**2. State-of-the-art**

**1. Introduction**

116 Recent Applications in Data Clustering

Frailty and Mild Cognitive Impairment are common and inevitable conditions in the elderly citizen population defined as premature or accelerated physical and mental declines. These conditions are often an early indicator of more severe states, such as Alzheimer's disease. Control (delaying or decelerating) of the onset and progression of MCI/frailty is becoming one of the major tasks of global efforts in maintaining the functional independence and quality of life of the globally growing elderly population. In 2016, for the first time in history, estimated majority of the world population can expect to live into their sixties and beyond. Beside the global initiatives, strategies and action plans on healthy ageing conducted by the relevant key organizations such as World Health Organization (WHO), or United Nations (UN), the growing market trend for the so-called "silver economy" sector is booming. The increase of population aged 65 and over is projected to reach 28.1% of the whole population in the EU by 2050, they have a spending power higher than the generation segment aged 18 to 39, and they account for approx. 60% of total expenditures in the US and 50% in the UK, generating demand for new services and products, ranging from personalized care to age-friendly technologies and other solutions that enable the maintenance and prolongation of healthy, independent lives. Technologies and systems supporting innovative ways of influencing people's behavior

and lifestyles at all ages also present a significant economic and business opportunity.

predictive and preventive models in geriatric or healthcare practice [1, 2].

Research presented in this paper is mainly the part of **City4Age project** (www.city4ageproject.eu) that develops age-friendly **Cities and Environments in** deployments in six different prominent pilot smart cities in EU and beyond—Athens, Birmingham, Lecce, Madrid,

Geriatric practice has in this aim utilized different standardized instruments based on traditional data collection methods (administration of questionnaires, meter-based measurement or direct observation in controlled conditions) which are in most cases intrusive and demand citizens' presence in geriatric centers and a lot of time for data collection. More importantly, these methods do not enable real time monitoring of behavioral changes (e.g., data from questionnaires are collected on semi-annual or annual intervals) and thus prevent predictive and preventive interventions. Finally, data collected from such method is often subjective or incomplete. With the goal to overcome the stated drawbacks**, many technological instruments and methods emerged**, aiming to automate as much as possible the detection and mitigation of behavior deteriorations and anomalies. Particularly, the recent development and expansion of wearable technologies/devices and Internet of Things (IoT) has enabled the build-up of infrastructures of smart devices that collect vast and heterogeneous volumes of various sensory data in smart cities. These social and technological infrastructural advances potentiate the public health and prevention aspects of smart cities, transforming the urban public health from a reactive to a predictive system. In the specific area of support for (active and healthy) ageing, the transformation and progress direction of particular interest is the expansion of concept of **ambient-assisted environments**, from currently predominant implementations in residential and social indoor spaces (homes, elderly care/community centers) to outdoor and public environments. However, there is still a large gap between potential and actual IoT data exploitation because of many challenges that have to be overcome before putting data driven

World Health Organization (WHO) had recognized the importance as well as human and economic impact of age-friendly environments and launched the age-friendly Cities and Communities Programme that introduced the terms in 2006/2007, as the foundation initiative aimed for local and metropolitan governance and development levels. The European Commission (EC) supports the pursues of goals and objectives of age-friendly environments and sustainable development by numerous different instruments, primarily through R&D funding programs such as Horizon 2020 or specialized Active and Assistive Living (AAL) Programme. Important higher-level EC initiatives that foster innovation and concentrate stakeholder efforts are the recently established:


These efforts increased research efforts in the area of smart city IoT data analytics through different projects.

inclusive and collaborative system supporting health equity. The clinical focus of the project is on chronic respiratory (asthma) and metabolic diseases (type 2 Diabetes), developing risk stratification models based on risk factors in each urban location (pilot deployment in five global cities—Barcelona, Birmingham, Paris, New York and Singapore), taking account of biological, behavioral, social and environmental risk factors, community resilience and well-

Temporal Clustering for Behavior Variation and Anomaly Detection from Data Acquired…

http://dx.doi.org/10.5772/intechopen.75203

119

IGERT project, ended in 2016, funded by the US National Science Foundation grant, comprises a multi-disciplinary doctoral training programme focused on designing and studying health-assistive smart environments, with particular emphasis on automatic monitoring and analysis of human health and behavior, unsupervised data-driven detection of activity/ behavior and lifestyle changes, potential simulation/prediction of human behavior and activi-

Recently exhaustive and comprehensive reviews about temporal clustering algorithms and applications are published [5–7] and thus we will focus here only on the concepts that are closest to this research. As said behavior recognition, change and anomaly detection can be modeled naturally with clustering algorithms. Clustering techniques that allow grouping objects into homogenous groups where objects in the same group are similar (intra-cluster distance is low) and objects between groups are dissimilar (inter-cluster distance is high). Since the definition of clustering is based on the notion of similarity it is utterly important to define the notion of similarity and types of similarity measures. Unlike stationary data, time

Similarity in time: this is the simplest form of similarity with the assumption that instances that are close in time have similar values. This is a naïve assumption and is used for bench-

Similarity in shape: these similarities disregard the time of occurrence of patterns. Using this definition, clusters of time-series with similar patterns of change are constructed regardless of time points—for example, extracting groups of elderly citizens who have a common pattern in their visits to pharmacy, regardless when these pharmacy visits occur in time-series. Dynamic time warping (DTW) is one of the mostly used dissimilarity measures of this type. Structural similarity: the types of metrics used to find similarities in changes of time series. This is done by building models like AutoRegressive Moving Average (ARMA) or Hidden Markov models (HMMs), and the similarity is measured between parameters of models. In case of intra-activity behavior variations, structural similarity could recognize patterns such as: after three weeks of decrease of outdoor\_time, citizen has registered increased outdoor\_

Main challenges for the Data Science and Analytics in the related research in City4Age coincide with main generic challenges for the potential of collected IoT data from smart cities for health (and other personal) monitoring—volume and diversity of collected data is huge and

ties, and enhancement of human physical and cognitive abilities [4].

series have several aspects of similarity [7]:

time, and this behavior is repeated every month.

