=Paper=
{{Paper
|id=Vol-3367/paper4
|storemode=property
|title=ChAALenge: An Ambient Assisted Living Project to Promote an Active and Health Ageing
|pdfUrl=https://ceur-ws.org/Vol-3367/paper4.pdf
|volume=Vol-3367
|authors=Dimitri Belli,Paolo Barsocchi,Edoardo Gabrielli,Davide La Rosa,Vittorio Miori,Filippo Palumbo,Dario Russo,Gabriele Tolomei
|dblpUrl=https://dblp.org/rec/conf/aiia/BelliBGRMPRT22
}}
==ChAALenge: An Ambient Assisted Living Project to Promote an Active and Health Ageing==
https://ceur-ws.org/Vol-3367/paper4.pdf
ChAALenge: An Ambient Assisted Living Project to
Promote an Active and Health Ageing
Paolo Barsocchi1,† , Dimitri Belli1,*,† , Edoardo Gabrielli2,† , Davide La Rosa1,† ,
Vittorio Miori1,† , Filippo Palumbo1,† , Dario Russo1,† and Gabriele Tolomei2,†
1
Institute of Information Science and Technologies, National Research Council (CNR), Via G. Moruzzi 1, 56124, Pisa (PI),
Italy
2
Sapienza University of Rome, Piazzale Aldo Moro 5, 00185, Roma (RM), Italy
Abstract
The rapid growth of older population in the next years will lead to the rapid growth in the demand
of health care, resulting in an increasing difficulty in managing hospitalizations and in a prohibitive
grow of costs for medical care. In this context, chronic heart failure emerges as one of the most difficult
problems to be treated, especially in advanced age, and a major cause of hospitalization and death. The
Project ChAALenge aims at facing the problem by proposing a proactive approach based on pervasive
monitoring and artificial intelligence. The goal is to promptly stepping, before the pathology onset, with
effective suggestions ranging from the request of medical examination to the adjustment of lifestyle.
The current paper presents the mid-term results of the ongoing project, introducing the sensors, the
middleware and the candidate artificial intelligence techniques constituting the predictive system of the
older adults’ health status.
Keywords
Ambient Assisted Living, Health Monitoring System, Health Ageing, Anomaly Detection
1. Introduction
The EU 65 or more year-old population bracket is estimated to grow faster in the next years, if
compared to the population bracket aged between 15 and 64 year-old, reaching the 50% in 2060.
Since the health care demand grows with the age, according to the previous estimation it follows
that such demand in the next years will grow proportionally [1]. In the plethora of diseases that
can occur in a chronic form in the older population, heart failure is one of the major. Since its
first methodical documentation in early nineties, heart failure has increased in the new century
AIxAS’22: The Third Italian Workshop on Artificial Intelligence for an Ageing Society, November 28th - December 2nd,
2022, Udine (UD), IT.
*
Corresponding author.
†
These authors contributed equally.
$ paolo.barsocchi@isti.cnr.it (P. Barsocchi); dimitri.belli@isti.cnr.it (D. Belli);
gabrielli.1693726@studenti.uniroma1.it (E. Gabrielli); davide.larosa@isti.cnr.it (D. La Rosa); vittorio.miori@isti.cnr.it
(V. Miori); filippo.palumbo@isti.cnr.it (F. Palumbo); dario.russo@isti.cnr.it (D. Russo); tolomei@di.uniroma1.it
(G. Tolomei)
0000-0002-6862-7593 (P. Barsocchi); 0000-0003-1491-6450 (D. Belli); 0000-0002-9573-3615 (E. Gabrielli);
0000-0001-8424-0762 (D. La Rosa); 0000-0001-9778-7142 (F. Palumbo); 0000-0003-3409-6189 (D. Russo);
0000-0001-7471-6659 (G. Tolomei)
© 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073
CEUR Workshop Proceedings (CEUR-WS.org)
to become the more frequent cause of hospitalization and mortality in the ageing population [2].
The mean age and diagnosis is 76 year-old, and the percentage of incidence increases rapidly in
the transition to old age. Specifically, incidence is 2% in people aged 40-59 years, 5% in people
aged 60-69 years, and 10% in people with more than 70 year-old [3]. Despite the introduction of
effective cures, the mortality rate still remains too high [4]. To date, heart failure is the cause of
20% recovery of more than 65 year-old, and the likelihood of death after the diagnosis goes from
the 20% of the first year to the 50% over the next 4 years [5]. In Italy, the hospitalization for heart
failure mainly involves older people, and it has been demonstrated that there is a progressive
increase of hospitalization for people at high risk of developing chronic heart failure as they
age [6]. Effective and efficient solutions to the problem stand on the promotion of a healthy
lifestyle and an active and balanced ageing. However, classical health treatment are inadequate
for the majority of older heart patients due to difficulties in constant monitoring, inadequate
assessment of clinical profile, and poor communication between healthcare workers.
