Biomedical Signal Processing and Control 92 (2024) 105963

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

journal homepage: www.elsevier.com/locate/bspc

Enhanced decision-making in healthcare cloud-edge networks using deep

reinforcement and lion optimization algorithm
S.S. Saranya a, *, Palagati Anusha b, S. Chandragandhi c, O. Kiran Kishore d, Nakka Phani Kumar b,
K. Srihari b
a
Department of Computing Technologies, School of Computing, SRM Institute of Science and Technology, Kattankulathur campus. Chennai, 603203, India
b
CSE, SNSCT, India
c
AP / AI and DS Karpagam institute of technology coimbatore, India
d
Dept of CSE, SNS College of Technology, Coimbatoreo, India

A R T I C L E I N F O A B S T R A C T

Keywords: Addressing the limitations of traffic-centric approaches in cooperative cloud-edge networks, this paper in­
Artificial intelligence troduces an adaptive deployment strategy for FDNN using LOA. However, the oversight of network structure
Neural networks hampers load-balancing efficiency in medical data categorization. To overcome this, a novel DRL-LOA approach
Fuzzy deep neural network
integrates Deep Reinforcement Learning (DRL) with LOA, considering both network structure and traffic-related
Evolutionary algorithm
Soft computing
data. This DRL-LOA considers both network structure and traffic-related data for optimization tasks. The DRL
Machine learning employs the Using Graph Convolutional Network (GCN) to extract the network architecture data, the node vector
is created by fusing it with traffic-related features. The deep Q-network uses this node vector to anticipate the
best rewards and determine which FDNN task is carried out at the edge nodes. In light of this choice, the FDNN
classifier is deployed to categorize the medical information. In the end, the simulation findings show that the
DRL-LOA significantly enhances optimization, showcasing a 35.7% reduction in power usage and a of 27.9% %
decrease in latency on 20 edge nodes in cloud-edge systems. These findings underscore DRL-LOA’s efficacy in
optimizing medical data categorization and addressing load-balancing challenges in cooperative cloud-edge
networks.

1. Introduction techniques by boosting the ability to control contemporary ecosystems

and spread a variety of novel resources in an adaptive manner. From this
Generally, IoT is a model for constructing and establishing web ac­ perspective, IoT-based cloud/edge computing can be broadened for the
tivities using modern technologies. It suggests that having a lot of less development of novel services in healthcare devices [3]. In fact, these
powerful devices, like wearable sensors, is preferable to having few systems are beneficial in providing powerful devices to monitor and
powerful smart devices, like laptops, cellphones, and so forth, electric retrieve the patient’s health-related reports from a remote distance. IoT-
appliances and so on. Recurring physical systems such as automobiles based clinical technologies have been utilized to accumulate critical
have been wisely designed using microprocessors accompanied by sen­ data such as recurring changes in physiological state measures over time
sors to provide desired outcomes [1]. As a result, the connected devices and to update the seriousness of illness over time. It helps physicians to
facilitate computation and interface knowledge over the demands of recognize a variety of illnesses at the beginning phases [4].
concentrated solutions. In the past few years, IoT technologies have been Machine learning models offer an important function in deciding the
encouraged to integrate with cloud-edge computing systems for moni­ vast array of samples. These models include neural network methods,
toring patients in various regions. grouping and classification methods [5]. In IoT, there have been a vast
IoT techniques are utilized in cooperation with the cloud to improve number of sources to create the crucial samples and so designing a
performance in terms of memory, cost, processing speed, and energy powerful data processing system has been complex. To avoid this
efficiency [2].Additionally, cloud/edge computing gains from IoT problem, a Context-aware Data Fusion Technique (CDFT) [6] was

* Corresponding author.
E-mail addresses: [email protected] (S.S. Saranya), [email protected] (P. Anusha), [email protected] (S. Chandragandhi), o.
[email protected] (O. Kiran Kishore), [email protected] (N. Phani Kumar), [email protected] (K. Srihari).

https://doi.org/10.1016/j.bspc.2024.105963
Received 1 December 2023; Received in revised form 5 January 2024; Accepted 27 January 2024
Available online 16 February 2024
1746-8094/© 2024 Elsevier Ltd. All rights reserved.
S.S. Saranya et al. Biomedical Signal Processing and Control 92 (2024) 105963

