgent need to push the AI frontiers to the network edge so as to fully unleash the potential of the edge big data.
Therefore, the paper aims are threefold: (i) to introduce a novel Edge Intelligence architecture; (ii) to exploit several interrelated AI techniques that may be involved in an Edge Computing solution; and, (iii) to present a novel fulledge platform, called eLifeCare, which is enhanced by the In-Edge computation of AI-based techniques to perform reliable decision-making activities in a high complexity scenario such as the healthcare domiciliary hospitalization in an AAL fashion.
This paper is organized as follows. Section 2 provides an overview of related work and technologies which were investigated as background knowledge. Section 3 describes the architecture of our proposal, taking into account all the requirements coming from di erent disciplines of AI, and introduces the eLifeCare platform. Section 4 describes the possible scenarios of application speci cally designed for our approach, such as Healthcare in AAL. Finally, Section 5 discusses the proposed framework and concludes the paper, outlining future works. 2
In this section we will cover all the basics and recent state-of-the-art of Edge Intelligence and AI-based reasoning tasks useful to present the architecture of our proposed platform. 2.1
Pushing the AI frontier to the edge ecosystem that resides at the last mile
of the Internet is highly non-trivial, due to the concerns on performance, cost
and privacy. Essentially, the physical proximity between the computing and
information-generation sources promises several bene ts compared to the
traditional cloud-based computing paradigm, including low-latency, energy-e ciency,
privacy protection, reduced bandwidth consumption, on-premises and
contextawareness [
Edge Intelligence does not necessarily mean that the AI model is fully trained
or inferred at the edge, but can work in a cloud-edge-device coordination manner
via data o oading. As shown in Figure 1, as the level of Edge Intelligence goes
higher, the amount and path length of data o oading reduce. As a result, the
transmission latency of data o oading decreases, the data privacy increases
and the WAN bandwidth cost reduces. However, this is achieved at the cost
of increased computational latency and energy consumption. The surge of IoT
devices makes the Internet of Everything (IoE) a reality [
Regarding Edge Intelligence scenarios speci cally in the area of AAL, [
Recently, in [
Our research is broadly situated in the latter purpose of Edge Intelligence.
Further, our contribution is placed in a scenario where the convergence of
multiple AI-technologies allow us to automate complex decision-making activities
resulting from a multi-strategic inference approach. In fact, the scope of Edge
Intelligence should not be restricted to running AI models, referring exclusively to
Machine Learning (ML) or Deep Learning (DL) models [
The Semantic Web of Things (SWoT) is an emerging paradigm in ICT, joining
the Semantic Web and IoT. On the one side, the Semantic Web initiative aims
at allowing software agents to share, reuse and combine information available
in the World Wide Web [
SWoT environments are intrinsically dynamic: the availability of hosts, data
sources and services can vary frequently and unpredictably, due to device and
people mobility, battery limitations and wireless networks unreliability [
SWoT may enhance also ML classi cation tasks by merging ontology-based
characterizations of data distributions with non-standard reasoning for a
negrained event detection [
In AI, Abstract Argumentation is a very simple but also very powerful formalism
to reason over con icting knowledge. It studies the acceptability of arguments
based purely on their relationships and abstracted from their content. An
argument is a set of assumptions (i.e., information from which conclusions can
be drawn), together with a conclusion that can be obtained by one or more
reasoning steps. Given a problem to solve (making decision, reasoning with
uncertain information, classifying an object), arguments are di erent from proofs
in that they are defeasible, that is, a type of non-monotonic reasoning in which
the validity of their conclusions can be disputed by other arguments in the light
of new evidence. Then, Argumentation is the process by which arguments and
counterarguments are constructed and handled. Handling arguments may
involve comparing arguments, evaluating them in some respects, and judging a
constellation of arguments and counterarguments to consider whether any of
them are warranted according to some principled criterion [
Basically, Abstract Argumentation, introduced by Dung [
Building appropriate argumentation formalism can cause much more concern
than expected. Attention must be paid to avoid the risk of violating some
natural rationality postulates [
0:6 0:7
0:5
0:3
process. Here, generating proper argumentation structures is the key to
obtaining reasonable and consistent output. Dung's original formalism for abstract
argumentation has been extended along many lines giving rise to a large and
thriving literature in AI (see [
A Bipolar Weighted AF (BWAF ) [
This representation has been chosen as the most convenient and suitable
to reason over partial and/or inconsistent knowledge conveyed by devices in
an Edge Intelligence system. In fact, giving the sensors, actuators, and other
edge devices the faculty of arguing about the information they are conveying,
it follows that the whole IoT network and Edge infrastructure would become
smarter and more reliable. Devices would become able to perform operations
such as processing data on Edge so as to produce higher-level information and
autonomously deciding their own course of actions toward the achievement of
their (individual or collectively shared) goals [
In this section we rstly present a generic Edge Computing architecture that exploits several AI techniques to optimize healthcare in a SHE autonomous setting, and then we focus on eLifeCare, that is our platform in which AI on Edge is actually performed. We have to deal with and Edge Computing architecture so that processing can be done at the devices (i.e., end-nodes), or at the gateways (i.e., edge nodes). This will reduce unnecessary data tra c and processing latency, and is important for applications such as critical patient monitoring and analysis. Focus is placed on designing and developing edge nodes, with associated end-nodes for various patient monitoring applications. In this complex scenario, we deal with a large scale of heterogeneous devices as edge nodes in an IoT environment, which, at di erent levels of abstractions and di erent roles, communicate and convey di erent kinds of information: 1. End-nodes / Device Layer: { Simple, Complex sensor; { Mobile, Wearable, Embedded devices; { Actuators. 2. Edge nodes / Edge Layer: { Gateway, Sink devices; { Fog, Decider nodes. 3. Cloud Datacenter / Cloud Layer.
