Edge-AI in LoRa-based Health Monitoring: Fall Detection System with Fog Computing and LSTM Recurrent Neural Networks

Abstract

Remote healthcare monitoring has exponentially grown over the past decade together with the increasing penetration of Internet of Things (IoT) platforms. IoT-based health systems help to improve the quality of healthcare services through real-time data acquisition and processing. However, traditional IoT architectures have some limitations. For instance, they cannot properly function in areas with poor or unstable Internet. Low power wide area network (LPWAN) technologies, including long-range communication protocols such as LoRa, are a potential candidate to overcome the lacking network infrastructure. Nevertheless, LPWANs have limited transmission bandwidth not suitable for high data rate applications such as fall detection systems or electrocardiography monitoring. Therefore, data processing and compression are required at the edge of the network. We propose a system architecture with integrated artificial intelligence that combines Edge and Fog computing, LPWAN technology, IoT and deep learning algorithms to perform health monitoring tasks. In particular, we demonstrate the feasibility and effectiveness of this architecture via a use case of fall detection using recurrent neural networks. We have implemented a fall detection system from the sensor node and Edge gateway to cloud services and end-user applications. The system uses inertial data as input and achieves an average precision of over 90\% and an average recall over 95\% in fall detection.

Full text

Edge-AI in LoRa-based Health Monitoring:
Fall Detection System with Fog Computing and
LSTM Recurrent Neural Networks
J. Pe˜
na Queralta1, T. N. Gia1, H. Tenhunen2and T. Westerlund1
1Department of Future Technologies, University of Turku, Turku, Finland
2Department of Electronics, KTH Royal Institute of Technology, Stockholm, Sweden
Email: 1{jopequ, tunggi, tovewe}@utu.fi, 3hannu@kth.se
Abstract—Remote healthcare monitoring has exponentially
grown over the past decade together with the increasing pen-
etration of Internet of Things (IoT) platforms. IoT-based health
systems help to improve the quality of healthcare services
through real-time data acquisition and processing. However,
traditional IoT architectures have some limitations. For instance,
they cannot properly function in areas with poor or unstable
Internet. Low power wide area network (LPWAN) technologies,
including long-range communication protocols such as LoRa,
are a potential candidate to overcome the lacking network
infrastructure. Nevertheless, LPWANs have limited transmission
bandwidth not suitable for high data rate applications such as fall
detection systems or electrocardiography monitoring. Therefore,
data processing and compression are required at the edge of
the network. We propose a system architecture with integrated
artificial intelligence that combines Edge and Fog computing,
LPWAN technology, IoT and deep learning algorithms to perform
health monitoring tasks. In particular, we demonstrate the feasi-
bility and effectiveness of this architecture via a use case of fall
detection using recurrent neural networks. We have implemented
a fall detection system from the sensor node and Edge gateway to
cloud services and end-user applications. The system uses inertial
data as input and achieves an average precision of over 90% and
an average recall over 95% in fall detection.
Index Terms—IoT; Edge Computing; Healthcare Monitoring;
LoRa; LPWAN; RNN; LSTM; Fall Detection;
I. INTRODUCTION
Health monitoring plays an important role in disease diag-
nosis and treatments. For instance, electrocardiogram (ECG)
monitoring or fall detection systems can help to detect ab-
normalities and send messages to caregivers about the abnor-
malities in real-time. Recently, fall detection systems using
wearable devices are widely used because of several advan-
tages such as light-weight, low-cost, energy efficiency and
non-intrusiveness [1]–[4]. These wearable devices often collect
3-dimensional (3-D) acceleration or 3-D angular velocity or
both of them from a human body. The devices then transmit
the collected data to a gateway which forwards the data to
cloud. However, there are still drawbacks in these systems.
For instance, they cannot function properly in many scenarios
like areas with unstable or lack of a Internet connection.
LoRa is one of the most prominent Low Power Wide Area
Network (LPWAN) technologies [5]. The LoRa modulation
scheme characterizes for enabling long-range and low power
transmissions [6]. LoRa is widely used in many IoT applica-
tions from farming, agriculture monitoring, flood detection to
metering in smart cities [7]–[9]. The potential of LoRa can be
leveraged mostly in areas with poor connectivity or lack of
infrastructure, but also in dense urban environments to reduce
the number of access points and power consumption. Taking
the above considerations into account, Lora seems to be a good
candidate to overcome the limitations of the existing cloud-
based healthcare monitoring systems that rely on traditional
WAN technologies such as Bluetooth or Wi-Fi.