recommended which accumulates and pre-processes the samples from suggestions for further work.
IoT devices. In preprocessing, the edge-based outlier subtraction scheme
has been employed for labeling the unlabeled attributes in the accu­ 2. Literature survey
mulated samples. Then, the samples from multiple IoT devices have
been merged using the CDFT according to the context. Further, the Deep A concept for a hybrid fog-cloud computing medical decision-
Neural Network (DNN) structure was applied to categorize the samples support system that protects privacy was created by Liu et al. [11]. In
to recognize and treat a variety of illnesses. this model, a lightweight data mining scheme was used by the fog server
The integration of DNNs into the cloud computing systems highlights to steadily forecast patients’ medical status in the real world. The
their deployment in various real-world applications. It covers DNN identified irregular signs were broadcast for the duty of prediction, to
complexities across architectures, emphasizing the necessity and ad­ the cloud server. In order to achieve a safe single-layer neural network
vantages of utilizing cloud computing’s high-speed, high-performance for the fog servers, a unique secure outsourcing inner-product method
capabilities. Additionally, it reviews existing DNN applications in cloud was created. Additionally, any activation functions were carried out via
systems and addresses the challenges while offering guidance for privacy-preserving piecewise polynomials computation in a multilayer
improved and new deployments [7]. Though it merges the samples, it neural network. Furthermore, a privacy-preserving rate estimation al­
causes high overhead due to more features. So, an Improved IPCA has gorithm was applied to reduce the computation cost. But, its decision
been applied which mines the relevant features and reduces the ratio was not effective since it was not able to handle highly sophisti­
dimensionality of samples for classification tasks [8]. Conversely, it was cated functions.
not effective in edge devices since it takes more processing time. As a Masood et al. [12] designed a new computer-aided decision support
result, an autoencoder-based lightweight DNN was applied as the model for identifying the pulmonary nodule using a 3D Deep Convolu­
dimensionality reduction scheme. The DNN encoder has been equipped tional Neural Network (DCNN). In this model, a new median intensity
at the edge to decrease the feature dimensions. Such features have been projection was employed for leveraging 3D data from CT images and a
forwarded to the cloud for categorization purposes. Still, the DNN new multi-region proposal network was developed for choosing the
training was not appropriate for optimizing task allocation at the edge possible Region-Of-Interest (ROI). Also, this model was incorporated
devices. into the cloud system to increase efficiency. But, its accuracy was
This Reinforcement Learning (RL) application within Intelligent- comparatively not effective for recognizing micro nodules.
healthcare (I-health) systems, exploring its potential across edge intel­ A smart electronic gastroscope model reliant on the cloud-edge
ligence, core networks, and treatment strategies. It aims to understand cooperation system was presented by Ding et al. [13]. To identify the
RL’s role in transforming healthcare systems towards intelligent and lesion or locate abnormal frames in this model, a Tinier-YOLO method
cost-effective solutions [9]. Thus, a collaborative cloud-edge computing was created based on the k-DSC units in the edge computing system. To
model has been developed for implementing the FDNN classifier into improve the modeling efficiency, the YOLOv3 algorithm used in the
edge devices [10]. The main objective of this work is to enhance load cloud system was then integrated with the lesion ROI division. On the
balancing and optimize task deployment within collaborative cloud- other hand, the choice of threshold values utilized for calculating the
edge computing frameworks, particularly in healthcare data categori­ similarity determines how accurate it is.
zation. By incorporating both traffic-related factors and crucial network Abdel-Basset et al. [14] presented a new model using computer-
structure information, the aim is to refine the load-balancing strategy propped prognosis and IoT for recognizing type-2 diabetes patients.
during the classification process, improving overall performance in Initially, a diagnostic fog-based IoT platform was executed to forecast
handling healthcare data within edge servers. The focus lies in type-2 diabetes patients. Also, a novel decision-making system was
leveraging this dynamic deployment scheme, integrating LOA, to make designed for type-2 neutrosophic numbers with the VIKOR scheme. But,
informed decisions about task allocation in edge servers based on it needs to design deep learning models to enhance accuracy.
comprehensive considerations of network structure and traffic-related Bhatia et al. [15] designed a potent model for home-centric urine-
factors. based diabetes surveillance. Four distinct processes were carried out in
However, these methods’ technical intricacies are critical to establish this model to identify the diabetes-oriented urine virus: diabetic data
a foundational understanding for the proposed work. Describing CDFT’s collection, recognition, diabetic mining, and decision-making. In addi­
outlier subtraction technique, IPCA’s feature mining process, and the tion, the Recurrent Neural Network (RNN) carried out a probability test
architecture and training specifics of the FNN classifier would provide of urine-based diabetes forecasting based on the degree of diabetic
necessary groundwork, facilitating a comprehensive explanation of the sickness. Furthermore, the self-organized map task conjectured the
proposed integration of DRL and LOA for optimizing task allocation in possibility of urine-based diabetes. But, its robustness was less if the
edge nodes. Incorporating these details would better elucidate the sig­ number of samples was high.
nificance of integrating DRL and LOA, illustrating how they augment the Tuli et al. [16] constructed an enhanced system known as HealthFog,
capabilities of the existing methods, specifically addressing the limita­ which applies the deep learner in edge devices to predict heart syn­
tions encountered previously in edge-based processing and task alloca­ dromes automatically. According to this system, the medical services
tion optimization. were served as the fog service with the help of IoT networks and the
Therefore, in this article, integration of the DRL and LOA is proposed heart patients’ details were managed. Afterward, a fog-assisted cloud
to choose which duty the edge nodes will carry out.This newly designed system was created to assess the prediction efficiency. Conversely, its
DRL-LOA extracts and optimizes both network structure and traffic- robustness was not effective because it was not able to achieve cost-
related information during optimization. The network structure infor­ optimized execution.
mation is mined by applying the GCN and it is combined to produce a Wang et al. [17] developed a computing resource estimation method
single node vector using the traffic-related data. The deep Q-network using intelligent ensemble learning. In this method, accuracy and rela­
uses this node vector to determine which FDNN job is carried out at the tive error-based pruning strategy were utilized for guaranteeing the
edge nodes and to forecast the best rewards. In light of this ruling, the effective ensemble of base classifiers such as SVM, decision tree and K-
FDNN classification is executed for categorizing healthcare records. nearest neighbor. Besides, a feature capability-based forward search
The remainder of the document is ready as follows: The research on feature choice method was applied to eliminate repetitive attributes.
the IoT-cloud concept in clinical usage is covered in Section II. The DRL- But, it was not suitable for cloud-edge systems under dynamic scenarios.
LOA-based deployed FDNN classifier for health-related data classifica­ Also, it needs to decide on optimum resources to enhance the large-scale
tion is described in Section III, and its testing effectiveness is shown in cloud-edge paradigm.
Section IV. This study is summarized in Section V, along with Shekar Goud, D et al. [18] builds a aims to simplify medical device