Recently, the American College of Cardiology and the American Heart Association published
a series of guidelines for the evaluation and treatment of chronic heart failure [7]. They propose
a four-stage classification approach, as follows:
• Stage A: high heart failure risk without apparent structural abnormalities of the heart.
• Stage B: structural abnormalities of the heart without past heart failure trauma.
• Stage C: structural abnormalities of the heart and current or past heart failure symptoms.
• Stage D: refractory symptoms to standard treatments.
Such classification highlights the need of large-scale treatment strategies to face the under-
estimation of the progressive nature of the heart failure as pathology related to advancing
age. Fundamental requirement to pursue such goal is an integrated system capable of seam-
lessly monitoring and collecting vital parameters of the user. One solution is provided by the
Care@Home methodology and the OMNIAPLACE 1 platform developed by eResult [8]. The
project aims at implementing a highly specialized health care monitoring system capable of
providing a customized mechanism of health status forecasting of older people, that enables
technologies to undertake measures to prevent the onset of the disorder. The goal is to anticipate
the heart failure onset by proactively providing to either the patient directly concerned, or his
general practitioner, insights on the worsening of the patient’s health status. The clues may
highlight the need of specialized visits, suggest a change from a dissolute lifestyle to a healthier
one, and so on. To allow ad personam solutions, the project introduces the concept of virtual
badge, which consists in an highly customized profile of the patient built on data collected by
sensors. The Project ChAALenge aims to evolve the Care@Home methodology, by proposing
a system able to promptly detect worsening of the older user’s health status and consider, as
above, different stages for the aggravation. To date the project is still in its infancy, for this
reason we do not have rich physiological data nor environmental and periodical information yet
(e.g., coughing, blood oxygen percentage weight and so forth). In the present work we focus on
the description of the main components of the system and the preliminary analysis performed
over synthetic time series data to test the current state of our health monitoring system.
1
Omniacare official website: https://www.omniaplace.it/en-us/Solutions/OmniaCare
1.1. Related Initiatives
In the last decades, a growing number of research projects and commercial products have dealt
with the development of e-Health solutions to promote Active and Health Ageing. Several EU
funded projects have worked on Information and Communication Technology (ICT) solutions
to early detect risks in different aspects of older people’s lives. For example, the main ambitions
of the EU H2020 My-AHA project2 are the early risk detection and intervention to support
healthy aging both in the physical and cognitive domain. The project relies on the deployment
of AAL sensors, wearable devices, and smartphones, and through big data analysis to engage
users in improving their lifestyle. EU H2020 projects GrowMeUp3 and Radio4 provide integrated
and services through robotic-based approaches to encourage older persons to stay active longer.
PreventIT5 and REACH6 are focused on monitoring the users’ physical activity through the
deployment of wearable and/or ambient sensors. Additional European research projects on the
coaching of older adults which feature common elements to ChAALenge are: NESTORE [9],
The CAPTAIN System [10], COACH Council of Coaches [11], and HOLOBALANCE [12]. All of
these projects focus their attention to specific sets of domains of the user’s life and on providing
coaching systems to assist them.
The project ChAALenge is intended to overcome the limitations of vertical solutions with a
strong holistic approach by addressing several aspects of the user’s lifestyle and by considering
them from the user point of view and offering a framework that in full is able to automatically
exchange information with integrated and intelligent third-party systems.
The rest of the paper is structured as follows. Section 2 gives details of the project and its
goals. Section 3 introduces the middleware. Section 4 introduces the main ML techniques
considered so far, gives insights about anomaly detection in multivariate time series, and briefly
present mid-term results. The last Section draws conclusions and provides suggestions for
future advancements in the topic.
2. The Project ChAALenge
Heart failure in older adults represents a real problem of health care that, other than the obvious
troubles for the affected, is capable of absorbing an ever increasing volume of public/private
financial resources and manpower. The project ChAALenge introduces a novel approach to
face the issue, by leveraging the synergistic use of ICT and Machine Learning (ML). The goal is
to prevent both onset and worsening of the disease by taking advantage of an Ambient Assisted
Living (AAL) system, to monitor the patient’s habits in their everyday life environments, to
build a customized profile to be used to identify changes in their routines that may suggest an
exacerbation of their health status. Yet maintaining timely responses in critical situations, the
project focuses on the development of a middleware infrastructure to ease the communication
between sensors and actuators for the collection of health information and a software with highly
2
http://www.activeageing.unito.it/en/home
3
https://cordis.europa.eu/project/rcn/194088/es
4
http://radio-project.eu/
5
http://www.preventit.eu/
6
http://reach2020.eu/
customized predictive skills capable of foresee, and then avoid, the onset of acute issues. As the
health status gets worse, the patient is encouraged to contact their own general practitioner or
cardiologist based on the severity of the aggravation. For instance, by providing to the patient,
their relatives, or directly to the attending physician reports from which infer the need of more
frequent or specific controls, suggestion for a healthier lifestyle, or warnings indicating potential
risk to health conditions.