However, Lora cannot support high data rate (i.e., 250 kbps
in theory). In practice, the data rate is much lower such as
a few bytes per message and a few messages per day, due to
strict regulations of Lora duty cycle (i.e., approximate 1% duty
cycle). It is challenging to satisfy both requirements of high
data-rate applications (i.e., fall detection based on wearable
devices) and Lora duty cycle regulations.
In this paper, we propose an advanced architecture com-
bining Edge computing, Fog computing, LoRa, and IoT-based
technologies. The proposed architecture can inherit the benefits
of these technologies to enhance quality of service. The
proposed architecture can help to overcome the limitations
the existing health monitoring IoT-based systems (e.g., fall
detection or ECG monitoring IoT-based systems) and satisfy
both requirements of high data rate-applications and Lora duty
cycle regulation. We demonstrate the proposed architecture
via a use case of fall detection. In addition, we propose and
implement Edge-AI algorithms based on neural networks at
Edge gateways for improving quality of service. In particular, a
system with the proposed architecture and Edge-AI can detect
human fall cases more accurately and dynamically in different
scenarios.
The remainder of the paper is organized as follows: In
Section II, we present related work in wearable sensors-
based fall detection and implementations of systems relying
on artificial intelligence. Section III introduces the proposed
system architecture; while Section IV describes the imple-
mented prototype and analyses experimental results. Section
V concludes the work.

II. RE LATE D WO RK AND MOT IVATIO N
Many efforts have been devoted to develop fall detection
IoT-based systems [10], [11]. Pivato et al. [12] introduce a
fall detection system which uses a wearable device to collect
and send 3-D acceleration data to a gateway for fall detection.
In [13], authors utilize a smartwatch to collect and send
acceleration to a smartphone which acts as a gateway for
processing data. When a fall case is detected, the smartphone
sends a notification message via 3G/4G to caregivers. Ngu
et al. present a smartwatch-based IoT fall detection system.
The system is able to run Support Vector Machine and Naive
Bayes machine learning algorithms to create the fall model
and detect a fall case with a high level of accuracy. Noury et
al. present an extensive survey of literature on fall detection
[1]. The authors summarize that many proposed methods had
an accuracy of nearly 100%, which is essential in applications
where a person’s life might be at risk. However, in practical
scenarios, this figure decreases dramatically. Therefore, they
defended the necessity of a common framework for accurately
evaluating fall detection systems and protocols.
More recently, researchers have been applying deep learning
techniques for fall detection using both active and passive sen-
sors. In [14], the authors use image processing to detect falls
based on video feeds. They use convolutional neural networks
(CNN) and a fully connected neural network for extracting
features and classifying situations, respectively. Their method
has an accuracy ranging from 90% to 96%. We propose the
use of a wearable device, instead, as the number of scenarios
where it can be used is broader, requires less amount of data to
be analyzed and similar performance can be achieved. Other
authors have explored the use of cameras or radio waves [15],
[16].
Fakhrulddin et al. develop a fall detection method for
body sensor networks using CNNs [17]. The authors use data
from two accelerometers as input to the network and obtain
an accuracy varying from 75% to 92% depending on the
dataset used. In our work, we defend that recurrent neural
networks (RNN) are a better alternative to CNN because of the
importance of time as a factor in the decision process. RNNs
have been effectively used by Musci et al. for designing an
accurate fall detection method tested over the SisFall dataset
[18]. The authors developed a method for analyzing data
online in real-time that was executed with the aid of GPUs. In
our work, we focus on similar methods for resource constraint
single-board computers that operate as edge gateways.
In summary, higher accuracy and adaptability to a wider
range of situations can be achieved with deep learning when
compared to threshold-based fall detection or SVM classifica-
tion. In particular, recurrent neural networks show promising
results in fall detection applications. However, previous works
focus on the analysis part or assume that cloud computational
capabilities are available. This requires full sequences of raw
data to be transmitted to the cloud and increases the alert
latency. Instead, we propose the use of lightweight analysis
algorithms that can run on edge gateways. At the same time,
this enables the system to be deployed in rural areas or
scenarios with poor connectivity as the amount of data that is
transmitted over the Internet to cloud servers can be decreased
several orders of magnitude.
