Добавил:
Sekretar
kiopkiopkiop18@yandex.ru
t.me/Prokururor I Вовсе не секретарь, но почту проверяю
Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз:
Предмет:
Файл:Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_5234_Библиотеки_им_академика_М_И_Перельмана
.pdf
294 Computational Intelligence Algorithms
19.3 CHALLENGES IN DIGITAL HEALTHCARE
IMPLEMENTATION FOR AD
Providing solutions for AD with digital healthcare systems undergoes substantial
difculties like privacy of data, security and integrity of the systems, validations,
ethical concerns, and cost. The author Pyrrho et al. [25] stated that to protect the
sensitive data of AD patients, there is a need for securing electronic health records
and digital health platforms. Merging previous healthcare systems also faces challenges; for example, many digital devices do not rely on outdated digital platforms
that cannot be integrated easily with modern devices and systems. These complexities in achieving effortless integration were discussed by Herrmann.T et al. [26] and
have highlighted the necessity of benchmarked protocols for improving interoperability. As stated by Doll et al. [27], stakeholders have to collaborate collectively and
follow to common data standards in order to achieve interoperability. Usability and
accessibility are essential, particularly for AD patients who may experience cognitive impairment. User-centered design principles may enhance the effectiveness of
digital health interventions for AD, according to a study by Dabbs et al. [28]. Grande
et al. [29] discussed the importance of consistent standards along with the complexity of the standardized environment around digital health technologies.
19.3.1 COMPUTATIONAL COMPLEXITY
To provide digital healthcare solutions for AD, dealing with computational complexity is a major concern, as it involves in dealing with large amounts of heterogeneous
data that include clinical data, records, image data of scanned reports, genetic biomarkers information, and data from wearable devices and sensors. Liu, Y. et al. [30]
have stated that computational complexity of integrating data from different devices
is a signicant challenge. Zhang et al. [31] have highlighted the challenges and difculties in developing algorithms that deal with high processing power and are cost
efcient.
19.3.2 CHALLENGES IN FOG, CLOUD, AND DEEP LEANING TECHNOLOGIES
DL, fog computing, and cloud computing technologies bring various opportunities in
treating AD. However, they face the following problems for detecting and diagnosing AD.
DL model training: For AD diagnosis and treatment, DL models must be
trained on the data provided. In the words of Li et al. [32], the complexity
and high dimensionality of AD data, like CT scan images and biomarkers
data, require the use of complex neural networks that are computationally
intensive.
Real-time processing: By bringing immediate information processing capabil-
ities near to the information source, fog computing aims to reduce latency.
However, due to their constrained processing capability, DL models are

hard to implement in fog nodes. The difculty of implementing highly
resource-intensive DL algorithms in fog environments without reducing
performance has been brought into focus by Shen et al. [33].
Data Privacy and Security: Providing security for distributed servers in cloud and
fog environments is a challenging task. Kim et al. [34] discussed the signicance of deploying strong encryption techniques and protected communication
protocols to secure patient data from security breaches and illegitimate access.
19.3.3 NETWORK REQUIREMENTS
Low latency requirements: The main goal of fog computing is to minimize
the latency value by processing the data at edge level. Despite that, implementing complicated CNNs is a challenging task with respect to latency
reduction. Kumar et al. [35] has discussed optimization of DL algorithms
for achieving minimum latency.
Bandwidth constraints: Transferring large amounts of data that are gathered
from various sources, including edge computing devices, the fog environment, and the cloud layer, overburdens bandwidth. To establish uninterrupted data transfer and data processing, it is crucial to harness bandwidth
effectively. Guo et al. [36] have highlighted the necessity for effective data
compression and transmission techniques to resolve bandwidth limitations.
295 Optimizing Digital Healthcare for Alzheimer’s Disease
19.3.4 INFRASTRUCTURE REQUIREMENTS
Resource allocation: Efcient resource allocation in distributed environments,
especially in cloud computing and fog computing, plays a major role in
supporting computational need and support for running complex DL algorithms. Chen et al. [37] have discussed various challenges of dynamic
resource allocation and load-balancing techniques to achieve optimal efciency and resource efciency.
Scalability: AD-related data are huge, and to adjust to an increase in the
volume of data, scaling up cloud and fog infrastructure is challenging task.
Li et al. [38] focused on the need for scalable and elastic system designs that
can manage large-scale applications without deteriorating performance. By
carrying real-time data processing abilities near to the data source, fog computing seeks to lower latency. However, because of their constrained processing power, DL models are difcult to implement in fog nodes. The challenge
of implementing resource-intensive DL algorithms in fog environments without sacricing performance is brought to light by Wen et al. [39].
19.4 DEEP LEARNING NETWORKS AND OPTIMIZATION
ALGORITHMS IN ALZHEIMER’S DISEASE
Due to its capacity to assess involved and high-dimensional data, DL networks
emerged as a key element in studying and dealing with AD. Liu et al. [40] stated
that by knowledge of the spatial hierarchies of brain pictures, CNNs have shown