2.1. ICT essential
The user’s conditions in everyday situations and the support of specific services can deeply
affect their quality of life, independently of the health status. Accordingly, it is important the
enhancement of their everyday living environment (i.e., home) as primary place of care, for
instance, through the support of information technology and AAL. Based on such assumptions,
home care strategies perfectly fit the Mark Weiser’s ambient intelligence vision [13]. To be
effective and efficient, technologies should guarantee some requirements like: the integration in
living environments (embedding); the exchange of information with each other (interoperability);
the detection of both the user and the context in which he/she operates (context-aware); the
adaptation to new scenarios and the user’s needs, including changes (proactivity, adaptivity, and
customization); the ability of hiding to the user the complexity they are made of, disappearing
into living environments in a non-intrusive way (transparency). The effectiveness of the health
status prediction depends on the knowledge of the patient, their current health status, habits,
and changes. Since we are different from each other, individualization and customization become
the basic components in the patient’s dynamic profiling. Constant monitoring and virtual badge
(i.e., profile customization) are the enabling requirements for the detection of worsening in
the health status of the patient, preamble of the heart failure onset. This requires the use of
efficient methods and techniques for collecting, transmitting, receiving and processing data.
In the following we introduce the main sensor nodes made available by project partners, by
reserving the description of the middleware to the next section.
2.2. Monitoring Parameters
To better select the most significant behaviors and habits to be monitored for heart failure
onset detection, we held multiple working sessions with an expert cardiologist. Based on
the technologies made available by the project partners, the physician suggested potential
parameters to be monitored based on their own clinical experience rather than through classic
clinical parameters made available by official prevention and treatment protocols. Table 1
gathers meaningful parameters to be monitored, their descriptions, the time duration, and the
associated sensor technologies in use.
The patient health monitoring system comprehends the following cyclical steps:
• collection of information of both patient’s health status and environment by means of a
monitoring service software module;
• data processing and data analytic through an information management module;
• introduction of ML techniques and use of a decision-making application inferring the
diagnosis based on data processing results;
Table 1
Environmental Monitoring for Detecting Health Status Worsening
Parameter Variation over time Duration Device(s) and Sensor(s)
(minimal)
Weight Sudden increase of body weight 10 days OMRON - VIVA weighing
scale
Cough Increase of night coughing 30 days Audio robot UnivPM, micro-
phones
Stairs Time required to ascent and descent stairs 30 days App WN Lab (CNR-ISTI),
passive infrared sensor,
proximity sensor, indoor
positioning system
Daily rest time, especially after physical activ-
Rest 30 days Pressure sensor under the
ity like ascending and descending stairs
pillow/chair/sofa, proxim-
ity sensor, App WN Lab
(CNR-ISTI), indoor position-
ing system
Reduction over time of the distance traveled
Distance 30 days Cyclette eResult for Serious
on foot, by bicycle, etc.
traveled Game
Reduction over time of peripheral oxygen sat-
SpO2 10 days Non-invasive pulse oxime-
uration
try sensor technology by
CNR-IMM
Bathroom Increase in daily access to the bathroom 20 days App WN Lab (CNR-ISTI),
passive infrared sensor, in-
door positioning system
technology
• translation from decisions to commands and forwarding to output devices (i.e., home
automation actuators capable of changing the current environment settings based on
processed data and the patient’s reactions).
The system functioning can be interpreted as a never-ending loop, whose steps are sequen-
tially executed each time occurs an update in the state of a device belonging to the monitoring
domotic environment.
2.3. Scouting for additional sensors
Table 1 introduces just a part of the sensors made available by all project partners, along with
the list of significant parameters to be monitored for the detection of heart failure onset. Other
sensors are expected to be employed during the progress of the project, based on the need to
collect information on patient’s vital parameters, and according to the application scenarios
that will be taken into account. Currently, the additional equipment made available by project
partners consists of various advanced tools, mainly for noninvasive monitoring of health status
parameters. The sensory node Indoor Activity Analytic, for instance, is in charge of detecting
Figure 1: Enclosured RPS monitoring unit (Source: CNR-IMM).
falls and recognize the Activities of Daily Life by means of position and postural joints tracking.
Another sensory unit available is the Radar Physiological Sensing (RPS). The node is composed
by an ultra-wide-band Radar PulsON® P410 series with annexes software modules for signal
processing and control, hosted on a real-time processing mini-PC unit. The enclosure of the
monitoring unit is inside a shielded box, with a single opening and silicon coating to mitigate
reflection effects that may affect the intensity of radar signals.
Figure 1 shows an image of the RPS monitoring unit. The information sensed by RPS are hearth
and respiratory rate. The sensing operation is made in a noninvasive way by leveraging radar
signal processing and ML techniques. The procedure comprehends wizard calibration, band-pass
filtering, clutter noise attenuation, signal quantization and sampling, cardio-respiratory signal
reconstruction through parametric optimization, signal extraction and heart rate estimation.