2
S.S. Saranya et al. Biomedical Signal Processing and Control 92 (2024) 105963

usage via an IoT-based solution, prioritizing affordability and accessi­ robustness of existing models. Motivated by these shortcomings, the
bility to therapy for patients. It utilizes Elephant Herding Optimizations authors aimed to bridge the divide by integrating deep reinforcement
with Convolutional Neural Networks (CNNs) to distinguish between learning (DRL) and Lion Optimization Algorithm (LOA), leveraging both
healthy and unhealthy behavior patterns. Using a scoring function like network structure and traffic-related information for enhanced decision-
fuzzy logic, the severity of irregularities is determined, targeting easier making in cloud-edge healthcare systems. This fusion aimed to address
home-based monitoring at a reasonable cost. the limitations encountered in previous studies, particularly in opti­
He, Y et al [19] introduces Edge CNN, hybrid edge-cloud architecture mizing task allocation, ensuring robustness in dynamic environments,
for agile healthcare data learning. It develops a deep learning model for and improving the accuracy and efficiency of medical data categoriza­
ECG inference, optimizing for low-latency diagnosis on edge devices. tion processes.
Additionally, it employs a data enhancement technique using deep
convolutional generative adversarial networks to expand ECG data. 3. Proposed methodology
Experimental results showcase Edge CNN’s superiority in I/O pressure,
architecture cost, and diagnostic accuracy compared to traditional cloud In this section, the proposed DRL-LOA-based decision-making model
systems. for medical data categorization in cloud-edge computing systems is
Gu, M et al [20] uses transfer learning and data augmentation ap­ described in brief. The schematic overview of this decision-making
proaches to provide a cloud-edge computing model for intelligent skin system is depicted in Fig. 1. Initially, the patient’s health-related data
lesion diagnosis. It attempts to increase classification accuracy in the is monitored by the IoT devices and pre-processed by the CDFT to merge
face of unbalanced data and a small number of training examples. To them. Such fused data is transmitted by way of edge nodes to the cloud
overcome these difficulties, a modified cross-entropy loss and model servers. Next, in order to compute different network structure data as
fine-tuning are used in the suggested approach. The approach’s higher well as traffic-related information, a cloud server and edge nodes work
performance over current methods is demonstrated by the experimental together. The network structure information is extracted by the GCN,
results, indicating major breakthroughs in accurate skin lesion whereas the traffic-related information is extracted by the LOA. After
diagnosis. that, all this information is combined together to obtain the node vector.
The paper [21] proposes leveraging mobile edge computing for This node vector is trained by the deep Q-network to decide the nodes
efficient sensory data gathering from social robots within IoT networks. (either cloud or edge nodes) to execute the task of the FDNN classifier.
This method aims to enhance energy efficiency, prolong network life­ As a result, the health information is categorized to recognize a variety
time, and facilitate informed decision-making by enabling edge nodes to of illnesses appropriately.
periodically collect sensory data from their associated social robots.
The study [22] proposes a novel architecture for IoT, aiming to
enhance job offloading and resource allocation. It employs a subtask 3.1. Preliminaries
pool strategy for job offloading and integrates the Moth-Flame Optimi­
zation (MFO) algorithm with Opposition-based Learning (OBL) as Let the cloud-edge network be a directed tree-like graph G = (M, E)
OBLMFO to distribute resources efficiently. Additionally, it leverages with every vertex m ∈ M being edge nodes wherein the FDNN classifier
blockchain to ensure transaction data accuracy, optimizing resource will be executed and every edge e ∈ E defines the connection between