296 Computational Intelligence Algorithms
auspicious results in recognizing early hints of AD, which are needed for detecting
and tracing the disease’s course. Research by Vaswani et al. [41] stated that RNNs can
calculate exactly how AD patients’ cognitive cost will progress, a strength that delivers
signicant evidence for individualized handling regimens. According to Schraudolph
et al. [42], converters can be used to incorporate multimodal data, such as inherent,
imaging, and clinical data, to increase the accuracy of AD diagnosis and prediction.
The optimization of DL models through iteratively changing model constraints to
minimize the loss function is a joint application of gradient descent and its derivatives,
such as stochastic gradient descent (SGD), minibatch gradient descent, and adaptive
techniques of Adam. Snoek et al. [43] emphasized the success of optimization procedures in circumstances with profuse adding resources. Zhang et al. [44] illustrated
how Bayesian optimization can be used in hyper parameter tuning of DL models, with
which signicant improvements can be achieved in accuracy..
19.5 METHODOLOGY
This section illustrates the proposed methodology for detecting AD.
19.5.1 DATASET DESCRIPTION
In 2004, to investigate AD, the ADNI research project was launched. The dataset
consists imaging les, including fMRI, sMRI scans, and PET scans. It has three
classes, namely cognitively normal (CN), MCI, and AD.
19.5.2 DEEP FEDERATED LEARNING CNN SCHEME (DFLCNNS)
This chapter demonstrated the architecture for classifying AD, as shown in
Figure 19.1. The architecture has three different layers, namely the edge com-
puting layer, the fog computing layer and the cloud computing layer. Initially,
the AD patient’s data (blood samples, ECG, and MRI scan) are collected from
various local laboratories such as at individual clinical hospitals. These data are
maintained in an individual electronic health record database for further processing. The entire process of data ofoading is carried out at the edge computing
layer with several diagnostic tools and bio medical instruments used for diagnosis.
As far as the FL is concerned, the collected data must be trained locally with
DFLCNNS at the individual hospitals, which are connected to a centralized cloud
server present at the main hospital that acts as a head for AD detection. The local
training is done at the fog computing layer.
As it is depicted in Figure 19.1, the local clinics that have undergone training
generate weights for the aggregated model, which is trained at the main hospital
that is directly connected to the cloud server. With this local training and aggregation, FL is implemented for achieving optimality by training smaller sample
datasets received at the edge devices. Training models with smaller sample datasets at local processing nodes results in reducing processing time at the aggregated
model when it is dealing with large samples of data. Local training is done at

297 Optimizing Digital Healthcare for Alzheimer’s Disease
FIGURE 19.1 Architecture for Alzheimer’s disease detection using deep federated learning
with optimization.
the fog computing layer, which is near to edge devices, and model aggregation
is carried out at the cloud computing layer. This architecture DFLCNNS uses
an EfcientNetB2 DL model for classication of AD. To achieve better accuracy
and other performance metrics, the architecture uses the Grey Wolf Optimizer
(GWO) optimization algorithm for feature extraction at local training and model
aggregation.

298 Computational Intelligence Algorithms
H
× W
h
× C
l
H
dth
19.5.3 EFFICIENTNETB2 ARCHITECTURE FOR ALZHEIMER’S DISEASE CLASSIFICATION
Implementation and effectiveness are enhanced via the CNN architecture known as
EfcientNetB2. The graphical representation is shown in Figure 19.2. The layers and
functions of the architecture can be expressed quantitatively:
•
Input layer: Receives an picture of size
ght
are denoted by height and width and also lr represents the number color
channels (E.X.
).
• Stem block:
, where
dt
r
ght
and
FIGURE 19.2 Efcient B2 architecture.