Because of the very wide frequency spectrum in which operates, the module shows strong
penetration ability of the impulse radio, resulting in detecting the position of the patients
even in presence of obstacles (e.g., walls). Moreover, the use of very short radio pulses (i.e.,
ps order) allows to maintain a very low power spectral density (<40dBm/MHZ), restricting
in this way jamming with other radio signals. Ultimately, the hybrid architectural nature of
the module, capable of simultaneously operate as both radar sensor and radio transceiver,
opens up the opportunity of applications in Body Area Network scenarios. Another couple
of sensory nodes made available for the project are the Visual-based Vital Signs & Emotion
Detector and the Wearable Vital Signs Monitoring & Fall Detection. Briefly, the Visual-based
Vital Signs & Emotion Detector [14] is in charge of estimating vital parameters such as heart
and respiratory rate, as well as oxygenation and facial expression of the monitored patient.
The monitoring of the emotional status of the patient, additionally to vital parameters, gives
further clues about the patient’s psychophysiology status. The node comprehends Logitech
C920 HD Pro webcams, which are very lightweight devices with 78° fixed diagonal field of view,
auto-focus and automatic lighting correction. All such characteristics make the node versatile
for the installation in a everyday life environment. The Wearable Vital Signs Monitoring & Fall
Figure 2: Smartex WWS for monitoring heart and respiratory rate.
Detection [15], instead, is responsible for fall event detection, posture monitoring, and other
vital parameter for the assessment of health status and stress conditions. To the purpose, the
sensory node may be composed of one of three little invasive and comfortable tools, that is:
the Zephyr Bioharness37 , the Smartex Wearable Wellness System (WWS) shirt/belt8 , and the
ShimmerECG kit9 .
All the three tools, equipped with sensors and short- mid-range communication interfaces,
are capable of collecting and wirelessly exchange information on heart rate, respiratory rate,
and static/dynamic chest acceleration. The systems present similarities and differences with
each other. For instance, both WWS and Bioharness3 share data through wireless interfaces
but, while the former takes advantage of textile sensors to collect vital parameters, the latter
leverages the plethysmography technique for sensing operations. Smartex WWS is present in
the form of either T-shirts (all sizes) or a chest belt (adjustable). The sole difference is in the fit.
Both models have two textile electrodes for detecting the heart rate and a piezoresistive sensor
for detecting the respiratory rate. In the T-shirt case, the sensors are connected to a circuit
board properly located in a slot inside the clothing. Figure 2 reports the circuit board and both
the WWS T-shirt and the WWS belt. Briefly, the ShimmerECG kit integrates an accelerometer, a
plethysmographic sensor (to be placed on the finger or earlobe), pre-gelled electrodes to detect
heart/respiratory rates and a Bluetooth communication interface for data transmission. The
manufacturing company provides both a chest belt (see Figure 3) and a wristband equipped
with the above sensors to monitor the interested parameters. The synergistic use of all the
previous tools enable to collect posture, fall event detection, respiratory and heart rate. Last but
not least, among other (invasive) sensors made available by the project partners, it is worth
7
https://www.zephyranywhere.com/system/components (accessed on 24 June 2022).
8
https://www.smartex.it/wearable-wellness-system (accessed on 24 June 2022).
9
https://shimmersensing.com/product/shimmer3-ecg-unit-2/ (accessed on 24 June 2022).
Figure 3: Shimmer cardio kit for monitoring heart and respiratory rate.
remembering the Glucose Skin Sensor that by means of an adhesive, waterproof, deformable
and sensorized patch allows the indirect monitoring of blood glucose (by sweat sampling) [16].
As stated above, based on the need of additional vital parameter information, other sensors may
be considered as the project evolves. It is worth to point out that the input core of the project is
based on physiological data. Environmental information are temporarily considered subsidiary,
as well as less invasive sporadic measurements like, e.g., the number of daily bathroom accesses
or the resting time after climbing the stairs.
3. The Middleware
One of the main objectives of the project is to design a software architecture able to implement
personal monitoring through indoor/outdoor sensors. The architecture must interface with
intelligent devices, collect the data provided by sensors, transmit and store them for their use
by analysis algorithms. The proposed architecture is based on the experience accumulated in
previous projects, involving different scenarios such as: i) the development of a platform for
monitoring older users’ lifestyles to contrast sedentary, malnutrition, and cognitive decline [17];
ii) the implementation of a fully interoperable and context-aware domotic system [18]; iii) the
design of a communication platform to integrate mobile and wearable devices with the existing
pervasive environments [19, 20]; iv) the development of an unobtrusive monitoring system to
evaluate of the user’s sleep quality [21].