distribution in the IoT landscape. nodes. The FDNN classifier is executed on any node in G. In this system,
In this study [23] introduces HPSOSSA, a hybrid optimization al­ the FDNN classifier is represented as a 3-tuple c = (a, l, d) which in­
gorithm merging PSO and the SSA. This method aims to enhance task dicates the efficiency of classification. The variable a(l) ∈ [0, 1] is the
scheduling efficiency in healthcare services reliant on cloud computing ratio of features that the classifier can learn. Also, Om = 〈oi 〉 represents
within the IoT framework, addressing critical issues influencing request the group of choices on node m where oi ∈ {0, 1} and oi = 1 denotes that
fulfilment time, service quality, and cost-effectiveness. the node m executes ith task of classifier. This decision-making step can
The paper [24] introduces an Artificial Neural-inspired Whale be formulated as the optimization dilemma.
Optimization method to address challenges in e-healthcare services Initiating from a null set S = ∅ (both cloud and edge nodes will not
reliant on cloud computing. It aims to optimize real-time computational execute the FDNN), the execution of ith task of classifier on the optimum
tasks with deadlines, enhancing task completion within the stipulated node m* is iteratively scheduled. To reduce the computation burden of
time frame. Specifically, the study presents two models: one for reli­ the decision-making process, the DRL model is developed which may
ability estimation and the other employing a whale optimization tech­ train network structure information of the graph. This information is
nique alongside a neural network-based binary classifier for healthcare combined with traffic-related information to create the node vector. For
application solutions. classifier execution, this node vector is input into the deep Q-network in
The study [25] aims to improve management structures in health­ order to acquire the Q-value.
care, especially in light of the coronavirus pandemic, by incorporating
IoT-cloud technology into a new schema for hospital information sys­
tems (HIS). In order to address issues and take advantage of opportu­ 3.2. Network embedding
nities in the contemporary healthcare system, it develops sophisticated
models for the database administration, data processing, and notifica­ Because of modifying G with m number of nodes, it is anticipated that
tion components inside the HIS utilizing the DEMATEL protocol. the Q-factor must consider the network structure and the present partial
The paper [26] introduces an Enhanced African Vultures Optimiza­ solution S for optimizing the number of variables [8]to execute the
∑
tion Algorithm-based Task Scheduling Strategy (E-AVOA-TS) tailored FDNN classifier. Therefore, oi ∈Om oi = 1 for every node m ∈ S where
for fog-cloud computing. This method aims to optimize task scheduling each node is connected in accordance with the graph structure. By using
by minimizing makespan, cost, and energy consumption, addressing this, while deciding the classification execution in the node m, both
issues related to resource utilization, response time, and energy effi­ network structure and traffic-related information are essential.The GCN
ciency. It incorporates a strategy where each village learns from its model is utilized for embedding G.
neighbors, employing the Best Worst Method (BWM) to prioritize tasks This network embedding model can determine a p-dimensional
and mitigate the sensitivity of task delays. feature embedding μm for all nodes m ∈ M. Particularly, a recursive
The identified research gaps within the literature review served as scheme is utilized for interpretation of the nodes. In all iteration of
catalysts, compelling the authors to define their current objectives. computation, the node m accepts data from its children nodes and
These gaps highlighted limitations in accuracy, scalability, and computes the recent μm as:

3
S.S. Saranya et al. Biomedical Signal Processing and Control 92 (2024) 105963

Fig. 1. Overview of the Presented Decision-Making Model in Cloud-Edge Systems for Medical Data Categorization.

