Ambient Assisted Living environments use different sensors and actuators to enable their endusers to live in their preferred environments. Unlike smart homes, where a target audience is usually a family unit, standard Ambient Assisted Living end users are care receivers and care providers. This article describes an approach based on the fog computing paradigm to detect sleep apnea in an Ambient Assisted Living context unobtrusively. The edge nodes process and detect local activities of daily living events and have direct control of the local environment. The fog nodes are used to further process and transmit data. The cloud is used for more complex and anonymous data computation. This research shows that sensors, which are unobtrusive and do not interfere with users' daily routines, can be successfully used for pattern observation.
Advancements in cloud computing and the
Internet of Things (IoT) have had a positive
impact on pervasive computing and can
improve Ambient Assisted Living (AAL)
solutions. Fog computing is a newer discipline
that brings an opportunity to fill in some gaps
and improve many aspects of cloud-based AAL
systems, mainly by increasing user privacy if
used correctly [
Technology for monitoring, assisting, and
improving personal health has improved
considerably with affordable wearable and
unobtrusive sensors, cloud computing, and
improved Internet connectivity. The presence
and rapid growth of the Internet of Things (IoT)
paradigm has also impacted how people
monitor their health. Most current wearable
devices can monitor heart rate and physical
activity. More appliances come with Internet
connection capability, and smart sensors are
becoming increasingly common. The data
obtained with unobtrusive sensing can give a
more detailed picture of the care receivers'
health and personal habits. In that way,
technology directly impacts elderly and
disabled people’s ability to remain at home and
live more independent lives [
A typical AAL goal is to enable care
providers to have technology-enabled
continuous monitoring of care receivers. It
reduces care costs on the one hand and
increases care efficiency on the other hand
[
A single device, individual, or group of
unobtrusive sensors present in the ELE can
provide input on a limited set of health aspects.
Smartwatches and health trackers can track
body temperature, heart rate, walking or
running; environmental sensors can detect
temperature, humidity, fall detection, and
movement within the home. A more holistic
picture of these devices and sensors can be
provided if connected to the cloud, where all the
data is analyzed for more robust data processing
techniques. By using data from many users,
machine learning (ML) algorithms can learn
and predict health hazards and find correlations
between the environment and human health
[
While IoT cloud-based computing benefits
are visible both in research and daily use, there
are many drawbacks when it comes to personal
health care data clouds. The most significant
ones are the following:
• The lack of security of IoT devices and
companies' un-proper practices that gather
and abuse personal data have made
consumers more proactive in protecting
their data [
In this paper, we identify multiple benefits of fog computing in the typical AAL scenario and propose an architecture that would make them possible. The AAL fog-based architecture is described in Section 2 of this paper. The proposed architecture benefits are illustrated with the experiment presented in Section 3, and Section 4 concludes the paper.
Fog computing adds a layer to the cloud
computing architecture. However, it should not
be interpreted as an extension of the cloud. Fog
computing spans to adjacent physical locations.
It supports online analytics and various
communications networks in performing
distributed computing [
Data is generated by sensors that can be wearable or body sensors and peripheral or environmental sensors on the first layer. Data can also be generated from external sources such are: social networks, clinical center information systems, or medical databases. Data collected by the sensors can include vital signs, personal habits, or environmental factors. External data sources can provide different information, including medical check results, medical databases for diagnostics, and similar.
The fog layer gathers the sensor data, processes them, and passes either processed or portions of raw data to the cloud. The devices directly connected to the sensors are called edge nodes. Aside from collecting data, they can take action with the user. Each LAN environment can have one or more edge nodes, depending on the application requirements and scale. In elderly care facilities, data for multiple tenants could be processed on the same edge nodes. These actions can include providing feedback to the person to take their medicine or to start exercising. They can directly interact with the environment, such as: activating the humidifier or regulating the room temperature, controlling electrical appliances, and cutting-off for water, gas, and electricity in case of an emergency. The fog network usually has a more limited capacity than the cloud for data computation and cannot do complex machine learning and feature extraction.
However, fog nodes could be able to run algorithms developed by machine learning. As the machine learning system improved and evolved, regular updates could be pushed down to the fog network to improve sensor data patterns. Using this methodology, ADL detecting ML could receive continuous data and improve the detection rate. Events that take a brief time, such as when a person falls, can be detected by the fog nodes using the latest ML model improved in the cloud.
The cloud layer assembles and processes the
data from multiple sources and creates machine
learning models. The feature extraction is done
at this layer as well. Data from the fog and
external sources is collected and processed by
the data fusion component [
The service layer is the product of the system. Knowledge obtained by analyzing the data is used for services, including creating customized recommendations for diet and exercise, improving diagnostics systems, providing updates to the health providers, and adding additional information in medical databases.