To monitor a person inside and outside their living environment is necessary to enrich the
spaces they frequent and their body with intelligent devices capable of measuring quantities
relating to their surroundings and their person. The software then transforms the measured
quantities into signals comprehensible to information systems for their management and
storage in appropriate databases. After having identified the use cases and chosen the hardware
components (i.e. sensors and actuators), the software architecture must be designed and
developed with the capability of:
• Acquiring the quantities detected by intelligent devices: The software must be able to
interface with each intelligent device via special gateways capable of exploiting the
devices’ outward communication mechanisms. The mode of interfacing depends mainly
on two factors: the transmission medium and the communication protocol. Transmission
medium can belong to two categories: wireless and wired. Typical examples of wireless
transmission media are Bluetooth, NFC, Wi-Fi and InfraRed. Typical examples of wired
protocols are Ethernet, USB and Thunderbolt. Transmission protocols, on the other hand,
define how two or more entities communicate. The best known, standardised and open
protocols are HTTP, TCP, UDP, SOAP and UPNP;
• Carrying out possible data format conversions and/or normalisation: The data from the
sensors may not conform to the model used by the platform. For example, the data
might be expressed in a different unit of measurement from the one used or need further
processing, such as normalisation, for their use;
• Storing environmental and patient data: Every single value retrieved by sensors must be
stored to not lose the information over time. Typically, data are stored in specific software
structures capable of organising and quickly retrieving historical data (databases) using
tools that interface information storage systems.
The interactions between the designed middleware platform and the main hardware and
software actors are shown in Figure 4. The sensing node is a data emitting entity characterized
by a set of properties that periodically produce information with a predefined format (e.g. tem-
perature, motion and pressure sensors). The actuator is a service provider entity characterized
by a set of properties on which an action can be invoked and from which a result can be received
in response. The application is a software module able to both provide a service, to generate
data and to actively search for sensors and actuators available on the platform.
Once the data has been collected, intelligent algorithms must retrieve and process them,
combining environmental and patient body information to identify changes in the patient’s
state of health. For this purpose, it is necessary to design the middleware with the ability to
also retrieve historical environmental and patient data, interfacing the algorithms with the
information storage systems containing the measures.
In this scenario, at the centre of the architecture, a software component that works as
middleware (Figure 5) must implement a network infrastructure to interconnect devices using
different communication protocols with the health services. To this end, each communication
standard is represented as a sub-network connected to the middleware core through modules
that work as gateways (e.g. KNX Gateway, Bluetooth Gateway, ZigBee Gateway) between the
wired and wireless device busses and the domain logic of the middleware core. The main task of
the middleware is to route messages between gateways correctly. In this way, devices can send
and receive commands and notifications from/to other devices belonging to sub-networks. In
addition, the middleware uses a complete abstraction of heterogeneous technologies to represent
devices, services, interactions and events. The abstraction layer acts as a common language
for middleware components interoperability and is mainly composed of two sub-languages:
one describes the characteristics of the devices, the available functions (services), the processes
through which interactions with other devices must take place, and the models of standardised
data types that provide a suitable intermediate representation to allow information marshalling
between heterogeneous technologies; the other describes the messages exchanged throughout
the framework. Messages that can be of two types: (i) command, when requesting the execution
Middleware
announce
sensor
<> Physiological
sensor
manage
sensor Sensing node
<>
remove
sensor
send data
Environmental Environmental
actuator sensor
manage
service
<>
<>
remove service Application
announce Serious game
Actuator service
provide service
search active Machine
sensors Learning
Robot search active
services Algorithm
<>
<>
invoke service subscribe to
service
Figure 4: Interactions between the main actors and the middleware platform.
of a service belonging to a device; (ii) event, both when a change of state occurs in a device or
when the device sends the periodic updates of detected values.
Once the values are received from the devices, the middleware stores them in a database
and, if needed, it sends them immediately to the ML algorithm for their real-time processing.
However, ML algorithms may also request historical patient data to the middleware that, in
turn, queries the database to produce the required information to the requester. Moreover, the
middleware provides a platform able to run on both desktop and mobile environments ensuring
the privacy of the transmitted user data by operating secure connection channels.
4. Predicting Variations on Health Status
To make customized predictions about the health status of each patient, the data collected
through sensors must be processed by an ML algorithm. To date, ML models considered for tests
fall into the domain of unsupervised and semi-supervised learning. The rationale is that making
predictions on health status of a patient is a domain with no (or with a very restricted number
of) labeled data. Accordingly, we temporarily ruled out binary and multi-class techniques. The
current focus is on unsupervised learning techniques, with a specific interest for multi-scale,
attention-based autoencoders that leverage the synergistic use of convolution and recurrence.
Health Services
Machine Learning
Client SOA
Algorithms Mobile Services
Communication Interface
Middleware
User’s Premises
KNX Gateway Bluetooth Gateway
…
Sensor/Service
Descriptor
Figure 5: Software middleware architecture with sensors and health services.
The former is able capture spatial information of the feature maps produced by input data.
