Physiological status and physical activity, social interaction, cognitive and emotional status, and nutrition in older people are the key target areas addressed by the NESTORE project. It is aimed at developing a multi-domain solution for users, able to prolong their functional, social, and cognitive capacity by empowering, stimulating, and unobtrusively monitoring, in other words, \coaching" the user's daily activities according to a well-de ned \Active and Healthy Ageing" life-style protocols. Besides the key features of NESTORE in terms of technological solutions, this work focus on the preliminary research carried out in the context of algorithms for modelling and pro ling target individuals with the aim of developing an e ective dynamic Decision Support System.
Ageing population is growing faster in EU [
The objective of NESTORE is to develop a virtual companion that, like the mythological Nestor, can give advice to older people so that they can maintain their well-being and their independence at home, based on experience and on understanding the current situation. The experience of NESTORE is based on well-grounded psychological and behavioural theories in conjunction with relevant know-how on the ageing process, while the current user's situation is understood on the basis of a comprehensive system of sensors able to monitor the di erent key parameters of the user. An intelligent system, deployed on the cloud and leveraging Decision Support (DS) logic delivers \advise and coaching", which is o ered via the companion, embodied in a smartphone or an intelligent tangible object, according to the user's preferences and interests. NESTORE provides coaching and personalisation in ve crucial domains (called henceforth NESTORE target domains) of the Active Ageing process: i) physiological status; ii) physical activity; iii) social interaction; iv) cognitive and emotional status; v) nutrition.
In this paper, we present the core component behind the coaching activities
suggested to the user by NESTORE: an intelligent and innovative Decision
Support System (DSS). It is able to analyse the user's behaviour, tracking its changes
and its compliance to active ageing guidelines, and providing personalized target
behaviours toward the adoption and maintenance of a healthy lifestyle. A DSS
can be de ned as a computerized information system used to support
decisionmaking in which the characteristics of an individual are matched to a
computerized knowledge base [
The personalisation is built upon dynamic models fed with ve well-being
dimensions, while the DSS selects, processes, and updates indicators by
learning from past choices. In particular, NESTORE implements novel algorithms
for detecting and monitoring of important indicators related to user status and
behaviour. These algorithms are able to adapt to personal needs, emotional and
behavioural patterns, thanks to the inclusion of the well assessed Selection,
Optimization and Compensation (SOC) model strategy [
The algorithm infrastructure comprises unsupervised and semi-supervised algorithms for the inference of user's behavioural pro le and the anomaly detection. User's trends in the ve well-being dimensions will be described in one single semantically annotated model giving the possibility of inferring the necessary information to understand the peculiarities of each user among the other users and their self in di erent scenarios thus generating the appropriate feedback.
The NESTORE DSS is based on a three-layer structure: i) a short-term
analysis that analyses data on a daily basis; ii) a long-term analysis that looks
at trends and is able to detect change and adapts the coach in the long term,
following the changing needs of people as they age; iii) a combined short- and
long-term analysis to provide a personalized mix of activities for nally sending
personalized plans to the Coach when appropriate. The recognized trends are
combined in the DSS with context reasoning to provide robust recommendations
and correlations. Behavioural theories leveraging SOC and HAPA [26] models
will be embedded in the algorithms. In NESTORE, we will also adopt the
socalled \emergent" modelling perspective. With an emergent approach, the focus
is on the low-level processing: sensory data are augmented with structure and
behaviour, locally encapsulated by autonomous subsystems, which allows an
aggregated perception in the environment [
The rest of the paper is structured as follows: Section 2 shows the main components of the NESTORE Decision Support System with Section 3 describing its user pro ling process. Section 4 illustrates how the pro les are used in the DSS, while Section 5 draws the conclusions. 2
During recent years, various researches have investigated and developed new
solutions in the area of DSS. This is due to the emergence of personalised medicine
and the enhanced ability to build tools aimed at predicting personalized risk and
advice systems. Currently, most DSSs provide decision support for particular
diagnostic or therapeutic tasks such as ensuring accurate diagnosis, improved
prognosis and theragnosis, screening for preventable diseases in a timely
manner or averting adverse drug events. As regards NESTORE's eld of interest,
there have been attempts to decision support for telecare but few of these have
achieved user-speci c personalisation. The current work done in this area can
be classi ed according to their purpose as systems that:
{ Give advice, recommend care plans and trigger alerts. People at this
age require individualised care plans so that they could maintain their health
taking into account the idiosyncrasy of each individual. Care plans can give
details of dietary requirements, activity levels, targets for physical activity,
blood pressure and other tests.
