1. Introduction
The advent of smart devices and rapidly evolving communication technologies, has enabled the formation of the Internet of Things (IoT) environment. The IoT paradigm intends to connect and exchange information and user data between devices, physical environment and the individual. This translates into a smart, connected and interactive environment for an individual, thereby improving the quality of life. The devices could be computers, phones, wearables, home appliances, infrastructure and vehicles [
1,
2,
3]. Therefore, any device which operates even with an ON/OFF switch can be integrated into an IoT environment. The IoT environment also allows for connecting devices with limited memory, power and CPU.
Figure 1 shows how different components and users are interconnected in an IoT paradigm [
1,
4].
Advancements in sensor design have also enabled the rapid evolution of smart devices for personalized applications which include communication, health and fitness monitoring, virtual environments, autonomous transportation and smart homes. Considering the aspect of connected healthcare, the development of telehealth systems has resulted in coining of the term IoMT (Internet of Medical Things), which is a subset of IoT. The IoMT environment focuses on delivering clinical services to an individual via connected devices such as smart phones, wearables and infrastructure (see
Figure 2). These services include [
5]:
Remote health monitoring via telecommunication network.
Use of mobile health monitoring equipment and applications.
Doctor-patient consultation via interactive technology.
Continuous monitoring using smart devices for elderly and critical care individuals.
Our study is based on the use of wearables for home-based health monitoring in an IoMT environment. Wearables are devices embedded with accelerometers, gryoscopes, light and pressure sensors, for capturing and analyzing streaming physiological data from an individual during daily activity. Unlike smart phones or tablets, these devices can be comfortably worn on different body regions throughout the day, and can be used for various applications such as fitness monitoring, behavior tracking and vital signs analysis for critical disorders such as stroke, falls or seizures [
6].
From our prior survey [
6], we found that many currently available wearables such as Apple Watch
TM and FitBit
TM have embedded sensors for collecting and analyzing basic human activity parameters such as step counts, pulse rate, temperature and sleep times for fitness awareness. We also investigated into their respective SDKs (software development kits), which described how physiological data is collected, analyzed and shared with service providers for decision generation. In recent times, many clinical studies have been conducted to explore the validity of using wearables for physiological data analysis for disease or disorder detection. For example, accelerometer-based wearables have been used to study daily activity monitoring in individuals suffering from neuromuscular disorders, and validate their outputs with clinical standards [
7].
As per a survey [
8], considering that only about 90 out of 600 currently available wearables are being used for medical applications, we can see a clear potential for their usage in long-term, home-based health monitoring applications. Even though these numbers present a promising future for wearable-based health monitoring solutions, our review indicates that there still exist some crucial hurdles before implementing health monitoring devices and applications in real-time [
6]. These include:
Focusing on developing physiological signal analysis algorithms which promote edge computing approaches [
4,
5,
6,
9]. That is, the data acquisition, compression and analysis must be done at the device level without having the need to transmit long, streaming data to cloud services. This would lead to optimization of cloud resources by minimizing usage for data storage and analysis. The idea of edge computing is to help in optimizing on-device memory and power usage, thereby increasing operating efficiency and throughput [
5,
9].
In addition to this, there is also a need for data acquisition standardization with respect to data formats and communication protocols [
10,
11].
Ensuring seamless Internet connectivity across users, devices, infrastructure and services.
Developing safe, non-invasive and comfortable wearables embedded with sensors for collecting and processing physiological data in a remote setting.
Meeting these challenges, could not only establish a set of standards with respect to device manufacturing and developing new communication protocols, but would also promote the development of novel data acquisition and storage algorithms in wearables. Since the most common sensor currently used in wearables is the accelerometer [
6,
8], we focus our study on activity monitoring applications. Note that wearables embedded exclusively with accelerometers are termed as actigraphs [
12]. In the following section, we will discuss actigraphy applications, data acquisition and signal analysis.
2. Actigraphy
Actigraphs measure human body displacement in single or tri-axial directions, and have been used extensively in calculating gross motor activity for different applications. They are miniature devices which record and store motion data, which could then be further used for performing offline analysis. Actigraphs have been used by researchers in numerous clinical and consumer studies such as fitness monitoring, calorie consumption, sleep/wake activity analysis and for rehabilitation therapies in disabled individuals. To cite a few examples, actigraphy studies have been conducted in the following domains:
Home-based sleep staging [
13,
14,
15].
Analyzing movements in individuals suffering from Parkinson’s and Alzheimer’s disease [
16,
17,
18].
Monitoring home activity of military personnel experiencing post-traumatic stress disorder (PTSD) [
19].
