Architecture of a Non-Intrusive IoT System for Frailty Detection in Older People

Abstract

Early detection of frailty is one of the main challenges in the world we live in. Being aware of physiologic and behavior changes can predict and prevent the onset of mental complications for older people. By using modern technologies, one can get insights, which may help detection of different pathologies. In this paper, preliminary results for a novel system to detect early signs of frailty are presented. A prototype was developed and tested in laboratory conditions after requirements and functional capabilities were defined. The main advantages of the proposed architecture are the usage of commercial off-the-shelf (COTS) components and custom mechanisms of security to assure a high level of confidentiality and integrity of user-specific data. Other original elements are its easy-to-use and non-intrusive characteristics.

1. Introduction

The progress of science and technology has impacted life expectancy, leading to significant increases in lifespan. Today we are facing new challenges as the elderly population is increasing, which leads to the necessity for finding new ways to improve their quality of life. The trend is switching to a preventive style of monitoring health, based on processing and extracting information related to various existing or future diseases.

Data collection and processing to obtain relevant information about people’s health status has become a topic of interest to the scientific community in recent years. The concept of Internet of Things (IoT) has gained a lot of attention considering its advantages: pervasiveness, interoperability, easy access and affordability. There are many published papers related to IoT-based systems that are real-time health monitoring systems that help people to improve their quality of life, provide them technologies and services that support their daily activities and help them to live longer and remain independent at home. Wu et al. demonstrated a design for a wearable patch sensor. The sensor measures many physiological metrics, such as photoplethysmography and blood pressure [1]. Akkas et al. proposed a WBAN-based architecture for biomedical applications. The system uses patient data, such as the oxygen ratio, pulse rate and plethysmogram [2]. Ni et al. introduced a wearable device with low power consumption to detect the vibration of heart signals. This device helps the medical apparatus in healthcare institutions monitor critical patients, aiding in rapid intervention and saving lives [3] Asghari et al. proposed a monitoring system for patients in IoT-based medical environments. The monitoring scheme is used to gather predetermined data about the patients and predicts their future diseases [4] Chui et al. proposed a monitoring system for patient behavior considering big data obstacles. In addition, they discussed the challenges that are related to the security, privacy and interoperability in the healthcare research field [5].

The inclusion of sensors used in the medical field in consumer-friendly devices such as mobile phones and smart watches provides easy access to data collection. Qingguo Chena have shown [6] that the existence of some diseases of the respiratory system can be indicated using the level of oxygenation of the blood. In a similar fashion, the use of the ECG can predict the appearance of some cardiovascular problems [7].

One of the most important insights about overall health of a person is frailty, defined as a geriatric condition characterized by deteriorations in physical and mental processes of a person, being considered a syndrome with impact to unfavorable health outcomes [8]. Fried et al. developed a frailty phenotype theory with five criteria [9], which can be measured by using smart systems.

The reversible nature of frailty and recent evidence of the efficiency of intervention strongly argue for frailty detection. This study is a part of a European Eureka project (Project URL: https://cinnamon-project.eu/en/home-page/, accessed on 30 March 2023) where the term frailty will be used in the broader meaning of the concept, encompassing physical frailty, cognitive frailty and social withdrawal. This project aims at developing a secure, cost-effective, unobtrusive and ethically acceptable device, in which one of the objectives is to allow early detection of risks associated with ageing to enable earlier intervention to ameliorate their negative consequences.

Given the diversity and availability of commercial off-the-shelf (COTS) wearables and devices, we decided that usage of smart watches has benefits in terms of easy usage and costs. The cost for a wearable starts from 20 euros, depending on its performance and features. The proposed smartwatch’s price was 120 euros, with a discounted price in 2023 starting at 70 euros. Wearable devices are used to keep up with health and fitness status. Such a system is an electronic wearable device that uses sensors to monitor health-related metrics such as orientation, movement, rotation, heart rate and blood oxygen level. These devices collect data and convert it into steps, calories, sleep quality and general activity that a person performs throughout the day. The use of these devices has the following advantages: ease of use, non-invasive nature in a person’s life, low costs, availability and small size.

This paper presents the design and development stages of an IoT system, a novel approach in terms of communications and security. The system is projected to detect early signs of frailty in elderly people. After a literature review, it was decided that the best approach for the project’s purpose was to use an accelerometer as main sensor and compute accelerations to detect signs/patterns related to frailty. In this paper the novel characteristics of the architecture are emphasized, while the testing of the system in real case scenarios, for detecting frailty will be subject of a future paper.