{ Are based on daily life activities. Changing routines for people in their
60's is not the best way to achieve motivation. That is why it is important
to adapt the recommendations to users' current behaviour. In [
Most of the DSS analysed before cover only a narrow eld of medical knowledge or only part of the relevant factors for preventing the elderly decline are treated and transferred into the DSS. In other words, so far, the inference techniques cannot represent the rich variety of elements that a professional could recommend to a speci c person. Furthermore, in the analysed works, due to limitations in the user interface, the advice of DSS relies only on computable input data, which represent just a small proportion of the information required to make decisions. It is extremely di cult for the user to determine whether the input data adequately represent a potential problem. Most of them fail to represent common-sense knowledge and have no real understanding of the user's problem.
In NESTORE, the DSS is intended to help older people to compile useful information about their lifestyle in order to identify proper actions and make decisions to improve or maintain a healthy life. One of the primary objectives in NESTORE project is to develop a DSS so that the users can obtain fast, reliable, personalised, and directly applicable advice. Suggestions are delivered in form of coaching plans, which are divided into pathways composed of di erent coaching activities and training activities. The DSS and, concretely, a user pro ling module will be in charge of proposing the coaching plans and recommendations that better t each user based on extracted attributes. The information the nal NESTORE DSS will use is: { Models describing the NESTORE target domains; { Recommendations and guidelines; { Behavioural models and intervention techniques; { Existing knowledge from domain experts and other evidence-based sources.
User pro ling is one of the key steps in the recommendation processes since it is essential for extracting user characteristics and predicting how much a user will like an item.
As depicted in Figure 1, the user pro le and user preferences feed the DSS engine with the necessary inputs to select the most convenient coach plan for each user. In this paper, we focus on the personalisation side of the DSS, mainly embodied in the user pro ling component. It describes the way we will pro le the users with the nal aim of selecting the recommendations and coaching plans that better t the user. User pro ling can be de ned as the process of identifying the data about a user interest domain. This information can be leveraged by the DSS to better understand the user needs and, thereby, provide personalized recommendations.
The process to build the user pro ling is foreseen as follows: Step 1. Personas are designed to analyse the di erent types of information that we will need to personalize NESTORE recommendations.
Step 2. The nal set of Personas is analysed and a list of attributes is extracted from it.
Step 3. The list of attributes is complemented with other items that NESTORE domain experts believe that are important for the personalisation procedure.
Step 4. Di erent user pro ling methods are analysed. A twofold user pro le is implemented: static and dynamic.
Step 5. The data ow for recommending coaching plans is designed and di erent use cases where user pro ling will be used are envisaged.
Step 6. User pro ling module is implemented and integrated in the NESTORE
DSS. 3.1
The Inmates Are Running the Asylum [
In NESTORE, the process of creating Personas was based not only on previous research projects prepared for the development of user pro les but also on an iterative process to facilitate the transversal cooperation between the different NESTORE partners and key agents. All this process was based on the importance of re ecting the idiosyncrasies and realities to develop useful pro les for the implementation of the system.