III. SYSTEM ARCHITECTURE
We propose a five-layer system architecture consisting of
wearable devices (sensor layer), smart edge gateways (edge
layer), LoRa access points (fog layer) and cloud services
(cloud layer) and end-user terminals (application layer). The
proposed architecture can be used for different health monitor-
ing applications such as cardiovascular or diabetes monitoring.
Sensor nodes in the proposed architecture can collect differ-
ent types of data including e-health (e.g., electroencephalog-
raphy (EEG) electrocardiography (ECG), electromyography
(EMG), and blood pressure) and contextual data such as tem-
perature, humidity, and air quality. The combination of both
e-health and contextual data helps to improve the accuracy of
disease diagnosis and analysis. The collected data is sent via
Bluetooth Low Energy (BLE) to an Edge gateway for data
processing.
At Edge-gateways, artificial intelligence algorithms are ap-
plied to enhance quality of service. For instance, the AI-
service can detect a human fall with a high level of accuracy.
Then, the results are sent to a Lora-based access point for
storing in a distributed manner and processing with some
advanced algorithms. Finally, the processed data or results
are sent to cloud servers for final data processing and global
storage. This mechanism helps to reduce latency of sending
a large amount of data via LoRa network. In addition, this
can ensure that the bandwidth can be efficiently utilized
and LoRa duty cycle regulations are satisfied. Furthermore,
Edge-gateways can provide many advanced services such as
distributed storage, security, and localization. These services
altogether with services implemented at Fog-assisted LoRa-
based access points help to improve the quality of healthcare
services. For instance, a Fog-cloud-based push notification
service sends real-time messages to caregivers in case of a
fall or an abnormality. Due to the scope of the paper, Fog
services are not discussed. More detailed information of the
Fog services is discussed in our previous papers [19]–[22].
The proposed architecture can help to reduce the computa-
tional load on the sensor nodes by switching heavy compu-
tational tasks from sensor nodes to Edge-gateways. In case
of fall detection systems, data is processed with advanced
algorithms such as AI-based activity categorization algorithms.
Only results such as generic activity status are sent to the Fog-
assisted access points which then transmit to cloud servers.
End-users can use terminal applications to access results stored
in cloud servers.
IV. EXP ERIME NTAL AN ALYS IS
In order to test the feasibility of the proposed architecture,
we have implemented the complete system for the use case
of fall detection. Wearable devices equipped with inertial
measurement unit (i.e., MPU9250 3-axis accelerometer, 3-axis

SENSORLAYER
SENSORLAYER
EDGELAYER FOGLAYER CLOUDLAYER ENDUSER
APPLICATION
LoRa Access Points
Distributed Storage
Global Storage
Cloud Services
Web/Mobile Application Servers
Bio-signals analysis
Fall detection
Activity status
Fig. 1. Proposed System Architecture
gyroscope, and 3-axis magnetometer) to collect and transmit
the data to an Edge-gateway via Bluetooth Low Energy (BLE).
The sensor node is equipped an AVR 8-bit MCU and supplied
with 3.3 V. The gateways have been implemented using a
Raspberry Pi 3 Model B running Ubuntu Mate Desktop with
a Dragino LoRa shield with communication via SPI. The
LoRa access point is also implemented with a Raspberry Pi
single board computer and a LoRa shield, directly connected
to the Internet. For this experiment, we have used raw LoRa
with customized data format and encryption, instead of using
LoRaWAN as the link a network layer over LoRa. This
allows us to customize further the transmission frequencies
and packet structures. Data from the inertial measurement unit
is analyzed at the edge gateway and then only information
about the status of the patient and instant notifications in case
of a fall are transmitted over LoRa. A PostgreSQL database
is used as a cloud storage solution and the web monitoring
application for end-users is implemented using Django and
Apache on CentOS.
The deep learning algorithms have been trained and eval-
uated with a public dataset. However, we have implemented
the wearable sensor node in a way such that the data format is
equivalent to that of the used dataset. Data is normalized and
preprocessed before the analysis step so that data from dif-
ferent sensors can be used with the same algorithm, enabling
flexibility in the design of the sensor node.
A. Edge-AI: LSTM RNN for Fall Detection
We have implemented a recurrent neural network (RNN)
with long short-term memory layers (LSTM) cell layers. A
recurrent neural network is, essentially, a special case of
densely connected neural networks where time is introduced
in the form of connections across consecutive time steps.