The Technology Readiness Level (TRL) for our current research is 4, i.e., technology validated in a laboratory environment. In future work, we are aiming to reach TRL 6, which means demonstrating functionalities of our system in relevant usage conditions (industrial usage), to prove that our hypothesis is true. This paper is structured as described below. In Section 2, System Design, we present previous results for the proposed system, its general topology, followed by a description of results from the design process. The proposed architecture is presented with a description of each module. Section 3, Results, describes the system development and the testing phase of initial prototype in laboratory conditions. Discussion addresses the conclusions from the experiment and further work.

2. System Design

To implement our IoT system, a questionnaire about user requirements was developed and conducted to determine the functional and non-functional requirements [10]. The role of this questionnaire was to fill the gaps in knowledge of end-user input into such a system and to have a broader view of the end-user needs and concerns. The results showed that security, easy-to-use and non-intrusive aspects of this system architecture are the basic properties needed for a new generation of frailty detection and monitoring devices. Data privacy and confidentiality were also identified as important non-functional requirements.

For the design process of this system, we conducted a comparison between specifications of some of the most popular smartwatches on the market, to find suitable components for our proposal. Their features are described in terms of collected data, battery life and sensors included, and presented in

Table 1. Comparison of popular smart bracelets.

Manufacturer	Device	Battery Life	Sensors	Main Processed Data
Fitbit	Versa	up to 7 days	3-axis accelerometer, optical heart rate monitor, 3-axis gyroscope, ambient light sensor, vibration, orientation sensor.	Activity and exercises, heart rate, steps, distance, floor, calories, sleep, profile (e.g., weight, height).
Apple watch	Series 6	18 h	Blood oxygen sensor, electrical heart sensor, GPS, accelerometer, gyroscope, ambient light sensor, altimeter	Steps, calories, sports (including underwater), heart rate, ECG, SPO2, time asleep, fall detection
Garmin	Forerunner 745	up to 7 days	GPS, altimeter, gyroscope, accelerometer, pulse OX acclimation, heart rate sensor.	Steps, calories, sports (including underwater), heart rate, personal running maps, SPO2, fall detection, recovery.
Samsung	Galaxy 3 watch	24–36 h	Accelerometer, gyroscope, barometer, electrical heart rate, optical heart rate, ambient light	Steps, calories, sports, heart rate, ECG, SPO2, VO2 Max, sleep stages, fall detection
Polar	Ignite	up to 5 days	Accelerometer, heart rate monitor, compass, gyroscope, ambient light	Sleep tracking, steps, heart rate, calories, distance, stress, VO2 Max. swim tracking

. Security and privacy vulnerabilities, specific to smart watches, are a constant pursuit in cybersecurity domain. In our study design phase, we have considered papers related to monitoring applications, security in IoT devices and hardware design. We did a comparison between some of the most popular fitness tracker, and we found that in comparison with other vendors, Fitbit employs the most effective security mechanism in their products.

Related to security, after an extended analysis, Cyr et al. showed that the Fitbit environment provides a reasonable level of privacy [11]. Yoon et al. also conducted a study where they checked what data could be retrieved in the Android environment from the Fitbit app and tracker, concluding that the exposed data is not enough be considered as a significant issue [12].

A topology used for data collection involves sensors, processing units and communication protocols. The raw data are captured by the electrical sensor as output voltages and are transmitted to the data acquisition unit (DAU). Depending on the resolution of the sensor, the DAU applies a physical filter to improve the noise performance to raw data and transmits it to the data processing unit (DPU), which is an assembly capable of preprocessing and storing information, in the form of a data structure. In most implementations, the accelerometer and data acquisition unit are embedded in a single device, which is connected to the data processing unit via a direct data bus.

In Figure 1 the high-level design of our system is illustrated, including the direct connection to the Internet. In this architecture, the smart bracelet obtains the accelerometer data and preprocesses them, to improve overall data accuracy and reliability. Data are transmitted using a short-range communication protocol (e.g., Bluetooth, ZigBee, Wi-Fi) to a smartphone, which acts like a communication gateway, as it packs the data and transmits it using a two-way Internet connection to a cloud environment. In the cloud, data are stored and processed for obtaining relevant information, which is transmitted back to the mobile phone in the form of an insight, for example, the number of steps over a period.

2.1. A Novel Architecture for Frailty Data Collection Using a Fitbit Smartwatch

The non-functional requirements gathered from the conducted questionnaire identified a number of necessary properties, which required new components to be added in our architecture. Firstly, non-intrusiveness required our system to be easily integrated into daily activities. Secondly, our system must be easy to use; taking into account the age of subjects, we had to choose a low-cost device with enough capabilities to allow us to record relevant data for frailty detection.