The research was developed consulting the main European demographic public resources to detect the core characteristics of the elderly population, but also to be aware of the possible heterogeneity in this target group. Personas were designed by the co-design experts and piloting countries to include their privileged view of the real users needs and preferences. There were also taken into account the co-design experts considerations to include their privileged view of the real users needs and preferences. Domain experts considerations were also taken into account in order to introduce valuable information to enrich the global understanding of the potential NESTORE users. Another valuable feedback was obtained from the Forum Advisory Stakeholders (FAS). Suggestions and questions pointed out by FAS members were re ected in a new version of users pro le and personas document. This fruitful cooperation had, as a result, a large list of pro les (n=24). This contribution aimed to re ect the heterogeneity from the European contest.
Three tools were created to help in the process of re ning pro les. Firstly, it was created a checklist with key questions to be asked to co-design experts and pilot teams. The main purpose was to select the nal personas systematically and guide experts in the evaluation of each pro le to detect those who have more capacity to be more informative or descriptive for technical researchers and developers. Secondly, it was produced a document based on a table with two tabs, one for comparing and grouping the di erent pro les and a second tab for merging and de ning 8 contexts. This tool helped to re ne the status, preferences, and attributes. Finally, the third tool was a diagram that presents three important aspects (personal and environmental characteristics and possible pathways). This schema was crucial to highlight the needs and preferences of personas' pro les which will determine the possible elections of pathways of real users. Finally, it was proposed to create a card template to re ect the main characteristics of each pro le. This task helps to be systematic and gain consistency to build pro les.
The use of the diagram tool showed in Figure 2 helped to de ne two main
aspects to be included in the re nement of Personas, in accordance with the
Cooper de nition [
In NESTORE's pro les, it has been suggested indirectly the end goals and life goals by means of the description of personas' daily activities and main interests. For example, in some pro les spending time with family, to be involved in cultural or voluntary movements, etc. Figure 3 provides an example illustrating the building process to de ne each pro le.
NESTORE's card model includes general information such as gender, age, country or socio-economic status, and more speci c details about how many people live in the home, the main characteristics of the living space (size, existence of stairs, balcony or garden), where they live (urban or rural), web connection level, if they have pet or not. Environmental information (weather and humidity that could a ect their activities in daily living) is also provided. Finally, Personas' status in relation to the di erent domains is provided with a level, a de nition of the status and target with a narrative description that includes information about preferences and values. Figure 4 shows an example of the cards of NESTORE Personas.
Personas present diversity in relation to the di erent NESTORE target domains, in order to have a wider spectrum that could enrich the views and understanding of potential needs and preferences of future end users.
Physiological status and Physical activity domain Although NESTORE users are de ned as healthy older people, it is relevant to include di erent type of health conditions (not severe chronic diseases) very common and prevalent in the elderly population. There are acute illnesses or health conditions that could determine behaviours or a ect system functionalities, and because of this, domain experts pointed out the need to consider some conditions in the health status. Since NESTORE pathways are based on the user needs to maintain or improve a de ned physiological status, Personas included pro les with di erent physical activity levels and several behavioural targets. According to this, Personas have a wide range of physical activity level and pro les with high (2 pro les), medium (5 pro les), and low activity (3 pro les) are included. Similarly, Personas include pro les who need to improve aerobic activities such as walking but do not need stretch exercise as well as aerobically t subjects who need to increase the frequency of strength activities.
Social interaction domain Personas were de ned to describe di erent living conditions, even though the majority of pro les are characterized by medium or high levels of interaction. Personas are retired or working part-time, or taking care of grandchildren or other family members. Also, there are very active proles, involved in volunteering activities, hobbies (music, reading, travel, etc.), doing cultural or training activities. But it was decided to also include perceptions of some loneliness in some pro les that could a ect the perception of quality of interactions with others. We also de ned the use of social networks. Pathways considered in the social interaction domain were de ned to maintain or improve a persons social opportunities or skills.