The critical features that should be satisfied
by the system include security, privacy, high
availability, and interoperability. Security and
privacy [
In fog computing, the nodes nearest to the
devices are named edge nodes. In healthcare
systems, these nodes represent smart e-health
gateways. They act as a bridge for medical
sensors to cloud computing platforms. The
main requirement of a gateway is to support
various wireless protocols and inter-device
communication. Its role can be extended to
support several features such as acting as a
repository to temporarily store sensors’ and
users’ information and bring intelligence by
enhancing data fusion, aggregation, and
interpretation techniques. It is essential to
provide preliminary local processing of sensors'
data, which is the primary role of a smart
ehealth gateway. Smart e-health gateway can
tackle many challenges in ubiquitous healthcare
systems such as energy efficiency, scalability,
interoperability, and reliability issues [
Due to the privacy concerns and the technical aspects for scalability and interoperability, it is crucial to identify and trace the data flow in the system. Sensor data originates when sensors acquire measurement from the physical world. This measurement is represented by an electrical signal transferred to a controller that would interpret the signal. Some sensors are manufactured to include the electronic circuits to digitalize the reading, and some are even Internet-connected, enabling them to upload the data to a remote system directly. The sensor data is then passed to the local processing nodes. These nodes are part of the fog and can communicate to other layers of the fog. The data on these edge nodes is processed for local events detection.
Only the edge nodes or smart e-health
gateways should be able to get unfiltered raw
sensor data. The data that is passed on to other
layers of the fog is pre-processed [
Each of the services dealing with user data is logically independent and can be hosted on separate cloud platforms. The health provider service is independent of social media or medical research databases. The separation of the cloud can be implemented by separation on any level in the fog network. As the data is passed between layers of the fog network, several processing types can occur. Data processing tasks mostly would take place on the edge nodes or smart e-health gateways as the gateways would directly interface with the sensor network and receive raw sensor data. At this layer, we can identify the following types of tasks:
Data filtering is used to filter noise, invalid
sensor readings, and redundant information that
does not contribute to the desired information
the system should induce. Sensor data contains
valuable information. However, they also carry
non-deterministic errors such as motion
artifacts, data corruption issues, and unwanted
signals that are also significantly uploaded to
increasing storage requirements and power
consumption. Fog computing could play an
essential role in increasing efficiency and
reduce storage requirements for medical big
data solutions [
Anonymizing of data strips or replacing person-identifying information from data packets. When there is a requirement to separate patient/customer data, personal information is replaced with unique identifiers. This data can be passed on to the fog nodes for added security using an enterprise service bus (ESB).
Data fusion automatically transforms
information from different sources and points
into a representation that provides practical
support for automated decision-making.
Applying data fusion in gateways provides
several advantages: reduced data ambiguity,
extended coverage in space and time,
robustness and reliability, and increased data
quality. After data is fused, only final results are
transmitted through the network so the network
bandwidth can be efficiently utilized, and the
system can be more energy-efficient [
Data processing that can be done on any
layer of the fog network includes:
• Data compression is used to reduce the
amount of bandwidth required to transmit
the sensor network's information.
Compression can be lossy or lossless. Lossy
compression can be acceptable in many
cases, especially if the sensor data's
resolution is too high. Besides, the extra
information will not cause significant
improvements in the algorithms [
A common usage of sensor networks is to train machine learning models and enable different end-user actions. Depending on the number of sensors used, the number of features extracted from the sensor data, and the data generation rate, generating the model will most likely be done in the cloud due to the resources demand and a potential need to use data from other locations. On the other hand, the implementation of the model can and should be done on edge. As an example, we will consider a data flow model to detect sleep apnea using noninvasive sensors, illustrated in Figure 1.
The strong correlation between the two sensor types, shown in the diagram of recorded sensor data over 8 hours, is presented in Figure 4.
When motion is detected, the data from
multiple noninvasive sensors is processed on
the edge node. The local machine learning
model is run, and the possible occurrence of
sleep apnea is diagnosed. Periodical sessions
with invasive sensors or medical professionals'
observations can be carried out to label the data
set [
After anonymization of the data, it is packaged and sent to the cloud for additional processing. The data model on the cloud side is run to verify the outcome for the received data. If the model present in the cloud makes positive detection for the received data and if the data was previously labeled with a negative result by the edge node, then the updated model is sent back to the edge node, which in turn processes the data against the updated model.
Sensor data that does not suggest strong negative results are marked for further labeling if additional data such as monitoring from medical equipment or video that can be analyzed by a trained professional is available. Such feedback is periodically included to build the cloud model continuously. 3.1.
This section explains the classification
algorithms used for feature ranking and
construction classification models. The
accuracy was used for the comparison of
various classification models throughout the
system. One of the classification algorithms
used in our experiments is logistic regression
[
Random Forest (RF) [
The tree branching is performed by finding
the best split from the features on each node.
During classification, trees vote for the class,
and the majority class is eventually predicted.
Like RF, the Extremely Randomized Trees
(ERT) algorithm [
Both ERT and RF provide feature importance estimates, a property used for feature ranking and discarding of lowimportance features during the feature selection phase. We have used the feature importance estimates when training an ERT classifier due to its better speed than RF.