Routine of children diagnosed with autistic spectral disorder and ADHD (attention deficit hyperactivity disorder) [
20,
21].
Estimating the severity of sleep related movement-disorders such as periodic limb movements (PLMs) [
7,
22,
23].
Therapeutic rehabilitation of joint disabilities in war veterans [
24,
25].
Demographic studies for identifying differences in sleep patterns with respect to age, gender, ethnicity and sleep disorder prevalence [
26].
A variety of actigraphs are currently available in the market (see
Figure 3), and they are usually worn on wrist, waist or lower ankles for capturing human motor activity [
27]. Typically, an actigraph is able to capture motion data with a sampling frequency in the range of 16–3200 Hz, coupled with an A-to-D quantization of 6–16 bits per sample, depending on the manufacturer [
7,
12,
27,
28] .
The reader must note that, due to device property variability from one manufacturer to another, data analysis of the same activity captured from two different actigraphs, might yield different results. This infers that actigraphy analysis algorithms must be designed to be device-independent and customizable as per application [
6,
29]. Typically, an actigraph consists of the following components [
12,
30]:
Piezoelectric accelerometer for capturing motion/vibrations.
Signal amplifier coupled with an A-to-D converter.
low-pass filter to remove external vibrations.
Flash-memory to store sampled and filtered amplitudes.
Capacitive and rechargeable battery.
A micro-USBTM, serial or low power wireless interface to transfer data to a local computer.
The actigraph maintains a record of zero-crossings and minimal thresholds, and uses them to generate raw signal values from the motion. Most of the currently available actigraphy devices are able to record and store 24 h motion data for up to a week. Depending on the choice and application domain, actigraphs could be single axial or tri-axial. Note that, usually tri-axial devices are comparatively more sensitive than single axial ones, and may capture motion in scenarios which require real-time data analysis.
Figure 4 illustrates single and tri-axial actigraphy signals captured from two different actigraphs.
In case of tri-axial actigraphy data, our review of prior studies indicates that one must perform vector compounding of individual axial data before analysis, in order to simplify computations, and most importantly ensure that vibration information from all three directions is captured [
14,
31,
32]. For example, given a tri-axial signal
, its vector magnitude would be computed as,
In order to analyze an actigraphy signal, we must first run certain signal property tests to determine appropriate processing tools and techniques [
29]. Following
Table 1 highlights various tests and our observations on actigraphy data, computed in MATLAB
TM.
Before an actigraphy signal is analyzed to detect specific movements or patterns, it must be pre-processed in order to remove noise and artifacts. Conventionally, actigraphy signals undergo the following operations before analysis:
- (1)
A-to-D conversion in order to assign discrete amplitudes to specific movements [
29].
- (2)
As per our literature review, human activity is usually captured in the 0.3 to 6 Hz frequency range, and high frequency noise is captured around the sampling frequency. In order to remove the noise, a simple low-pass filter (Butterworth) is employed to capture movement data [
12,
14,
31,
32].
- (3)
Additional band-pass filters could be implemented in order to remove low frequency artifacts and noise.
- (4)
Depending on application, the actigraphy signal is annotated using time-stamps. For example, in many sleep studies, actigraphy data was clipped between “Lights-off” and “Lights-on” time periods, in order to ensure alignment with other clinical signals recorded in simultaneous PSG [
7].
Although most actigraphs are designed for long-term recordings, there are certain shortcomings in their data acquisition and storage methods, which need to be met in order to optimize their usage and implementation as standalone devices, or in smart wearables. These limitations could be:
- (1)
Actigraphs that sample data at higher frequencies (typically 100 Hz and above) along with a high quantization rate (typically 12–16 bits per sample), often lead to memory leakage and underutilization of battery life during recording.
- (2)
Manufacturer-based variability in sampling and quantization. This limits algorithms from being designed as device-independent tools [
27,
37]. Some actigraphs tend to sample movement data too infrequently, thus leading to information loss in the output raw signal.
- (3)
Many prior studies have been conducted on short-duration actigraphy datasets and did not require extensive memory and computational resources for analysis [
14,
22]. Translating these studies into long-term activity monitoring solutions is not feasible unless the actigraphy data is subjected to significant compression and segmentation at the source.
- (4)
Increased use of computational resources (local or cloud) during offline processing of long-term recordings. Conventionally, actigraphy data is captured and entirely transferred to a local computer or cloud for analysis. Our review indicates that in most studies, no prior data processing is done at the source to retain only meaningful information and discard redundant values.