Cognitive and emotional domain Personas were described to include a broad range of cognitive and emotional status. The majority of pro les (n=7) have a good cognitive status, but they could be worried to maintain it, or they could be worried about future memory loss. Because of this, the personas' pro les have interest in pathways such as: \maintain cognitive skills", \maintain/improve memory" or \maintain/improve daily mental skills". We introduced three proles with low/medium status that includes: memory loss, depressive symptoms, emotional or mood problems.
Nutrition domain The majority of participants have a well-balanced diet, but they want to improve some aspects as the diversity of menus, introduce some foods and nutrients such as proteins or bre from vegetables or fruits, or reduce others as cakes, fats, etc. Some of them need to increase the intake of water. Also, it was included two pro les with digestive problems, to help identify other needs and preferences that could a ect the diet behaviour or food selections. Two Personas are overweight, but their target was de ned to diversify menus and balance their diet because it is possible that in existence of overweight problems the user decides that he/she does not want to reduce body weight or fat mass. However, the NESTORE System will rstly encourage him/her to lose weight (explaining the bene ts, risk factors, etc.). If users continue interested in diet, then NESTORE system will understand their needs and preferences in order to propose a pathway that includes some activities which could encourage a behaviour change, if possible. Also, four Personas have a di erent diet pro le because one has food allergies, one has lactose intolerance, and two are vegan or vegetarian in order to introduce some diversi cation in pro ling. 3.2
After analysing Personas and complementing the information with domain experts, it is proposed a twofold user pro le: { Static pro le. It is formed by the status and preferences of the user and it is characterized by containing non-varying attributes. Concretely it includes demographic characteristics, attributes regarding the context where the user lives, physical and physiological aspects and baseline data of the various domains. { Dynamic pro le. It is built dynamically while receiving data from sensors, applications and contextual APIs. It is foreseen to receive daily indicators about the di erent domains and also contextual information (i.e. current weather conditions).
Static pro ling is the process of analysing a user's static and predictable characteristics. Users' static features comprehend factual data, such as the idiosyncrasy of their residence (e.g. do they live in a rural or in an urban area?), or their diet routines (e.g. is meat part of their diet?), as well as inter-individual di erences in the other NESTORE domains (marital status and perceptions of loneliness, cognitive functioning, physical tness, etc.). They also describe the environment and context of users. One of the uses of static pro ling will be the cluster of users, the resulting groups of which will be inputted into the DSS to make thoughtful recommendations. Considering that real data will not be available until the pilots take o , a data simulator has been implemented to cope with the absence of data creating, thus, solid fundamentals for the clustering process and the recommender system. Getting into detail, the static pro le simulator generates a population of users who is described by its fact-based properties.
Dynamic pro ling is the process of analysing data coming at run-time from the sensors and applications deployed in the NESTORE user's ecosystem. It describes the changing context of the user, which is the element that leads the personalisation process.
To build the static pro le of a user, not only the three well-known well-being
domains (i.e., physiological status and physical activity behaviour, cognitive and
social behaviour, and nutrition) need to be considered. The user's context is a
quite new feature in user pro ling that will help to characterise the situation
of the user. There are di erent types of contexts or contextual information that
can be modelled within a user pro le [
After deciding on the obtainable information to pro le the users, a collection of variables with the values they can take has been de ned. Those have been split per kind of contextual feature (demographic and environmental) and per wellbeing domain (physiological status and physical activity behaviour, nutrition, cognitive and mental status, and social behaviour). Besides, a category called activities has been added to include the user's routines and preferences. It should be noted that the following variables are not all the properties indicated by the domain experts. This is due to the uncertainty about the setting in the pilots at the current stage and for this reason, some of the variables have been temporarily dismissed since they may not be measurable.
Demographic information is to a great degree relevant to group people according to their culture and generation. Due to the scope of the DSS, there is no need of depicting the participant's culture. Table 1 shows the variables which best characterize users' demographic context.
The environmental context captures the entities that surround the user. These entities can, for instance, be services, temperature, light, humidity, noise, and people [24]. Table 2 displays a compendium of variables that provide contextual information about the environment of users.