The latter, in the form of attention-based Long Short-Term Memory (LSTM) cells, is able to
apprehend temporal aspects of the data streams. Of our interest is the Multi-Scale Convolutional
Recurrent Encoder Decoder (MSCRED) [22], an attention-based ConvLSTM neural network
developed to capture both temporal and cross-channel anomalies in Multivariate Time Series
(MVTS). The goal is to capture MVTS space-time features and detect temporal and cross-channel
anomalies on the basis of reconstruction errors, or loss, obtained as difference between the
original multi-scale signature matrices and the reconstructed ones. In this context the MSCRED
is applied to provide a feasible solution for the detection of small changes in human habits
and, in full, worsening in patients’ health status to promptly detect situations of heart failure
onset. Intuitively, the network captures spatio-temporal information from inter-correlations
of time series. The data dimensionality is guaranteed by the use of different window sizes,
packed three-by-three, that overlap to each other. The network performs well on harmonic
series and data sets made up of chronologically ordered sensor information, e.g., belonging
to a power plant. But there is no evidence of its proper functioning in the domain of health
care for older adults. The implementation of the framework requires some steps. Firstly, the
processing of health monitoring data to obtain correlation matrices from such weights to feed
the network. Secondly, the encoding of spatio-temporal information and construction of feature
maps through a convolutional encoder and an attention-based recurrent neural network with
LSTM cells. Thirdly, the use of a convolutional decoder to reconstruct the correlation matrices
and a square loss function to perform the end-to-end learning. Lastly, the computation of the
anomaly score as residual matrix obtained as the difference between the original matrix and
the reconstructed one. In case of missing data (due to, for instance, signal interference) we are
considering two possible solutions for real time processing: either remove the entire temporal
slot in which data are incomplete, or temporarily reduce the dataset dimensionality by dropping
the faulty channel and reconsidering the thresholds (if any) accordingly. Differently, for offline
statistical analysis we are considering to replace the missing values by applying techniques like
median, interpolation, last/next observation carried forward/backward and so forth. We do not
go into details of the MSCRED functioning, and we refer interested parties to [22].
4.1. Anomaly Detection in Multivariate Time Series
Briefly, the detection of anomalies in MVTS refers to the problem of identifying, within a
data stream of several chronologically ordered channels, rare and distinct schemes (or isolated
observations) that deviate from their normal trends. Such schemes are often labeled as outliers, to
point out their irregular behavior [23, 24]. Fundamentally, there are two main kind of anomalies
that can occur into a MVTS: temporal anomaly, which occurs within each single channel
when it deviates from its regular behavior (e.g., the increase in body temperature due to a
seasonal flu); cross-channel anomaly, which occurs as correlation between different channels
that present irregularities on their joint historical behaviors. Of the two, the cross-channel
anomaly is the more difficult to detect because of the common regular trend manifest by each
channel individually taken. To date, many MVTS anomaly detection techniques have been
investigated for different research areas. It is worth remembering that among the plethora of
MVTS anomaly detection algorithms developed so far, the main method families span from
classic ML techniques to Deep Learning [25, 26], from stochastic learning to statistical regression
[27, 28], and so forth. The issue of detecting the worsening of the patient’s health status to
prevent the onset of heart failure can be faced as an anomaly detection problem. The change in
health status may depend on a single factor or cross factors, and often this is due to imperceptible
changes in lifestyle habits. That is why we need a predictive system able to detect both types of
anomalies. In the next subsection we introduce some preliminary results obtained by applying
an unsupervised learning approach to a synthetic data set.
4.2. Preliminary results
To test the functioning of the selected ML model over physiological data, waiting to obtain real
environmental data as the project evolves, we run MSCRED on a synthetic dataset composed of
time series reproducing sinusoidal signals typical of health monitoring sensors. Generally, syn-
thetic time series data are used to both augment sparse data set information and test predictive
algorithms in absence of real world information [29]. The model has been trained to learn data
with no anomalies, while the test set has intentionally been enriched with perturbations in order
to test the inability of the neural network to properly reconstruct the input signature matrix in
their presence. Precisely, the data set is made up of one-hundred channels (i.e., sensors), each
one with a different frequency, and several time-steps. Temporal patterns of the synthetic MVTS
data are modeled as trigonometric functions, randomly selected for each signal. To generate the
anomalies a little perturbation is added to the generated signals. The conditional expression, as
Figure 6: Sample fragment of the perturbed data set obtained by plotting synthetic health monitoring
information. Anomalies have been highlighted with black spots.
introduced in [22], is as follows:
{︃
𝑠𝑖𝑛[𝑡 − 𝑡0 /𝑤] + 𝜆 · 𝜖, if 𝑘 = 0
𝑇 𝑆𝑖
𝑐𝑜𝑠[𝑡 − 𝑡0 /𝑤] + 𝜆 · 𝜖 otherwise
where 𝑇 𝑆𝑖 stands for the i-th time series channel, 𝑘 is a 0/1 token randomly selected, 𝑠𝑖𝑛 and
𝑐𝑜𝑠 [𝑡 − 𝑡0 /𝑤] are trigonometric functions applied to the ratio between time delay 𝑡0 ∈ [50, 100]
and frequency 𝑤 ∈ [40, 50] to reproduce periodic cycles, and the dot product in tail is a scaled
random Gaussian noise to simulate perturbation within the time series. Figure 6 shows a
fragment of the test set with five perturbations (highlighted by black spots) as part of the
synthetic health monitoring data. To train the model, also on the basis of better empirical
results, we chose the following setting: gap time 1, window size [10, 30, 60], max step 5, batch
size 32 and learning rate 0.0002.