( )
{ } 3.4. Deep Q-leaning network
μ(t+1)
m ←F Om , μ(t)
l ;Θ (1)
l∈Dm
In this framework, the reinforcement training i.e., deep Q-learning is
In Eq. (1), it updates the node embedding μm in each iteration t + 1.
utilized for training the optimal set of network structure with different
Dm denotes the collection of children nodes for m in network N and F
traffic-related data. Deep Q-Networks offer significant advantages in
defines the nonlinear function. The primary embedding μl at every
(t)
reinforcement learning scenarios. They efficiently handle complex
node is zero. decision-making tasks by learning optimal policies, combining neural
It can be shown from Eq. (1) that the structure of the network and the networks with Q-learning. It utilize experience replay, enhancing
embedding in the previous stage are necessary for the node embedding learning by reusing past experiences, ensuring better sample efficiency.
alteration task. A node can only access data through its offspring nodes. Their ability to approximate Q-values for state-action pairs leads to
The node will only receive all of the structural information of a sub-tree stable learning, tackling high-dimensional state spaces effectively.
that is rooted at it if there are sufficient iterations. Moreover, it employ target networks, mitigating Q-value estimation
The function F is designed to modify p-dimensional embedding μm as: errors and ensuring smoother convergence during training. These ben­
( ) efits collectively enable it to navigate complex environments, making
∑ (t)
μ(t+1)
m ←ReLU θ O
1 m + θ 2 μ l (2) them valuable for various applications requiring efficient decision-
l∈Dm making under uncertainty. The states, actions and rewards in this Q-
p×|Om | learning network are described the following:
In Eq. (2), θ1 ∈ R and θ2 ∈ Rp×p are the model variables and
ReLU is the Rectified Linear unit (ReLu(x) = max(0, x) ) utilized element-
• State: A state S is a collection of present selections of each node.
wise to its input. The 0–1 vector Om is the decision of the node m.
Because this embedding μm has included the decision, the state is
∑
represented as a p-dimensional space m∈M μm and this interpreta­
3.3. Parameterization of Q-Factor tion is applied in various tree structures.
• Actions: The action (m, i) is a 2-tuple where m is a node of G without
Once the embedding is completed, these network structure-related executing the FDNN classifier and i is the task of FDNN. The actions
features are integrated with the different variables related to the en­ are characterized as their respective p-dimensional node embedding
ergy use, latency and processing ability optimized by the LOA for μm . Because of the amount of tasks of FDNN is computed, the
determining the Q-factor. In fact, μm is utilized for m and the addition of outcome of Q-factor is the collection of Q-value of every task of
∑
each npde embedding m∈M μm is the partial solution p(S). To determine FDNN.
the Q-value of FDNN classifier of every node, the overall processing • Transition: The transition is predetermined. For a thoughtful
latency (T QP ), edge-to-cloud transfer latency (T EC ), cloud-to-edge response (m, i), the node m is assigned to execute the ith task of FDNN
( )
uplink transfer latency (T CE ), static energy use ESi,k of mj in task xk , classifier.
( ) ( ) • Rewards: This algorithm trains the DRL-LOA, therefore the reward
computational power use ECom i,k and the maximum latency T high . factor r(S, m, i) that the action (m, i) is considered at state S is
Therefore, the Q-factor Q(p(S), m; Θ ) is defined as: described as:
[ ([
∑
]) r(S, m, i) = r(S ∪ (m, i) ) − r(S) (4)
S Com
Q(p(S), m; Θ ) = θ3 ReLu θ4 μl , θ5 μm , T QP , T EC , T CE , Ei,k , Ei,k
l∈M
Also, r(∅) = αr1 (S).
]
,T high • Policy: According to the Q-factor, a deterministic greedy policy
π((m, i)|S) = argminQ(S, m, i) is utilized. Selecting action (m, i) cor­
(3) m∕
∈S,i
responds to inserting a node of G to the present partial solution and
In Eq. (3), θ3 ∈ R|Om |×(2p+2) and θ4 , θ5 ∈ Rp×p are the model parame­ obtaining a reward r(S, m, i).
ters, [•, •] is the fusion operator. This outcome of Eq. (3) is the
|Om |-dimensional vector which is the Q-value of execution of FDNN on An n-step Q-learning technique is utilized for training the approxi­
node m. It is complex to learn variables Θ = {θi }5i=1 because of inade­ mate optimum solution under different traffic-related information. To
quate training labels. Thus, these variables are learned by the rein­ conduct an end-to-end training of variables in Q(S, m, i; Θ), a mixture of
forcement training.