Routine activities of users, as well as their preferences, should be taken into account to provide personalized recommendations that adapt to their lifestyle. The variables that will be considered are shown in Table 3. less than 30; [30,60]; [61,120]; more than 120 Daily, 2-3 times per week; weekly
Walk, bike, swim, golf, dance, extreme sport, etc.
The factual data that best describes the physiological status of users mainly comes from their anthropometric characteristics, presented in Table 4.
The nutritional domain will basically be characterized by the dynamic pro le. Only the list of refused foods will be considered to create the nutritional static pro le, as it is presented in Table 5.
The variables that describe the cognitive and mental status of users can be found in Table 6.
Finally, the static social integration level can be characterized by the factual information contained in Table 7.
Various sensors and applications deployed in the NESTORE platform generate, at run-time, input data to the DSS. An environmental monitoring system is
Good; medium; low Positive and negative a ect; life satisfaction; depressive symptoms; cognitive functioning (test-based); memory failures (self-reported); loneliness; social integration
Good; medium; low Friends; volunteering/working; family Daily; 2-3 times per week; weekly; monthly; yearly
Friends; association; activism; volunteer deployed in the NESTORE user environment as an ensemble of wireless sensors able to sense the variables indicated by the domain experts in the relative NESTORE target domains. Furthermore, it has the aim of detecting the interaction of the user with the environment and monitoring the status of the environment itself (e.g., indoor air quality). Also, an innovative wearable device is expected to be worn by the NESTORE user. It is able to detect physiological parameters (e.g., heart rate, steps, distance, sedentariness, stairs, energy expenditure, etc) while the user performs the activities suggested by the NESTORE virtual coach.
Re ecting the separation of concerns of all the data generator deployed in the NESTORE environment, we call environmental device any sensor deployed in the user's vital space, while wearable the device worn by the user during his daily activities. As a further source of information about the user's status, we have derived data as result of a computation or fusing strategy and data coming from a direct input of the user, as questionnaires while interacting with the NESTORE coach. We call the latter soft data. Table 8 describes how the device types (wearable, environmental, and soft data) cover the variables indicated as needed by domain experts for each NESTORE target domain. NESTORE DOMAIN
VARIABLES
From the devices point of view, besides the wearable device, in the form of
a smart wristband, we plan to deploy a smart scale to collect anthropometric,
musculoskeletal characteristics and balance [
Physical Activity Behaviour Cardiorespiratory Exercise Capacity Cardiovascular System Respiratory System Strength-Balance-Flexibility Exercise Capacity Anthropometric Characteristics Musculoskeletal System Sleep Quality Energy Expenditure Nutrition Habits Cognitive Status Mental Status Mental Behaviour and States Social Behaviour
DEVICE TYPE
Wearable
ceived Signal Strenght Indicators (RSSIs) [
outdoor indoor Internet
NESTORE
Cloud Infrastructure
From the architectural and deployment point of view, in order to reduce the
e ort needed by the end user to install and use the environmental sensors, we
chose to adopt a Web of Things (WoT) approach. WoT is a computing
concept that describes an environment where everyday objects are fully integrated
with the Web. The prerequisite for WoT is for the \things" to have embedded
computer systems that enable communication with the Web. Such smart
devices would then be able to communicate with each other using existing Web
standards. Considered a subset of the Internet of Things (IoT), WoT focuses on
software standards and frameworks such as REST, HTTP and URIs to create
applications and services that combine and interact with a variety of network
devices. The key point is that this doesn't involve the development of new
communication paradigms because existing standards are used [
A general work ow illustrating how the two types of the previously described pro les are used in the DSS is depicted in Figure 6. As shown in the picture, we foresee to generate clusters or groups of users taking into account a number of static pro les generated with the simulator described in Section 3.2. Afterwards, experts will select the best type of coaching activities to be recommended to each group.