The residual matrix computed over test set data identifies 37 broken elements, obtained as
difference between the original correlation matrix and the reconstructed one. Such value high-
lights the network failure to correctly reconstruct the input matrix in presence of perturbations
and defines the anomaly score. In Fig. 7 are shown the anomalies identified by MSCRED as
broken elements of the residual matrix. The test corroborate the goodness of the candidate
neural network in detecting anomalies (i.e., potential variation in the patient’s health status)
over synthetic health data. Ultimately, it is worth noting that such a technique, being based
on the analysis of multidimensional time series information, can be applied to other domains
than the current one. Because of these results, we expect the same (or quite similar) behavior
of the algorithm over real data. The effort is to select the appropriate information capable of
obtaining and clearly interpret the internal sectors of the residual matrix.
Figure 7: Broken elements of the reconstructed residual matrix as anomaly score.
5. Conclusions
In this work we have introduced a summary description of the elements and mid-term results
of the project ChAALenge. In particular, we focused on sensor technology for collecting patient
health data, the interfacing middleware between sensors and actuators, and potential artificial
intelligence algorithms for health status prediction. Preliminary results based on a synthetic
data set were also presented to test one of the most promising ML algorithm, i.e., MSCRED.
Such results corroborated the functioning of the predictive model by identifying anomalies and,
in full, potential aggravations in the patients’ health status. Regardless the considered scenario,
being it a house where a single patient lives, or being it a nursing home where multiple patients
require cures, the algorithm should be able to provide ad hoc predictions. To do so, we are
considering an implementation able to train multiple copies of it over each single patient’s
personal data. With respect to the economic sustainability of the equipment required to pursue
the project goal, we are taking into account non-invasive solutions that enable the reuse of
both allocated devices and sensors. Currently, the project ChAALenge is halfway and still work
in progress. To properly evaluate the proposed solution, it is necessary to collect real-world
data to make inference in field-test. Pursued this goal, we will be able to corroborate the real
predictive skill of the model. Otherwise, there will be made architectural adjustments, also
according to the empirical results achieved for improving the efficacy, the effectiveness, and the
predictive capabilities of our system.
Acknowledgement
The project ChAALenge is funded by the Ministry of Economic Development of Italy via the
Sustainable Growth Fund. The latter aims at implementing the national specialization strategy
through the granting and disbursement in favor of research and development projects in the
area of Life Sciences.
References
[1] B. Rechel, E. Grundy, J.-M. Robine, J. Cylus, J. P. Mackenbach, C. Knai, M. McKee, Ageing
in the european union, The Lancet 381 (2013) 1312–1322.
[2] A. Groenewegen, F. H. Rutten, A. Mosterd, A. W. Hoes, Epidemiology of heart failure,
European journal of heart failure 22 (2020) 1342–1356.
[3] K. K. Ho, J. L. Pinsky, W. B. Kannel, D. Levy, The epidemiology of heart failure: the
framingham study, Journal of the American College of Cardiology 22 (1993) A6–A13.
[4] L. H. Curtis, D. J. Whellan, B. G. Hammill, A. F. Hernandez, K. J. Anstrom, A. M. Shea,
K. A. Schulman, Incidence and prevalence of heart failure in elderly persons, 1994-2003,
Archives of internal medicine 168 (2008) 418–424.
[5] N. Townsend, D. Kazakiewicz, F. Lucy Wright, A. Timmis, R. Huculeci, A. Torbica, C. P.
Gale, S. Achenbach, F. Weidinger, P. Vardas, Epidemiology of cardiovascular disease in
europe, Nature Reviews Cardiology 19 (2022) 133–143.
[6] G. Corrao, A. Ghirardi, B. Ibrahim, L. Merlino, A. P. Maggioni, Burden of new hospitaliza-
tion for heart failure: a population-based investigation from italy, European journal of
heart failure 16 (2014) 729–736.
[7] P. A. Heidenreich, B. Bozkurt, D. Aguilar, L. A. Allen, J. J. Byun, M. M. Colvin, A. Deswal,
M. H. Drazner, S. M. Dunlay, L. R. Evers, et al., 2022 aha/acc/hfsa guideline for the
management of heart failure: a report of the american college of cardiology/american
heart association joint committee on clinical practice guidelines, Journal of the American
College of Cardiology 79 (2022) e263–e421.
[8] G. Borrelli, M. Pistoia, P. Casacci, A. Leone, P. Siciliano, M. d. Tommaso, E. Vecchio,
M. Delussi, A. Greco, D. Sancarlo, et al., Care@ home: methodology, goals and project
experimentation activities, in: Ambient Assisted Living, Springer, 2015, pp. 307–316.