Accurately, for a convolution process,
+
c
=
˜
In
put
ker
(
(
)
)
m n
,
,
m
.
ww
2 3
1
,, w
=
=
299 Optimizing Digital Healthcare for Alzheimer’s Disease
where
P In .ker
abc put ,,,
,,
pq,
is the output features map,
ap.bq. d+ +
is the input image, and
abcd
bias
is the convolu-
tion kernel.
19.5.3.1 Mobile Inverted Bottleneck Convolution Blocks
Batch normalization, residual connection, depthwise convolution, and pointwise
convolution are included in every mobile inverted bottleneck convolution block.
With input M, for a mobile inverted bottleneck convolution block,
M = Rectified Linear unitBatch Normalized
Input
DepthwiseConv M
()
˝ Pointwiswe Convoulation ˝ RectifiedLinear unit
19.5.3.2 Global Average Pooling (GAP)
GAP decreases each feature map to a single value:
P = M
a H
1
H × W
ght dth
˜˜
H W
ght1 dth
W
a
ght
dth
1
This is the fully connected layer:
Here, m is the weight matrix.
19.6 GREY WOLF OPTIMIZER (GWO)
The GWO algorithm imitates the chasing behavior of gray wolves. It is used to improve
hyperparameters of the model. In this work, GWO is used for optimizing performance
of EfcientNetB2. The design of GWO algorithm is shown in Figure 19.3.
The accurate stages involved are
19.6.1 POSITION UPDATE
Gray wolves apprise their locations constructed on the place of
Cw. − G
w
w 1 i
1 1
Cw. − G
w
w 2 i
2
2
wolves:

300 Computational Intelligence Algorithms
=
w + w + w
3
A = 2.l v1 − l
2
vv
2
1
,
() ()
FIGURE 19.3 Grey Wolf Optimizer architecture.
where
decreases linearly from 2 to 0, and
19.6.2 OBJECTIVE FUNCTION
Classically, the objective function is the performance of negative metrics.
For hyperparameter optimization:
and
Cw. − G
w
w 3 i
3 3
1 2 3
new
= − AU.
i
are coefcient.
are random vectors in [0, 1].

301 Optimizing Digital Healthcare for Alzheimer’s Disease
19.7 RESULTS AND DISCUSSIONS
The following bar plot provides an accuracy comparison of four different DL models: ResNet, DenseNet, Inception v3, and Efcient Net.
From Figure 19.4, it is observed that 92.2% highest accuracy is achieved by
EfcientNetB2. An accuracy 89.0% is achieved by DenseNet, and ResNet has achieved
88.5% accuracy. Inception V3, when applied to the dataset provided, has recorded accuracy of 87.8%. This analysis shows that EfcientNetB2 is the most efcient of the models.
Figure 19.5 illustrates the loss comparison values for ResNet, DenseNet, Inception
v3, and EfcientNetB2. The plot in the gure shows that EfcientNetB2 attained
the lowest loss value of 0.29, which is effective for minimized prediction error and
contributes to the performance improvement and efciency. Inception v3 achieved
0.39 loss value, which is highest among the four DL architectures, showing poorer
performance than the other models. ResNet and DenseNet have attained 0.35 and
0.32 loss values, respectively. Thus, Figures 19.4 and 19.5 show that EfcientB2 has
the highest accuracy and lowest error values.
Figure 19.6 depicts the values of F1 score for different DL architectures for clas-
sication of the AD dataset. For the ResNet DL architecture, 0.87 is the F1 score,
DenseNet achieved 0.88, Inception v3 slightly slower than previous two architectures with a value of 0.86, and nally EfcientNetB2 got a 0.89 value F1 score. The
results shows that EfcientNetB2 as the most efcient model in terms of the F1
score, demonstrating its capability in balancing precision and recall. Interpretation
of the results shows that ResNet with 0.87 value and Inception v3 with 0.86 are less
efcient when compared with the other two architectures.
Figure 19.7 highlight EfcientNetB2 with ten hours training time as the most ef-
ciently designed model for training, succeeded by DenseNet at 12 hours, Inception
v3 takes 14 hours, and ResNet takes 15 hours for training. For cloud computing
FIGURE 19.4 Accuracy of EfcientNetB2.

302 Computational Intelligence Algorithms
FIGURE 19.5 Loss comparison of EfcientNetB2.
and fog computing environments, it is preferable to choose EfcientNetB2 with less
training time. These results highlight the balancing factors between EfcientNetB2
model complexity and training time.
The bar plot shown in Figure 19.8 presents testing times (in hours) for four DL
models: ResNet, DenseNet, Inception v3, and EfcientNetB2. EfcientNetB2 has
achieved the least testing time of 4.0 hours, which shows that EfcientNetB2 is not
just efcient in training but also outperforms with its testing speed.
Figure 19.9 depicts the model communication cost for the federated DL model.
When DL models are used in FL, the data are communicated between the central
FIGURE 19.6 Comparison of F1 score values.

303 Optimizing Digital Healthcare for Alzheimer’s Disease
FIGURE 19.7 Comparison of training time.
server and distributed clients. This communication includes parameter updates and
training details. As this communication affects the system’s performance, there is
a cost incurred with the process of communication. The plot shown in Figure 19.9
illustrates comparison of communication cost for the selected DL models. Due to
the concise architecture of EfcientNetB2, 0.9 GB is the cost associated with the
FIGURE 19.8 Comparison of testing time.
Соседние файлы в папке Библиотека им академика М.И. Перельмана