The recommendation process carried out in the DSS will follow the Health Action Process Approach (HAPA) model [25]. In this paper, we summarize the DSS personalization process in three main steps based on the phases described in the HAPA model and the three levels of coaching of NESTORE: Step 1. After acquiring the needed information about the user for building their static pro le (around 2 weeks), the system proposes a list of pathways that correspond to the detected weaker aspects (this is the result of the phases 1 and 2 of the HAPA model). The user selects one of the pathways to focus during the following weeks (this corresponds to phase 3 of the HAPA model).
Step 2. Each pathway has a prede ned list of coaching activities, but not all of them apply to all the users. In this step, the DSS selects the subset of coaching activities that better t the group where the user belongs to. Step 3. Training activities or recommendations related to the di erent chosen coaching activities are suggested to the user depending on the dynamic pro le data. The user, nally decides which activities he wants to perform.
In the following, we explain all the steps with a concrete example. After two weeks gathering information about the user through the sensing system and the coach, the pro ler module in the DSS constructs the static pro le shown in Table 9. It contains the necessary information to characterize the baseSocial (baseline) line of the user in di erent general aspects and in all NESTORE's domains of interest. This information allows the system to infer which pathways to propose and which groups does the user belong to.
Let's say that the example user belongs to group A. The system could interpret that the weakest points are physical activity and social domains, so it proposes the following pathways: { Improve tness level { Improve social activities { Retain healthy eating { Retain memory { Climb stairs (2 oors) { Track your steps { Nordic walk { Walk on the beach { Track your nutrition { Cook new recipes Assuming that the user selects \Improve tness level" as their main objective, the DSS takes the subset of activities of group A that belongs to this pathway and some other activities from the other pathways (step 2). For example: Then, the coaching phase starts. The DSS creates on a daily or weekly basis a dynamic pro le that will permit to personalize and contextualize the recommendations even more. An example of a dynamic pro le is listed in Table 10. Daily/weekly data Context
Total PA Level Nutrition level [...] Location Weather [...]
Low Low calcium intake Barcelona Sunny, 26°
The activities and recommendations proposed to the user throughout the day could be: { (At 8:30) Add more milk to your morning co ee! { Today its sunny, why dont you go for a walk on the beach? { (Afternoon) Go for a nice long walk with your dog! Finally, the user can decide which activity he wants to perform.
The NESTORE project addresses ve important domain a ecting their active and healthy ageing trajectories: physiological status and physical activity, social interaction, cognitive and emotional status, and nutrition. This is achieved by designing a multi-domain solution aimed at coaching older people toward an active and healthy ageing lifestyle protocol.
We presented the core component behind the coaching activities suggested by NESTORE: an intelligent and innovative Decision Support System. It is able to analyse the user's behaviour, tracking its changes and its compliance to active ageing guidelines, and providing personalized target behaviours for the adoption and maintenance of a healthy lifestyle.
The system is going to be deployed and validated in 60 pilot sites across Europe during the next year. At the current stage of the project, the overall system is under development leveraging the extensive research carried out in the context of algorithms for modelling and pro le the target users. The output of this process represents the core focus of the paper.
This work has been funded in the framework of the EU H2020 project \Novel Empowering Solutions and Technologies for Older people to Retain Everyday life activities" (NESTORE), GA769643. The authors wish to thank all the project partners for their contribution to the project. 23. Safety, O., Administration, H., et al.: Indoor air quality in commercial and institutional buildings. Maroon Ebooks (2015) 24. Schia no, S., Amandi, A.: Intelligent user pro ling. In: Arti cial Intelligence An
International Perspective, pp. 193{216. Springer (2009) 25. Schwarzer, R.: Modeling health behavior change: How to predict and modify the adoption and maintenance of health behaviors. Applied psychology 57(1), 1{29 (2008) 26. Schwarzer, R., Luszczynska, A.: How to overcome health-compromising behaviors: The health action process approach. European Psychologist 13(2), 141{151 (2008)