In an healthcare scenario, end-nodes can be distinguished in a further taxonomy, as shown in Figure 4: 1. Medical Devices : any device intended to be used for medical purposes, such as the diagnosis, prevention, monitoring, treatment, alleviation or compensation of a disease or an injury. 2. Ambient Devices : any type of consumer electronics, characterized by their ability to be perceived at-a-glance, such as motion sensors, cameras, smoke sensors, smart appliances, etc. 3. Interactive Devices : any mobile or stationary hardware component which enables the interaction between the human user and an application or the environment of the user, such as smartphones, speech recognition devices, wearable devices, etc.
Here an edge node can be nearby end-device connectable by device-to-device
(D2D) communications [
Focusing on AI, we envision a Edge Computing network capable of performing not only data analysis, classi cation, regression and/or clustering via ML/DL online training and inference, but also a wider range of AI techniques, adding a context-awareness and explainability/reliability of results. In this vision, ML/DL models are deployed in a hybrid mode which combines the decentralized training/inferring mode with a centralized revised/re ned mode.
As shown in Figure 5, the edge servers may train the ML/DL model by either decentralized updates with each other or centralized training with the cloud datacenter. The hybrid architecture is also called as Cloud-Edge-Device training due to the involved roles. As the hub of the ML/DL architecture is placed as close as possible to end-nodes and edge nodes, there is an improvement of performances regarding the training loss, convergence, privacy, communication cost, latency, and energy e ciency.
The structure of such a complex heterogeneous network of edge nodes may have a pyramidal topology, as in Figure 4. In this con guration, the nodes at the upper level are intended to have a di erent role since at a higher level of \abstraction" of the heterogeneous network, and may, therefore, have di erent tasks from the end-nodes, not excluding that they may include an exclusive partial knowledge over the entire network. Therefore, they can act as simple collectors of information deriving from groups of sensors, or from aggregators of such information, or they can make further processing steps on the edge network. A key issue in this complex scenario is that of the truthfulness and reliability of information gathered by groups of heterogeneous sensors that can: { detect con icting information within a end-nodes and/or edge nodes; { group information at di erent levels of abstraction; { perform partial processing at the Edge Layer.
In particular, an Edge Intelligence task could be performed either with partial ML/DL model training or with speci c inference and reasoning tasks on the data collected through semantic descriptions and ontological dictionaries. A certain degree of uncertainty or inconsistency is likely to arise from all the partial small amount of processed data by any type of device in the heterogeneous network.
Therefore, a key problem regards to evaluate the reliability (or justi ability) and explainability of the partial results from the entire network. Moreover, such information may be con icting, yielding the system unable to take autonomous trustable decisions. A computational model of argumentation, such as the BWAF, could solve two types of inconsistencies between con icting information: 1. in a \horizontal" way, it can resolve the con icts created by the end-nodes that transmit similar but con icting information; 2. in a \vertical" way, it can solve problems of con icting knowledge between the information (possibly aggregated) sent by the lower levels of the pyramid towards the higher levels, in which an inconsistency is found between partially disjointed knowledge between the end-nodes and edge-nodes.
The automatic building of an argumentation framework like the BWAF and its subsequent evaluation through argumentative semantics for the selection of acceptable arguments is a perfect solution in this context. Speci cally, endowing the information coming from end-nodes and edge nodes with semantic annotations allows a machine-understandable representation of information that can be exploited to automatically build the BWAF, which can give a representation of con icting and/or supporting arguments of the entire network. In particular, the non-standard inference method of semantic matchmaking between pairs of arguments is employed to de ne a weighted notion of relation between arguments (i.e., attack or support).
From the result obtained it will therefore be possible to be certain of the truthfulness and reliability of the information detected and processed at the Edge, in a heterogeneous and complex multi-IoT network, which can be crucial when a complex decision is automatically taken. 3.2
The eLifeCare platform is the Telemedicine system that contains the set of solutions and services provided by Exprivia for Teleconsulting, Telereporting, Telepresence and Telemonitoring to support all operators involved in the care and monitoring of the patient at home (see Figure 6).