[9] F. Palumbo, A. Crivello, F. Furfari, M. Girolami, A. Mastropietro, G. Manferdelli, C. Röcke,
S. Guye, A. Salvá Casanovas, M. Caon, et al., “hi this is nestore, your personal assistant”:
design of an integrated iot system for a personalized coach for healthy aging, Frontiers in
Digital Health 2 (2020) 545949.
[10] E. I. Konstantinidis, D. Petsani, G. Conti, A. Billis, V. Conotter, G. Chican, T. L. Lorenzo,
A. R. Rodríguez, S. H. Fraile, A. Beristain, et al., A new approach for ageing at home: The
captain system, Studies in Health Technology and Informatics (2019) 1704–1705.
[11] R. op den Akker, T. Beinema, O. Banos, D. Heylen, B. Bedsted, A. Pease, C. Pelachaud, V. T.
Salcedo, S. Kyriazakos, H. Hermens, Council of coaches a novel holistic behavior change
coaching approach, in: 4th International Conference on Information and Communication
Technologies for Ageing Well and e-Health, ICT4AWE 2018, SciTePress, 2018, pp. 219–226.
[12] I. Kouris, M. Sarafidis, T. Androutsou, D. Koutsouris, Holobalance: an augmented reality
virtual trainer solution forbalance training and fall prevention, in: 2018 40th Annual
International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC),
IEEE, 2018, pp. 4233–4236.
[13] M. Weiser, The computer for the 21 st century, Scientific american 265 (1991) 94–105.
[14] A. Manni, A. Caroppo, A. Leone, P. Siciliano, Video-based contactless mood detection
combining heart rate and facial expressions, in: AISEM Annual Conference on Sensors
and Microsystems, Springer, 2022, pp. 307–313.
[15] A. Leone, G. Diraco, P. Siciliano, Detecting falls with 3d range camera in ambient assisted
living applications: A preliminary study, Medical engineering & physics 33 (2011) 770–781.
[16] S. Sharma, Z. Huang, M. Rogers, M. Boutelle, A. E. Cass, Evaluation of a minimally invasive
glucose biosensor for continuous tissue monitoring, Analytical and bioanalytical chemistry
408 (2016) 8427–8435.
[17] F. Palumbo, D. La Rosa, E. Ferro, D. Bacciu, C. Gallicchio, A. Micheli, S. Chessa, F. Vozzi,
O. Parodi, Reliability and human factors in ambient assisted living environments, Journal
of Reliable Intelligent Environments 3 (2017) 139–157.
[18] V. Miori, D. Russo, C. Concordia, Meeting people’s needs in a fully interoperable domotic
environment, Sensors 12 (2012) 6802–6824.
[19] F. Palumbo, D. La Rosa, S. Chessa, Gp-m: Mobile middleware infrastructure for ambient
assisted living, in: 2014 IEEE Symposium on Computers and Communications (ISCC),
IEEE, 2014, pp. 1–6.
[20] G. Barbon, M. Margolis, F. Palumbo, F. Raimondi, N. Weldin, Taking arduino to the internet
of things: The asip programming model, Computer Communications 89 (2016) 128–140.
[21] A. Crivello, F. Palumbo, P. Barsocchi, D. L. Rosa, F. Scarselli, M. Bianchini, Understanding
human sleep behaviour by machine learning, in: Cognitive infocommunications, theory
and applications, Springer, 2019, pp. 227–252.
[22] C. e. a. Zhang, A deep neural network for unsupervised anomaly detection and diagnosis
in multivariate time series data, in: Proceedings of the AAAI conference on artificial
intelligence, volume 33, 2019, pp. 1409–1416.
[23] V. Chandola, A. Banerjee, V. Kumar, Anomaly detection: A survey, ACM computing
surveys (CSUR) 41 (2009) 1–58.
[24] R. Chalapathy, S. Chawla, Deep learning for anomaly detection: A survey, CoRR
abs/1901.03407 (2019).
[25] C.-L. Liu, W.-H. Hsaio, Y.-C. Tu, Time series classification with multivariate convolutional
neural network, IEEE Transactions on Industrial Electronics 66 (2018) 4788–4797.
[26] H. Zhao, Y. Wang, J. Duan, C. Huang, D. Cao, Y. Tong, B. Xu, J. Bai, J. Tong, Q. Zhang,
Multivariate time-series anomaly detection via graph attention network, in: 2020 IEEE
International Conference on Data Mining (ICDM), IEEE, 2020, pp. 841–850.
[27] J. Li, W. Pedrycz, I. Jamal, Multivariate time series anomaly detection: A framework of
hidden markov models, Applied Soft Computing 60 (2017) 229–240.
[28] P. J. Rousseeuw, M. Hubert, Anomaly detection by robust statistics, Wiley Interdisciplinary
Reviews: Data Mining and Knowledge Discovery 8 (2018) e1236.
[29] G. Forestier, F. Petitjean, H. A. Dau, G. I. Webb, E. Keogh, Generating synthetic time
series to augment sparse datasets, in: 2017 IEEE International Conference on Data Mining
(ICDM), 2017, pp. 865–870. doi:10.1109/ICDM.2017.106.