As stated in previous section, signal acquisition methods which promote an edge computing approach could overcome the afore-mentioned challenges in long-duration actigraphy data analysis and optimize device usage [
5,
6]. In the following section, we propose one such technique to pre-processing actigraphy data by performing data compression and denoising at the source. It should be noted that the proposed solution in this study is not an edge computing technique in itself, but rather focuses on optimizing data acquisition and storage which would then promote edge computing on the hardware.
Proposed Approach
In our review of actigraphy signals captured from different studies and applications, we found that employing a lower level of quantization to actigraphy data at the source, addresses a significant number of afore mentioned challenges. In this study, we propose a low-level encoding scheme which would improve actigraphy analysis in the following ways:
- (1)
Data compression at the source. The proposed encoding method intends to reduce the output actigraphy file size, thus enabling faster transfer and read time on a local computer.
- (2)
Signal normalization and denoising, which removes redundant and minute vibrations captured from highly sensitive accelerometers.
- (3)
SNR (signal-to-noise ratio) increase and enhancement of meaningful movement amplitudes in the signal.
- (4)
The proposed scheme also ensures operation across different types of actigraphs, thus promoting device-independency of this algorithm.
The reader must note that data compression might result an increase in energy consumption and latency at the source. But the proposed solution intends to reduce memory usage and optimize overall battery usage, which would balance-off these shortcomings.
Figure 5 illustrates the methodology implemented in this study.
In order to conduct a systematic investigation, we have conducted experiments on actigraphy data acquired from the following applications:
- (1)
Long-duration tri-axial actigraphy signals captured simultaneously with polysomnography in sleep studies [
28].
- (2)
Activities of Daily Life (ADL) dataset obtained from Dua et al. [
38].
- (3)
Vibroarthrographic signals captured from knee joints for osteoarthritis severity assessment [
39].
The reader must note that in case of long-duration sleep actigraphy signals, the proposed encoding scheme’s results have already been published in [
28] by Athavale et al., and hence we’ve shown the same results in this paper, to augment our experiments with daily activity [
38] and vibroarthrography datasets [
39].
For the reader’s reference, this paper has been further organized as follows: In
Section 3.1 we will briefly explain the datasets used in our experiments, along with actigraph and signal properties used in each study. Next, in
Section 3.2 we explain the proposed signal encoding scheme. Following this, we then proceed to check the validity of the proposed encoding scheme by performing simple machine learning and pattern classification of encoded signals, and comparing its results with those of raw actigraphy signals from each dataset, in
Section 3.3 . In the next
Section 4.1 and
Section 4.2 we present our experimental results from signal encoding and its validation. We finally conclude this paper with some critical discussions in
Section 5.
5. Discussions and Future Works
As evident from our investigation and experimental results, employing a very low-factor signal quantization greatly improves the device’s data handling capacity by ensuring enhanced SNR, high compression ratio and removal of redundant movement information from the actigraphy signal. The 3-bit encoding proposed in this study, works best in compressing actigraphy data at the edge of an IoT-type setup. Considering the nature of actigraphy signals as highlighted in
Table 1, the proposed encoding scheme addresses the transient, spiky information by retaining only significant movement amplitudes or true acceleration values. Movements which are very small are floored to zero in the encoding operation. Thus, redundant values and high frequency noise are removed in the encoded signal, which now contains only relevant movement information.
Although in this study we have used offline datasets, it must be noted that the objective of the proposed encoding scheme is to be applied at the recording source (i.e., on the device) in real-time. This supports an edge computing approach when coupled with activity-based adaptive segmentation techniques to extract regions of peak movements. The machine learning validation approach used in this study aptly supports the proposed encoding scheme as shown by the classification results in
Table 5. Further to this, we observe that the 3-bit encoding provides the highest activity recognition rate. From our study on different actigraphy datasets, it should be noted that the proposed encoding algorithm is device-independent and signal-independent, and could easily be ported onto any accelerometer-based wearable.
Current trends in IoMT and related device developments highly promote the edge computing structure in smart devices, as it would significantly reduce cloud burden, and ensure data privacy and security at the consumer end. Home-based health monitoring using an IoMT framework is a burgeoning market and would help in significant reduction of patient-doctor visits and associated healthcare costs. One way to encourage this trend is to use wearables and sensors, embedded with edge computing friendly algorithms, such as the one proposed in this study. This would also promote the clinical validation and development of tools for long-term monitoring of vital physiological parameters in not just chronically ill or elderly patients, but for the betterment of all individuals [
6,
43].
As part of our future work, we would like to test the proposed algorithm’s efficiency on commercially available wearables such as FitBitTM, Apple WatchTM as well as other generic actigraphs used in activity monitoring studies.