4
S.S. Saranya et al. Biomedical Signal Processing and Control 92 (2024) 105963

Fig. 2. Flowchart for Decision-Making System.

deep Q-network as presented in Fig. 2 and Algorithm 1. To decide a node r(S, m, i) +γminQ(S ∪ (m, i), m, i; Θ2 ) for a state S excluding the end state
for executing the task of FDNN and obtain an optimal reward, the
(m,i)

S′, Θ2 denotes the variable of the target network, it is fine-tuned at the

combinational dilemma is formulated as an n-step sequence decision
starting of each episode by the evaluation network, i.e.Θ2 = Θ1 . To
challenge. To resolve this problem, the measures are trained for the
avoid the correlation among the training data, the experience replay is
optimum range of every action, expressed as the anticipated total of
employed for modifying the function approximate with a batch of in­
likely rewards if the activity and the best course of action are taken into
stances from the dataset. The storage δ is occupied during prior step.
consideration.
Rather than conducting Stochastic gradient fine-tuning is performed on
(S* , (m* , i* ), r* , S* ∪ (m* , i* ) ) fromδ a random sample of tuples derived from δ, resulting in a gradient step in
error of the current result.
For a considered network structure, X various inputs factors (pro­
Algorithm: Deep Q-learning for DRL-LOA-based Decision-Making System
cessing ability, latency and energy use determined in [8]) are considered
for learning. An episode is utilized for denoting the entire sequence of Begin
Initialize experience replay storage δ;
executing FDNN task from a null state S to the end state S′. Also, decision Initialize random variables Θ1 and Θ2 = Θ;
procedure. All steps, deep Q-network modifies the function approximate Input the network structure G and traffic-related information;
variables via achieving a gradient step for reducing the squared error: f or(time < K)//K: Computation time
Set the data in network G;
(y − Q(S ∪ (m* , i* ), m* , i* ; Θ1 ) )2 (5) f or(episode < X)
Initialize the state S = ∅;
In Eq. (5), Θ1 denotes the variable of the evaluation network and y = (continued on next page)

5
S.S. Saranya et al. Biomedical Signal Processing and Control 92 (2024) 105963

(continued ) task of the FDNN classifier.

Algorithm: Deep Q-learning for DRL-LOA-based Decision-Making System The comparison among the different task execution decision-making
Θ2 = Θ1 ;
models reveals distinct trends in power usage concerning varying edge
f or(step < T)//T: number of iterations node counts is illustrated in Fig. 3. The LOA model shows a consistent
if (random(0, 1) < ∊ ) increase in power usage as the number of nodes escalates, starting at
(m, i) = argminQ(S, m, i; Θ1 ); 48.6% at 5 nodes and peaking at 37.8% with 25 nodes. In contrast, the
m∕
∈S
else UDTC model displays a continual decrease in power usage, beginning at
Random edge node m ∕ ∈ S; 55.3% and decreasing to 35.1% across the same node range. Similarly,
Random task i of FDNN classifier; the LEACH model also portrays a diminishing trend in power usage from
53.3% at 5 nodes to 40.1% at 25 nodes. Notably, the DRL-LOA2 model
endif
r = r(S ∪ (m, i) ) − r(S);
S = S ∪ (m, i); exhibits the most substantial reduction in power consumption,
Include tuple (S, (m, i), r, S ∪ (m, i) ) to δ; commencing at 40.3% and plunging notably to 23.6% with 25 nodes,
if (data ≥ N)//N: Number of input data indicating its superior efficiency in optimizing power utilization across
Sample random batch; all node configurations. Overall, the analysis underscores DRL-LOA2′s
(S* , (m* , i* ), r* , S* ∪ (m* , i* ) )from δ;
significant efficacy in minimizing power usage in comparison to the
if (step == (T − 1) )
y = r* ; other models, particularly evident increases. It achieves a minimization
else of 35.7% in power usage than the LOA for 25 edge nodes in a system,
y = r* + γ min Q(S* ∪ (m* , i* ), m* , i* ; Θ2 ); showcasing the efficacy of DRL-LOA in optimizing power consumption
(m* ,i* )
endif for load-balancing in cloud-edge networks.
Fine-tune Θ1 by stochastic gradient descent over Eq. (5);
endif
endf or
4.2. Latency
endf or
endf or It is the time considered to perfectly execute the task of the FDNN
Obtain Θ1 ; classifier by the edge nodes.
Cloud selects which task of FDNN classifier will be conducted at the edge nodes;
The latency comparison between various task execution decision-
End
making models illustrates distinctive performance variations concern­
ing different edge node counts are shown in Fig. 4. At 5 nodes, the LOA
Thus, by training the deep Q-network, the optimum reward is obtained model starts with 57.9% latency, while the DRL-LOA2 initiates at 48.5%.
to decide which task of FDNN which FDNN task will be performed at the As the number of nodes increases, the LOA model shows an increasing
servers in the cloud and which task will be performed at the edge node. trend in latency, reaching 17.8% at 25 nodes. In contrast, DRL-LOA2
consistently exhibits notably lower latency across all node configura­
4. Simulation results tions, starting at 48.5% and reducing significantly to 12.8% at 25 nodes.
Similarly, UDTC and LEACH models follow intermediate patterns, with
This section assesses the effectiveness of task execution decision- decreasing latency trends but higher values compared to DRL-LOA2.
making in cloud-edge systems for load balancing utilizing DRL-LOA Notably, DRL-LOA2 consistently outperforms other models, show­
[8] based on latency and power consumption.. The testing is conduct­ casing superior efficiency in reducing latency, with the most significant
ed in Javato assess the performance of DRL-LOA on cooperative cloud- reduction observed at 25 nodes, emphasizing its effectiveness in
edge systems. To simulate this process in the IoT, 25 edge nodes are executing FDNN tasks in cloud-edge systems while maintaining signifi­
built and the testing is conducted with the help of the Melanoma 384 × cantly lower latency rates than other models.
384 dataset. It is available on https://www.kaggle.com/datasets/cde
otte/jpeg-melanoma-384x384?select = test.csv. 5. Conclusion