The eLifeCare platform revolutionizes the approach to home patient care in that it provides the technological Edge Computing infrastructure and all the services necessary for the full, integrated management of all the care-giving processes and services, accessible and usable from any kind of IoT device: { Remote monitoring; { Telemedicine and Teleconsulting; { Medicinal product procurement monitoring; { Reporting and ling systems; { Patient's medical history in electronic folders.
The eLifeCare platform deals with the new Digital Health challenge that puts the patient at the centre and guarantees continuous services that improve quality https://www.exprivia.it/exprivia-resources/images/File/ yer-healthcare-052018/italiano/20180522 004-0 EXPITL H FS eLifeCare ITA.pdf of life and, at the same time, help to limit the costs of the local health authorities and hospitals by removing the patient from hospital. The platform is based on a WebApplication used to monitor and manage all patients in real time, acting as an operating intermediary between the patient and the medical team or specialist that is following the patient. The full-Edge architecture of the platform allows the medical team to perform AI-driven autonomous decisions for a single patient at home, according to his clinical record data on the treatment, vital parameters, diagnosis, etc. If necessary, it decides to modify the treatments, sharing speci c treatment protocols and consulting any reports in support of the treatment. The AI on Edge component of eLifeCare gathers data from devices and performs ML classi cation tasks to monitor the treatment process, endowing data with small pieces of semantic annotations inferred from healthcare ontologies on the edge nodes, and nally performing argumentation to discard the unjusti ed possible treatments in case of therapy modi cation.
The eLifeCare platform has speci c mobile applications for use by the patient that support active participation of the patient himself, the care giver, and the medical team allowing them the remote management of: { Care giving at the patient's home on the basis of the programme indicated in the Care Plans; { Measurement of vital parameters; { Videoconsulting sessions; { Patient geolocation; { Administration of pharmacological treatment; { Medicinal product procurement requests. 4
As a sample scenario, consider a SHE in which the patient at home has monitoring instruments and wearable technology, and is constantly monitored from the Edge Computing infrastructure of heterogeneous multi-IoT network devices. The Edge Layer coordinates and monitors the patient's care at home through active interaction designed to detect the care needs and critical factors involved in the care. The platform provides the operators involved with the information about the patient by opening a clinical record (medical history, clinical diary, treatment, vital parameters, etc.). The Cloud Layer acts as an intermediary: it receives any requests and/or alerts sent by the Edge Layer after the measurement of speci c vital parameters, and activates speci c operating protocols. If the Cloud-Edge-Device coordination of AI/ML models with SWoT annotation and inference and argumentation autonomously decides to make a change in the patient's treatment, the platform starts Teleconsulting or Videoconsulting session with the specialists that are following the patient to communicate the therapy variation and send a con rmation request. Using the platform, the specialist views the patient's clinical documentation and manages the care pathway.
As brie y depicted in Figure 7, the eLifeCare platform exploits AI on the Edge level to perform di erent kinds of reasoning an take complex autonomous
Fig. 7. Healthcare Scenario in AAL decisions at di erent layers of the Edge network. This is helpful in recognizing the severity, timeliness, and appropriateness of intervention among the factors determining the clinical outcome. Moreover, the platform acts as an alert systems, such as Early Warning Scores (EWS), which help in identifying speci c phases of illness and provide appropriate care. 5
Edge Intelligence, although still in its primary stage, has attracted more and more researchers and companies to get involved in studying and using it. This paper attempts to provide possible research opportunities about AI on Edge. Concretely, we rst discuss the relation between Edge Computing and AI. We believe that the synergy of multiple AI-technologies allow us to automate complex decision-making activities resulting from a multi-strategic inference approach. In fact, we state that Edge Intelligence should be the paradigm that fully exploits the available data and resources to optimize the overall performances of not only Machine Learning models, but also other inference and reasoning tasks, including Semantic Web of Things and Argumentation.
In particular, semantic-enhanced ML on heterogeneous data streams in the
IoT allows a mapping of raw data to ontology-based concept labels, providing
a low-level semantic interpretation of the statistical distribution of information,
while the conjunctive aggregation of concept components allows building
automatically a rich and meaningful representation of events during the model
training phase. Speci cally, the exploitation of non-standard inferences for
matchmaking enables a ne-grained event detection by treating the ML classi cation
problem as a resource discovery. In particular, Mini-ME [
We also showed how all these AI techniques promote and reinforce each other by presenting a novel Edge Computing architecture and the eLifeCare platform, speci cally designed for healthcare, outlining also a use case scenario in SHE.
Looking at future work, the progressive deployment of 5G networks will make available a backbone with signi cantly enhanced spectral e ciency, improved signal e ciency and signi cantly reduced latency, all factors that will facilitate new services in the healthcare domain. There will be a change of paradigm from \hospital based" to \distributed patient care".
This approach can then be re ned if dynamic networks are also taken into account, in which the Edge nodes change over time, introducing or eliminating information in the Edge network.