According to the aim and objectives emphasized in Section 1, we designed a system with a Fitbit Versa smartwatch in the Fitbit environment, using the official Fitbit Application Programable Interface (API). This allows third-party applications to record data from smartwatch sensors and use it for different purposes. In addition to an accelerometer, we were interested to collect data from a gyroscope, orientation sensor and heart rate sensor.

The general architecture consists of two components, each component using different service layers and modules.

• Fitbit environment (FE), presented in Figure 2, consists of:
  • the Fitbit Versa smartwatch
  • the Fitbit cloud and Fitbit API
  • a companion application, included in the Fitbit app on an Android Smartphone
• Data cloud server (DCS), presented in Figure 3, consists of:
  • a private virtual server (VPS) configured to serve as a NodeJS webserver, including the data processor (DP) and data storage (DS) modules
  • a MongoDB server instance, with bidirectional communications with DS.

Relations between the cloud-based data collector (Figure 1) and our novel architecture structure (Figure 2 and Figure 3) are as follows: the smartwatch with accelerometer and smartphone, as the communication gateway, are represented by the Fitbit environment (FE), while the cloud is the DCS.

The front-end interfaces in architecture were implemented in the Fitbit environment (as a Fitbit clock face application running continuously) and in the DCS (as a web platform with visual components).

2.2. Components and Modules

2.2.1. Fitbit Watch Application

An application was developed for the smartwatch that allows the collection of raw data directly from embedded sensors. To optimize the entire process, we created an application which is available in the Fitbit Gallery Application and can be downloaded directly from the official ecosystem (URL: https://gallery.fitbit.com/details/711fea15-b2f3-4792-a8f6-01b6c5339dd4, accessed on 29 March 2023). The application is hardware-platform-agnostic, which means it can be used on all Fitbit smartwatch models.

In our design, raw data are collected from the accelerometer along three orthogonal axes: x, y and z. The x-axis is parallel to the device screen, aligned with the top and bottom edges in left-right direction. The y-axis is parallel to the screen of the device aligned with the left and right edges in the up-down direction and the z-axis is perpendicular to the device screen facing up.

Further notable data collected by our system are from the gyroscope and orientation sensors. To monitor the heart rate, Fitbit Versa uses a photoplethysmography technology sensor.

2.2.2. Companion Application

The companion application is a native JavaScript application developed in the Fitbit environment using the official API. The companion acts like a shadow process, which is always waiting for instructions from other system components. Its main functions are:

• to establish the communication link between the smartwatch, Fitbit cloud and DCS, to transmit the data captured for storage and processing
• to act as the main node in the communications topology; it opens a bidirectional socket with the Versa using native Fitbit methods, as well as a secure communications socket with DCS, through the NodeJS server
• to connect to the Fitbit cloud for retrieving relevant user information (name, age, weight, userID), by using OAuth 2 tokens for authentication.

The companion application acts like a child process inside the Fitbit application. It is installed in the same package with Fitbit smartwatch application.

2.2.3. Fitbit Cloud

The role of the Fitbit cloud is to store information about the user of the application, as well as the data recorded by the bracelet, in a preprocessed and standardized format. Because we are using raw data, directly from sensors, data stored in the cloud are later used for comparative purposes, to validate the results obtained from processing in our system.

2.2.4. Webserver—Front-End Web Page Interface

The main user interface is presented in Figure 4, which shows different controls used for data recording. By selecting any activity, a recording slot session is opened. Using a NodeJS engine, it creates an asynchronous instance for each user connected at the system, so their activities in the application are separated.

For each session, authentication is done using user-specific credentials, by calling the companion application. After logging into the Fitbit cloud, the userID is extracted and the communications socket with the smartwatch is opened via the companion.

2.2.5. Data Processor (DP) and Data Storage (DS)

The role of the DP is to retrieve and type in activity-specific data, along with anthropology information (age, gender, height, and weight). It traverses the recorded data and marks with a specific flag (1/0) if one of the components of the data series is null. After processing the data, the DP calls the DS module to transmit the information to the database.

The DS component is intended to establish a secure connection to the MongoDB non-relational database instance. It assigns a transmission session number and performs the encrypted transmission.

2.3. Constructive Components of Architecture

2.3.1. Communication

From a software point of view, the main component in our system architecture is the NodeJS server, hosted on a VPS (virtual private server)—that includes DP and DS modules in Figure 3. It communicates with both the smartwatch and the MongoDB database. For multiple users, parallel connections are required to transmit data streams. For that purpose, the server is in a continuous state of waiting and opens a WebSocket each time it receives a valid request. Communications are fully secured with a Let’s Encrypt SSL certificate, generated by using OpenSSL tool.