4.1. Power usage This study presents a DRL-LOA-based framework designed to opti­
mize task allocation of the FDNN classifier within cloud-edge networks
It is the amount of power utilized by the edge nodes to execute the by leveraging network structure and traffic-related data. The

Fig. 3. Power Use vs. No. of Edge Nodes.

6
S.S. Saranya et al. Biomedical Signal Processing and Control 92 (2024) 105963

Fig. 4. Latency vs. No. of Edge Nodes.

methodology integrates extracted data into a unified node vector, Data availability
allowing a deep Q-network to make informed decisions regarding
optimal task execution at edge nodes. The subsequent deployment of the Data sharing not applicable to this article as no datasets were
FDNN classifier for medical data categorization is guided by these de­ generated or collected, but Melanoma dataset were analysed during the
cisions. The study’s outcomes demonstrate the significant advantages of current study.
DRL-LOA over traditional algorithms, exhibiting a noteworthy 35.7%
reduction in power usage and a 27.9% decrease in latency, particularly Acknowledgment
evident in scenarios involving 20 edge nodes within the cloud-edge
system. However, while showcasing efficacy in power and latency There is no acknowledgment involved in this work.
reduction, this research is limited in scope, primarily emphasizing these
metrics and potentially overlooking other essential performance mea­
References
sures crucial in intricate medical data processing scenarios. Future in­
vestigations should broaden evaluation criteria, encompassing a more [1] M. Aboubakar, M. Kellil, P. Roux, A review of IoT network management: current
comprehensive range of performance metrics beyond power and la­ status and perspectives, J. King Saud Univ.-Comput. Informat. Sci. (2021) 1–14.
tency. Additionally, exploring scalability aspects and accommodating [2] P.K. Malik, R. Sharma, R. Singh, A. Gehlot, S.C. Satapathy, W.S. Alnumay,
J. Nayak, Industrial internet of things and its applications in industry 4.0: state of
diverse complexities of FDNN classifiers could enhance the applicability the art, Comput. Commun. (2020) 1–18.
and robustness of the proposed DRL-LOA methodology in real-world [3] L. Greco, G. Percannella, P. Ritrovato, F. Tortorella, M. Vento, Trends in IoT based
medical data processing scenarios. In conclusion, this study un­ solutions for health care: moving AI to the edge, Pattern Recogn. Lett. 135 (2020)
346–353.
derscores the promising potential of DRL-LOA in optimizing task allo­ [4] S. Cao, X. Lin, K. Hu, L. Wang, W. Li, M. Wang, Y. Le, Cloud Computing-Based
cation within cloud-edge systems, prompting further exploration to Medical Health Monitoring IoT System Design, Mob. Inf. Syst. 2021 (2021) 1–12.
address diverse challenges encountered in medical data processing. [5] L. Cui, S. Yang, F. Chen, Z. Ming, N. Lu, J. Qin, A survey on application of machine
learning for internet of things, Int. J. Mach. Learn. Cybern. 9 (8) (2018)
Ethics Approval and Consent to Participate 1399–1417.
No participation of humans takes place in this implementation [6] S.S. Saranya, N.S. Fatima, Context aware data fusion on massive IOT data in
process. dynamic IOT analytics, Webology 17 (2) (2020) 957–970.
[7] K.Y. Chan, B. Abu-Salih, R. Qaddoura, A.Z. Ala’M, V. Palade, D.S. Pham,
Human and Animal Rights
K. Muhammad, Deep Neural Networks in the Cloud: Review, Applications,
No violation of Human and Animal Rights is involved. Challenges and Research Directions, Neurocomputing 126327 (2023).
Funding [8] S.S. Saranya, N.S. Fatima, IoT information status using data fusion and feature
extraction method, Comput. Mater. Continua (2021) 1–18.
No funding is involved in this work.
[9] A.A. Abdellatif, N. Mhaisen, Z. Chkirbene, A. Mohamed, A. Erbad, M. Guizani,
Reinforcement learning for intelligent healthcare systems: A comprehensive
CRediT authorship contribution statement survey, 2021. arXiv preprint arXiv:2108.04087.
[10] S.S. Saranya, N.S. Fatima, Load balancing of medical data classification in cloud-
edge network using optimization algorithm, J. Comput. Sci. (2021).
S.S. Saranya: Conceptualization. Palagati Anusha: Writing – orig­ [11] X. Liu, R.H. Deng, Y. Yang, H.N. Tran, S. Zhong, Hybrid privacy-preserving clinical
inal draft. S. Chandragandhi: . O. Kiran Kishore: . Nakka Phani decision support system in fog–cloud computing, Futur. Gener. Comput. Syst. 78
Kumar: . K. Srihari: Supervision. (2018) 825–837.
[12] A. Masood, P. Yang, B. Sheng, H. Li, P. Li, J. Qin, D.D. Feng, Cloud-based
automated clinical decision support system for detection and diagnosis of lung
Declaration of Competing Interest cancer in chest CT, IEEE J. Translat. Eng. Health Med. 8 (2019) 1–13.
[13] S. Ding, L. Li, Z. Li, H. Wang, Y. Zhang, Smart electronic gastroscope system using a
cloud–edge collaborative framework, Futur. Gener. Comput. Syst. 100 (2019)
The authors declare that they have no known competing financial 395–407.
interests or personal relationships that could have appeared to influence [14] M. Abdel-Basset, G. Manogaran, A. Gamal, V. Chang, A novel intelligent medical
the work reported in this paper. decision support model based on soft computing and IoT, IEEE Internet Things J. 7
(5) (2019) 4160–4170.
[15] M. Bhatia, S. Kaur, S.K. Sood, V. Behal, Internet of things-inspired healthcare
system for urine-based diabetes prediction, Artif. Intell. Med. 107 (2020) 1–16.
[16] S. Tuli, N. Basumatary, S.S. Gill, M. Kahani, R.C. Arya, G.S. Wander, R. Buyya,
HealthFog: an ensemble deep learning based smart healthcare system for