They are particularly useful in applications involving time
series data such as handwriting or speech recognition. Though
RNNs inherently store previous states, in practice they present
vanishing and exploding gradient problems in their raw form.
To reduce the impact of these problems, LSTM cells help to
decrease the vanishing gradient problem allowing longer-term
memory within the neural network [23].
We have used Keras and Tensorflow as its backend to
implement an LSTM RNN using Keras and Tensorflow as its
back-end [24]. We have run several tests and compared the
accuracy of the model with different network structures.
For training and assessing the accuracy of the models, we
have used the MobiAct Dataset [25]. The dataset contains
acceleration, angular velocity and orientation data. In particu-
lar, we use a subset containing two daily activities (standing,
lying) and four types of falls: forward-lying (fall forward from
standing, using hands for dampening), front-knees-lying (fall
forward from standing, first impact on knees), sideways-lying
(fall sideways from standing, bending legs), and back-sitting-
chair (fall backward while trying to sit on a chair).
We have implemented different RNN models with a variable
number of hidden layers and their sizes. The best results have
been obtained with three hidden layers, two of them LSTM
and one fully connected layer. The input data size is 10 points,
and the output of the network is a single value (probability
of fall occurring in the input data). Therefore, the proposed
model has a total of 5 layers with 2 dropout operations after
the LSTM layers.
B. Results
We have tested the efficiency of five different RNNs and
compared their accuracy in terms of precision (TP/(TP+FP))
and recall (TP/(TP+FN)), where TP,FP are true/false positives,
and TN,FN are true/false negatives.
Figure 2 shows the distribution of the results in the form
of boxplots, with 20 data points for each model. We started
the test with a simple RNN model (M1) with 2 hidden LTSM
layers with 23 neurons each, which allows us to obtain an
average precision of 91.90% ±4.46%, and an average recall
of 62.36% ±2.83%. Even though the precision is good, the
standard deviation in both cases is high and the recall is
too low. More robust results have been obtained with 30
neurons/layer (M2); 10 neurons/layer and an additional fully

M1 M2 M3 M4 M5
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Precision
Recall
Fig. 2. Precision and recall for the five different implemented models.
connected layer of 10 neurons (M3). An improved recall was
obtained when adding a dropout stage of 50% after the two
LSTM layers of M3 (M4). Finally, the best recall was achieved
when using two dropout stages after each of the LSTM laters
(M5), of 30% and 20% respectively. At the same time, the
standard deviation of the results was significantly reduced.
In summary, the best performance was obtained with a
neural network with 3 hidden layers and two dropout stages.
The final precision achieved was of 90.10% ±3%, and a
recall of 95.30% ±0.8%. This shows an improvement in
comparison to our previous work [2], [4], where threshold-
based fall detection was implemented. After the data analysis,
only 10 bytes of data have to be transmitted from the Edge
gateway to the Fog-assisted access point and to the cloud if a
fall is detected, including the unique device ID and the user
status. This reduces the bandwidth usage and allows hundreds
or thousands of devices to use the same LoRa access point
due to the infrequent transmissions. The LoRa transmission
has been tested in an urban environment in Turku, Finland. A
range of over 4km has been achieved with the LoRa access
point over a hill. This long-range has been achieved partly due
to the low buildings in the area. The transmission range can
be significantly extended for a system deployed in rural areas.
V. CONCLUSION
We have proposed a system architecture for health monitor-
ing with Edge and Fog computing. We have put an emphasis
on the use of LPWAN technology to enable deployment of
such systems in rural areas. Taking into account the network
capacity in these scenarios, robust data analysis and compres-
sion algorithms are deployed in the edge layer to lower the
size of transmitted data, improving the system latency.
We have shown how high accuracy in a fall detection system
can be achieved by means of an LTSM RNN implemented
to run on the edge gateways. This enables real-time alerts
and notifications, and waives the need for raw data to be
transmitted to cloud servers for online analysis. Furthermore,
when combined with BLE at the sensor node and LoRa
transmission from edge to fog layers, we are able to simplify
the sensor node design and potentially increase its battery life,
while allowing operation in areas with poor connectivity.