Additionally, we have also used the Support
Vector Machines (SVM) classifier [
The measurements from sensors can detect atomic actions or states. More complex actions are depending on the context, which recent measurements can determine. Therefore, the data needs to be first adequately segmented, and then feature extraction performed [30]. This study additionally discusses the window size impact on activity recognition. Generally, lower sensor frequencies entail longer windows. It is considered during our experiments by using different window lengths and analyzing the accuracy depending on them. The segmentation into windows, step 1 on Figure 5, was performed, thus excluding the border intervals when the activity changes from one activity to another.
Segmenting of streaming data into windows is performed in step 2 in Figure 5. Step 2 extracts the following types of features (the number of measurements within one window is denoted by n.): • Basic statistics results in 14 features per time series. • Equal-width histogram calculated with [log! +1] intervals, based on the Sturges rule [31]. It results in 5 to 8 features when the window length varies from 5s (25 measurements) to 20s (100 measurements). • Quantile-based features: first quartile, median, third quartile, interquartile ranges, and other percentiles (5, 10, 20, 30, 40, 60, 70, 80, 90, 95), also used in [32]. From onetime series, it generates 14 features. • Auto-correlation of the measurements within one sliding window [33]. Let τ denote the amount of shift, and its domain is defined as τ ∈ [1, ë n û] • For exponentially increasing values of τ in that range, classical autocorrelation and Pearson correlation are calculated. Additionally, it calculates both correlations using the first and second half of measurements within one sliding window. This results in 3 to 4 τ values when the window length varies from 5s (25 measurements) to 20s (100 measurements). • Pearson correlations between pairs of time series; For five-time series, this results in 9 features. • Linear and quadratic fit coefficients; There are two linear fit and three quadratic fit coefficients, yielding five features in total per time series. • As a result of step 2, 250 to 270 features are generated depending on the window length. In step 3 performs feature importance and drift sensitivity estimation is done. Next, step 4 performs coarse-grained feature selection, which tests a set of thresholds used to discard features with low importance or high drift sensitivity.
The system evaluates different feature sets by building classification models using the training dataset and evaluating them with the validation dataset. The test set is not utilized at this stage at all. Thus, only the feature set that results in the best classification accuracy is retained. To summarize, the purpose of this step is to significantly reduce the feature set size by discarding features with low importance or high data drift sensitivity.
The system evaluates different feature sets by building classification models using the training dataset and evaluating them with the validation dataset. The test set is not utilized at this stage at all. Thus, only the feature set that results in the best classification accuracy is retained. To summarize, the purpose of this step is to significantly reduce the feature set size by discarding features with low importance or high data drift sensitivity.
After the feature set is reduced, step 5 uses the training and validation sets to perform parameter tuning for the SVM.
Finally, step 6 evaluates different classifiers by building classification models with the training and validation dataset's union and evaluating it using the independent test set.
The duration of our experiment was 8 hours. The sampling rate was set to 10Hz, thereby measuring ten values from each sensor every second. We divided the dataset into three different subsets: training, validation, and testing. The training subset consisted of the first 45% records for each action, and the validation subset consisted of the next 25% records. The remaining 30% of records belonged to the test subset. When performing parameter tuning for SVMs and making feature selection, the training set was used to build models, and the validation set was used to evaluate their performance. Once this phase was completed, the final evaluation was performed only with the best feature set decided after the feature screening and using the most optimal parameters. The union of the training and validation sets was used to build classification models for making final predictions. The test set was used for building predictions and the performance evaluation.
As personal health becomes pervasive and the data generated by it increases in volume, fog computing offers a solution for many critical challenges. The added flexibility of the fog architecture enables better placement of computing and network resources. Smarter data flow could protect personal data, bandwidth cost could be reduced, and more scalable, secure, and interoperable systems can be designed. This paper illustrates those benefits by providing an experimental illustration of typical AAL service provided by fog-based health care ELE.
By using simple hardware, the AAL data was streamed to a cloud-based system, where it was fused. Using a systematic and automated feature extraction and selection process, we could extract robust and reliable features that facilitated building powerful classification models.
This work was partially financed by the Faculty of Computer Science and Engineering at the Ss. Cyril and Methodius University, Skopje, Macedonia and is supported by the networking activities provided by the ICT COST Actions IC1303 AAPELE and CA16226 SHELD-ON.
We also acknowledge Microsoft Azure's support for research through a grant providing computational resources for this work.
Learning Research, vol. 12, 2011, pp. 2825–2830. [30] O. Banos, et al., “Window size impact in human activity recognition,” Sensors, vol. 14, no. 4, 2014. DOI: 10.3390/s140406474, pp. 6474–6499. [31] H. A. Sturges, “The choice of a class interval,” Journal of the American Statistical Association, vol. 21, no. 153, 1926, pp. 65–66. [32] P. Siirtola and J. R¨oning, “Recognizing human activities user independently on smartphones based on accelerometer data,” International Journal of Artificial Intelligence and Interactive Multimedia, vol. 1, no. 5, 2012, pp. 38–45. [33] H. Mart´ın, A. M. Bernardos, J. Iglesias, and J. R. Casar, “Activity logging using lightweight classification techniques in mobile devices,” Personal and Ubiquitous Computing, vol. 17, no. 4, 2013. doi: 10.1007/s00779-012-0515-4, pp. 675– 695.