**3. City4Age analytic framework**

mark purposes in most of the cases.

being in cities.

Geriatric practice has in this aim utilized different standardized instruments based on traditional data collection methods (administration of questionnaires, meter-based measurement or direct observation in controlled conditions) and quantification and categorization of functional domains of daily life behavior and known frailty/MCI risk factors, such as Lawton IADL scale, Mini-Mental State Examination (MMSE), Fried Frailty Index, Nottingham Extended Activities of Daily Living, and numerous others. A comprehensive summary of such traditional generally **psychogeriatric instruments** and methods is provided in [3]. These instruments have evident major drawbacks, of late detection and problem identification (analysis and interpretation of questionnaires or conducting of exams can span intervals of months), and being generally ineffective, possibly subjective to a high degree, and costly for deployment.

The **City4Age project** (www.city4ageproject.eu), funded through the mentioned EC Horizon 2020 programme, is one of the pioneering efforts acting as a bridge between the mentioned two European Innovation Partnerships, EIP on SCC and EIP on AHA, contributing to specific and shared objectives and involving the committed participants from both Partnerships. The primary aim of the project is to enable fully Ambient Assisted age-friendly cities, through development and deployment of a range of ICT tools and services that will improve the unobtrusive early detection of MCI/frailty risks from heterogeneous IoT and smart city data sources at homes or on the move within the city, comprising the research and development work performed and results presented in this chapter as part of the work on the Data Analytics Platform. Coupled with the appropriate interventions—the developed tools will mitigate the detected risks as secondary aim. The developed system and components are being validated through in-situ deployments in six pilot smart cities.

Besides the City4Age project, there are numerous other related efforts in development of IoT driven systems for maintaining the functional independence and quality of life of the globally growing elderly population, or in development of data-driven health-related behavior recognition systems or platforms. Some of the recent relevant ones are the following:

The ActivAGE project (www.activageproject.eu), started in January 2017 and likewise funded through the Horizon 2020 programme, is a European multi-centric large-scale pilot on Smart Living Environments. The main objective is to build the first European IoT ecosystem across nine Deployment Sites (DS) in seven European countries, reusing and scaling up underlying open and proprietary IoT platforms, technologies and standards, and integrating new interfaces needed to provide interoperability across these heterogeneous platforms, that will enable the deployment and operation at large scale of active & healthy ageing IoT-based solutions and services.

Participatory Urban Living for Sustainable Environments (PULSE) project (www.pulseproject.eu), started in January 2016 and likewise funded through the Horizon 2020 programme, harvests open city data, and data from health systems, urban and remote sensors, and personal devices, to enable evidence-driven and timely management of public health events and processes, leveraging diverse data sources and big data analytics to transform urban public health from a reactive to a predictive system, and from a system focused on surveillance to an inclusive and collaborative system supporting health equity. The clinical focus of the project is on chronic respiratory (asthma) and metabolic diseases (type 2 Diabetes), developing risk stratification models based on risk factors in each urban location (pilot deployment in five global cities—Barcelona, Birmingham, Paris, New York and Singapore), taking account of biological, behavioral, social and environmental risk factors, community resilience and wellbeing in cities.

IGERT project, ended in 2016, funded by the US National Science Foundation grant, comprises a multi-disciplinary doctoral training programme focused on designing and studying health-assistive smart environments, with particular emphasis on automatic monitoring and analysis of human health and behavior, unsupervised data-driven detection of activity/ behavior and lifestyle changes, potential simulation/prediction of human behavior and activities, and enhancement of human physical and cognitive abilities [4].

Recently exhaustive and comprehensive reviews about temporal clustering algorithms and applications are published [5–7] and thus we will focus here only on the concepts that are closest to this research. As said behavior recognition, change and anomaly detection can be modeled naturally with clustering algorithms. Clustering techniques that allow grouping objects into homogenous groups where objects in the same group are similar (intra-cluster distance is low) and objects between groups are dissimilar (inter-cluster distance is high). Since the definition of clustering is based on the notion of similarity it is utterly important to define the notion of similarity and types of similarity measures. Unlike stationary data, time series have several aspects of similarity [7]:

Similarity in time: this is the simplest form of similarity with the assumption that instances that are close in time have similar values. This is a naïve assumption and is used for benchmark purposes in most of the cases.

Similarity in shape: these similarities disregard the time of occurrence of patterns. Using this definition, clusters of time-series with similar patterns of change are constructed regardless of time points—for example, extracting groups of elderly citizens who have a common pattern in their visits to pharmacy, regardless when these pharmacy visits occur in time-series. Dynamic time warping (DTW) is one of the mostly used dissimilarity measures of this type.

Structural similarity: the types of metrics used to find similarities in changes of time series. This is done by building models like AutoRegressive Moving Average (ARMA) or Hidden Markov models (HMMs), and the similarity is measured between parameters of models. In case of intra-activity behavior variations, structural similarity could recognize patterns such as: after three weeks of decrease of outdoor\_time, citizen has registered increased outdoor\_ time, and this behavior is repeated every month.