To enable message exchanges between our smartwatch application and the companion in the Fitbit application, we use the Fitbit messaging API. The implementation is identical on the smartwatch and companion and is required to successfully open a Message Socket connection before sending any message.

For data transfer between companion and server, we used secure a WebSocket, which is a continuous connection between a client and a server. It provides a two-way, full-duplex connection that works over https through a single TCP connection. Among the reasons behind the decision to use WebSocket instead of the http/https protocol were latency and WebSocket header size.

For a user to open an instance with the server and communicate on a separate, secure channel, an authentication with the Fitbit client account will be performed, both in the companion and in the Cinnamon App for extracting the client account that will be used to establish a single WebSocket connection. It that way, we are leveraging the security level of entire system.

2.3.2. Authentication

For user authorization, Fitbit uses OAuth 2.0, which requires the developed application to obtain an access token for the user to access their data. This access token is used to make HTTPS requests to the Fitbit API. Applications must only request permission for the resources they intend to access or modify. OAuth 2.0 refers to these permissions as domains. All Fitbit API endpoints require one or more domains, which are listed in each endpoint documentation. The application must specify a list of domains when redirecting the user to the authorization page. The issued access token will contain only the domains required by the application.

2.3.3. Database Collections

The database used in the system is MongoDB, a non-relational, document-oriented NoSQL database. The use of this type of database is motivated by the need to store JSON files, having as a system for organizing collections and indexes. Each user is assigned a specific collection, defined by userID. Another important reason for choosing MongoDB is the ability to quickly access information, considering the amount of data that is transmitted from the smartwatch to the server.

3. Results

After the requirements for system were defined, we wanted to check the feasibility of our system design. It was deployed it in a TRL 4 small-scale prototype, by developing each previously described component. It was evaluated and its performances checked in terms of multi-usage (if the system can work with multiple simultaneously users), data storage (how much disk space is required) and power consumption. By successfully building the prototype based on components presented in Section 2.3.1., Communication and Section 2.3.2., Authentication, new layers of security assuring data confidentiality were added: encrypted communications with servers using SSL certificates, custom and secured WebSockets and authentication using OAuth 2.0.

To determine early signs of frailty, we selected five mature similar projects as starting points for comparison. Most projects have wearable devices for main data acquisition and data processing units, with a focus on determining parameters related to gait, weight, locomotion and sleep. Similar systems are using a Fried frailty criteria approach, while others are using Groningen Frailty Indicator or have a proprietary frailty detection standard. A full comparison of five similar systems is presented in

Table 2. Comparison of popular frailty detection/monitoring systems.

Article Title	Year	Measurement Standard	Main Component	Parameters Used
EFurniture for home-based frailty detection using artificial neural networks and wireless sensors [13]	2011	Custom, defined by authors.	ICT system with multiple devices (smart pad, scale, pressure sensor, ultrasonic sensor)	weight loss, exhaustion, balance, slowness, weakness, reaction time, functional reach
Clinical frailty syndrome assessment using inertial sensors embedded in smartphones [14]	2015	Fried frailty criteria	Smartphone in a small neoprene sleeve	peak positive and peak negative acceleration values, means and standard deviations (SD) of accelerations x, y and z
Motor Performance and Physical Activity as Predictors of Prospective Falls in Community-Dwelling Older Adults by Frailty Level: Application of Wearable Technology [15]	2016	Fried frailty criteria	Wearable sensors (inertial sensors/accelerometer)	balance, gait speed, number of steps, stride length, stride time, posture duration (sitting, walking), locomotion outcomes
Validity of an accelerometer-based activity monitor system for measuring physical activity in frail older adults [16]	2016	Groningen Frailty Indicator (GFI)	A custom wearable device (DynaPort MoveMonitor), which consist of an accelerometer and an analysis module	accelerations for x, y and z directions of accelerometer (locomotion, sitting, standing and lying)
Wearable sensors and the assessment of frailty among vulnerable older adults: An observational cohort study [17]	2018	Fried frailty criteria	Pendant sensor (PamSys) with tridimensional accelerations	accelerations values to determine steps, periods of moving, sleep parameters
Proposed architecture in this paper	2022	Fried frailty criteria	A smartwatch with accelerometer, orientation sensor, gyroscope and heart rate sensor	accelerations, directions of orientation sensor, gyroscope’s axes and heart rate

.