Future work will include further improvement of our predic-
tion model and more extensive performance analysis. More-
over, we will evaluate our method with real-time data from the
implemented sensor node, and the system architecture will be
tested as a whole. Finally, we will expand our model to classify
different daily activities apart from fall detection.
REFERENCES
[1] N. Noury et al. Fall detection - principles and methods. In 2007 29th
Annual International Conference of the IEEE Engineering in Medicine
and Biology Society, pages 1663–1666, Aug 2007.
[2] T. N. Gia et al. Iot-based fall detection system with energy efficient
sensor nodes. In 2016 IEEE (NORCAS), pages 1–6, 2016.
[3] I. Tcarenko et al. Energy-efficient iot-enabled fall detection system with
messenger-based notification. In International Conference on Wireless
Mobile Communication and Healthcare, pages 19–26. Springer, 2016.
[4] T. N. Gia et al. Energy efficient wearable sensor node for iot-based fall
detection systems. Microprocessors and Microsystems, 56:34–46, 2018.
[5] U. Raza et al. Low Power Wide Area Networks: An Overview. IEEE
Communications Surveys Tutorials, 19(2), Secondquarter 2017.
[6] Semtech Corporation. Lora™modulation basics. , May 2015.
[7] A. J. Wixted et al. Evaluation of LoRa and LoRaWAN for wireless
sensor networks. In 2016 IEEE SENSORS, Oct 2016.
[8] D. Ismail et al. Low-power wide-area networks: Opportunities, chal-
lenges, and directions. In 19th ICDCN. ACM, 2018.
[9] K. Mekki et al. A comparative study of LPWAN technologies for large-
scale IoT deployment. ICT Express, 2018.
[10] A.K. Bourke et al. Evaluation of a threshold-based tri-axial accelerom-
eter fall detection algorithm. Gait & Posture, 26(2):194 – 199, 2007.
[11] J. Chen et al. Wearable sensors for reliable fall detection. In 2005 IEEE
Engineering in Medicine and Biology 27th Annual Conference, 2005.
[12] P. Pivato et al. A wearable wireless sensor node for body fall detection.
In 2011 IEEE International Workshop on Measurements and Networking
Proceedings (M&N), pages 116–121. IEEE, 2011.
[13] E. Casilari et al. Automatic fall detection system based on the combined
use of a smartphone and a smartwatch. PloS one, 10(11), 2015.
[14] Y. SunGil et al. An artificial neural network–based fall detection.
International Journal of Engineering Business Management, 10, 2018.
[15] K. Adhikari et al. Activity recognition for indoor fall detection using
convolutional neural network. In 2017 Fifteenth IAPR International
Conference on Machine Vision Applications (MVA), pages 81–84, 2017.
[16] Y. Tian et al. Rf-based fall monitoring using convolutional neural
networks. Proc. ACM Interact. Mob. Wearable Ubiquitous Technol.,
2(3):137:1–137:24, September 2018.
[17] A. H. Fakhrulddin, X. Fei, and H. Li. Convolutional neural networks
(cnn) based human fall detection on body sensor networks (bsn) sensor
data. In 2017 4th ICSAI, Nov 2017.
[18] Mirto Musci et al. Online fall detection using recurrent neural networks.
arXiv preprint arXiv:1804.04976, 2018.
[19] T. N. Gia and M. Jiang. Exploiting fog computing in health monitoring.
Fog and Edge Computing: Principles and Paradigms, 2019.
[20] M. Ali et al. Autonomous patient/home health monitoring powered by
energy harvesting. In IEEE GLOBECOM, pages 1–7. IEEE, 2017.
[21] T. N. Gia et al. Energy efficient fog-assisted iot system for monitoring
diabetic patients with cardiovascular disease. Future Generation Com-
puter Systems, 93:198–211, 2019.
[22] T. N. Gia et al. Fog computing approach for mobility support in internet-
of-things systems. IEEE Access, 6:36064–36082, 2018.
[23] H. Sak et al. Long short-term memory recurrent neural network
architectures for large scale acoustic modeling. In Fifteenth annual
conference of the international speech communication association, 2014.
[24] M. Abadi et al. Tensorflow: A system for large-scale machine learning.
In 12th {USENIX}Symposium on Operating Systems Design and
Implementation ({OSDI}16), pages 265–283, 2016.
[25] G. Vavoulas et al. The mobiact dataset: Recognition of activities of
daily living using smartphones. pages 143–151, 01 2016.