7
S.S. Saranya et al. Biomedical Signal Processing and Control 92 (2024) 105963

automatic diagnosis of heart diseases in integrated IoT and fog computing [22] M. Nematollahi, A. Ghaffari, A. Mirzaei, Task and resource allocation in the
environments, Futur. Gener. Comput. Syst. 104 (2020) 187–200. internet of things based on an improved version of the moth-flame optimization
[17] S. Wang, F. Zhu, Y. Yao, W. Tang, Y. Xiao, S. Xiong, A computing resources algorithm, Clust. Comput. (2023) 1–23.
prediction approach based on ensemble learning for complex system simulation in [23] K.M. Hassan, A. Abdo, A. Yakoub, Enhancement of health care services based on
cloud environment, Simul. Model. Pract. Theory 107 (2021) 1–17. cloud computing in IoT environment using hybrid swarm intelligence, IEEE Access
[18] D. Shekar Goud, B.B. Beenarani, C. Brijilal Ruban, R. Fathima, M.L. Bharathi, A. 10 (2022) 105877–105886.
Rajaram, et al., Deep learning technique for patients healthcare monitoring using [24] P. Gupta, S. Bhagat, D.K. Saini, A. Kumar, M. Alahmadi, P.C. Sharma, Hybrid whale
IoT body based body sensors and edge servers. J. Intelligent Fuzzy Systems, optimization algorithm for resource optimization in Cloud E-healthcare
(Preprint), 1-15. applications, Comput. Mater. Continua 71 (3) (2022).
[19] Y. He, B. Fu, J. Yu, R. Li, R. Jiang, Efficient learning of healthcare data from IoT [25] O. Nguyen, HSSCIoT: An optimal framework based on internet of things-cloud
devices by edge convolution neural networks, Appl. Sci. 10 (24) (2020) 8934. computing for healthcare services selection in smart hospitals, Adv. Eng.
[20] M. Gu, X. Li, S. Chen, Transfer learning-based intelligent diagnosis for skin lesion in Intelligence Syst. 1 (02) (2022).
cloud-edge computing networks, in: 2021 13th International Conference on [26] R. Ghafari, N. Mansouri, E-AVOA-TS: Enhanced African vultures optimization
Wireless Communications and Signal Processing (WCSP) (pp. 1-6). IEEE, 2021. algorithm-based task scheduling strategy for fog–cloud computing, Sustainable
[21] D. Ren, X. Li, Z. Zhou, Energy-efficient sensory data gathering in IoT networks with Comput. Inf. Syst. 40 (2023) 100918.
mobile edge computing, Peer-to-Peer Network. Appl. 14 (2021) 3959–3970.