We adapted our application to send data continuously. The packets recorded in the database contain the parameters from Figure 5, which represent an example of a record. Laboratory experimention was conducted in three data recording sessions, each one for a period of approximately 15 min. For each session, four non-frail people, each one with their own smartwatch, started recording data at the same time. The structure of a data packet is the following: four variables representing the directions of orientation sensor, three variables representing the accelerations x, y and z of accelerometer, three variables representing the gyroscope’s axes, one variable for heart rate and three variables for data packet identification (timestamp, session ID and user ID).

The server load is measured by a script which interrogates Ubuntu top command for server usage for each second and records it, while smartwatches continuously send data. The CPU usage and RAM loading were recorded, and a median value was calculated. In

Table 3. Results of three recording sessions.

Measured Insight	Session #1	Session #2	Session #3
Average CPU usage|system (%)	2.2%	2.1%	2.1%
Average CPU usage|user (%)	1.7%	1.9%	2%
Average RAM usage (MB consumed)	12 of 312 total	11 of 316 total	17 of 331 total
Total packets	3866	3647	4011
Packets size	292 kb	277 kb	302 kb

we can observe that our designed system is predictive and reliable, having similar averages for different recording sessions. During the testing phase, we did not record any functional errors.

4. Discussion

An IoT system with the ability of processing and storing data for elderly, to extract insights for early signs of frailty, was successfully designed, developed and tested in laboratory conditions. The design and functionalities of such a system are based on functional requirements determined in an early stage of our project. The initial requirements were satisfied: a system that is easy to use and non-intrusive and can assure data integrity and confidentiality.

Table 4. Comparison with other frailty detection systems.

Project	Sensor Location	Easy to Use (System Complexity)	Security	Non-Intrusive
[13]	Sensors are placed in different home appliances: a lamp, a chair and a carpet.	Hard to use: consists of multiple devices, with complex actions to trigger	Not addressed	Yes
[14]	Middle third of the sternum	Hard to use: system fixed with non-elastic tape around the trunk	Not addressed	No
[15]	Inertial sensors: 2 at shins above ankles, 2 at thighs above knees and 1 attached to lower back; 1 triaxial accelerometer inserted in a shirt	Hard to use: five sensors fixed directly on body, accelerometer manually inserted in a shirt	Not addressed	No
[16]	At the lower back, between and above the posterior superior iliac spine fixed with an elastic strap	Easy to use: attached on the body with a strap	Not addressed	No
[17]	Sternum	Easy to used: attached at neck	Not addressed	No
Proposed architecture in this paper	Wrist	Easy to use: attached at wrist	Addressed by standard and custom mechanisms.	Yes

illustrates the main characteristics of similar systems, from an initial requirements point of view. Compared to those systems and based on requirements identified, our system was projected to have custom security layers, along with traditional security measures. We are using a smartwatch, with an application which works as a regular clock, which is non-intrusive for a regular user. In our system, the sensor location is at the wrist. The downloadable and easy-to-install application installed in an accessible environment (Fitbit) ensures its ease of use. The clock face of our application is presented in Figure 6.

The proposed architecture has direct link communications between components, providing a peer-to-peer communication via secured channels. The usage of SSL certificate and tokens for authentication adds extra layers of security to prevent data to be accessed illegally. Custom WebSockets for each data transmission allows users to create short time communications channels, which are hard to be intercepted by third-party malicious applications for two reasons: duration of channel availability and encryption used to transfer data. Other relevant components from a security perspective are usage of OAuth 2.0 protocol, the separation of data by storing them independently in the database collection with no personal data recorded and the unique way of authenticating each user in our system: by using an ID which is claimed from a third party.

As it can be seen from the table, our system, as compared with other similar systems, have all the sensors placed into a bracelet which is easy to use and non-intrusive and addresses security from a design phase.

5. Conclusions

Development of IoT systems starting with requirements from end-users is challenging, having multiple constraints. We consider that is feasible to design and develop an IoT system to detect frailty with a COTS smartwatch as the main component. Our paper proposes an architecture of a system, which is easy-to-use and accessible. The costs for end-users are determined by prices of a smartwatch (starting from 20 euros, 120 euro in our developed system) and a smartphone (with a starting price of 100 euros). This architecture uses low resources and can be adapted to record and process massive amounts of data. The modularity of the applications (i.e., the software components are installable and easy to deploy), the simplicity of usage in the proposed system and the original characteristics emphasized in Section 2 and Section 3 highlight the strengths of our proposed architecture as a starting point for a frailty detection system.

Our future work will leverage the functions of our system by adding a machine-learning layer able to detect patterns for frail people, to find relevant and replicable frailty insights.