A New Neural Network Model for Prediction Next Stage of Alzheimer’s

Disease

Nour Zawawi

1,3

, Heba Gamal Saber

, Mohamed Hashem

and Tarek F. Gharib

Ain Shams University, Faculty of Computer and Information Science, Cairo, Egypt

Ain Shams University, Faculty of Medicine, Geriatric Department, Cairo, Egypt

October University for Modern Sciences and Arts, Faculty of Computer Science, 6th October City, Egypt

Keywords:

Feature Selection, Time-series Forecasting, LSTM.

Abstract:

Alzheimer’s disease (AD) is a brain-related illness; The risk of development is minimized when diagnosed

early. The early detection and treatment of Alzheimer’s disease are crucial since they can decrease disease

progression, improve symptom management, allow patients to receive timely guidance and support, and save

money on healthcare. Regrettably, much current research focuses on characterizing illness states in their

current phases rather than forecasting disease development. Because Alzheimer’s disease generally progresses

in phases over time, we believe that analyzing time-sequential data can help with disease prediction. Long

short-term memory (LSTM) is a recurrent neural network that links previous input to the current task. A new

Alzheimer’s Disease Random Forest (RF) LSTM Prediction Model (RFLSTM-PM) is proposed to capture the

conditions between characteristics and the next stage of Alzheimer’s Disease after noticing that a patient’s data

could be beneﬁcial in predicting disease progression. Experiments reveal that our approach beats most existing

models and can help with early-onset AD prediction. Furthermore, tests show that it can recognize disease-

related brain regions across multiple data modalities (Magnetic resonance imaging (MRI), Neurological Test).

Also, it showed decreased value in Mean Absolute Error and Root Mean Square Error for forecasting the

progression of the disease.

1 INTRODUCTION

Alzheimer’s Disease (AD) is a form of dementia

that impacts humans’ memory, daily activities’ per-

formance, and communication abilities(Association,

2019). Effective early diagnosis and treatment of AD

is of fundamental importance as it can reduce dis-

ease progression, and therefore, reduce the substan-

tial cost for health care. Recent research shows that

only 20–40% of individual cases will change to AD

within three years; This is a lower rate of exchange

reported in medical samples than in clinical cases (Ve-

muri et al., 2017). However, AD’s progression starts

several years before any symptoms become visible

and progressive (Association, 2019). Thus, identify-

ing high-risk patients who will convert to AD is sub-

stantial (Association, 2019; Edwards III et al., 2019).

As a result, early diagnosis is essential for making a

treatment strategy to slow down the progress. It is

where the disease altered from one symptom to an-

other. At the same time, current research mainly fo-

cuses on predicting the possibility that it converts into

another stage.

In recent years, the growth of neurodegenerative

disorders such as AD has gained much interest from

researchers worldwide to develop high-performing

methods for diagnosis, treatment, preventive thera-

pies, and target drug discovery. In the risk assessment

of conversion from MCI, these variables’ change rate

could represent an additional source of knowledge

(Perneczky, 2018). Researches made it possible to di-

agnose AD using advanced diagnostic tools and com-

bine markers from different clinical features (Davda

and Corkill, 2020; Tanveer et al., 2020). Combining

markers from different clinical features and investi-

gation modalities has been shown to maximize their

sensitivity and speciﬁcity in clinical use compared to

individual biomarkers. Unfortunately, current stud-

ies mainly focus on classifying disease states in their

current stage using MRI instead of combining multi-

ple features. As a result, these studies serve as proof

of concept without being tested in the real world.

Machine learning models can help to forecast the

disease progression(Yang et al., 2018; El-Sappagh

Zawawi, N., Saber, H., Hashem, M. and Gharib, T.

A New Neural Network Model for Prediction Next Stage of Alzheimer’s Disease.

DOI: 10.5220/0010892400003122

In Proceedings of the 11th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2022), pages 689-696

ISBN: 978-989-758-549-4; ISSN: 2184-4313

689

et al., 2020) accurately. Time-series forecasting mod-

els based on recurrent memory-based approaches are

used to examine extracting patterns from sequential

healthcare data and classifying data based on diag-

nostic categories. A variety of settings in healthcare

uses predictions(Piccialli et al., 2021). They ranged

from predicting future medical outcomes and diag-

nosis to univariate time-series predictions of monthly

expenditures of patients for medication. We can ﬁnd

that it has lots of applications in the healthcare in-

dustry. Time series forecasting has been a grow-

ing science subject due to its utility in real-world

applications, but technique development has been a

concern. In medical applications, time series fore-

casting models predict sickness progression, estimate

death rates, and assess time-dependent risk(Bui et al.,

2018). However, the large number of various method-

ologies available, each of which thrives in different

scenarios, makes selecting an effective model more

complicated.

This paper addressed several of the mentioned

issues for generating time-series predictions of AD

data. On the other hand, studies show that devel-

opments linked with Alzheimer’s disease can begin

more than 20 years before symptoms arise. The pro-

posed model can identify the stage of transforma-

tion throughout time. In contrast to earlier research,

a model using LSTM (Hong et al., 2019; Basher

et al., 2021) presents to predict the progression of

Alzheimer’s disease. It employs the time step data

acquired by a data preprocessing pipeline because the

time series data may alter the prediction. It predict

time progression of AD based on these data. The pa-

per is organized as follows: Related work and previ-

ous work is discussed in section 2. Section 3 illustrate

the proposed data and methods. Section 4 discusses

the proposed prediction model (RFLSTM-PM). Fi-

nally, experiments where the models sensitivity was

tested to different features and test its stability in var-

ious data sizes shown in section 5

2 RELATED WORK

Most of the studies on Alzheimer’s disease have fo-

cused on using medical imaging as the only fac-

tor. (Mart

ı-Juan et al., 2020) is a survey concen-

trating on longitudinal imaging data. It focused on

papers that have been published between 2007 and

2019. (Hong et al., 2019) introduce Long short-

term memory (LSTM) to predict the development of

AD. It carries out the future state prediction for the

disease, rather than the state of a current diagnosis.

While (Janghel, 2020) develops and compares dif-

ferent methods to diagnose and predict AD by using

MRI scans only. It implements one model which is

the convolution neural network (CNN). At the same

time, it uses four different architectures of CNN.

An embedded feature selection method based on the

least-squares loss function and within-class scatter for

selecting the optimal feature subset are proposed by

(Cai et al., 2020). The optimal subsets of features

used for binary classiﬁcation are based on a support

vector machine (SVM). Also, deep learning technol-

ogy was discussed by (Bi et al., 2019). It focused on

the problem of automatic prediction of AD based on

MRI images. It applies two main steps: 1-implement

the unsupervised CNN for feature extraction. 2- uti-

lizes the unsupervised predictor to achieve the ﬁnal

diagnosis.

According to our knowledge, (Grassi et al., 2019)

and (Liu et al., 2020) are the only work that em-

ploys more realistic and affordable data for diagnosis.

First, (Grassi et al., 2019) use a weighted rank aver-

age grouped by different supervised machine learning

methods to predict 3 years conversion. Only a limited

set of diverse characteristics are used to make pre-

dictions. The employment of algorithmic decision-

making tools is the key beneﬁt. While,(Liu et al.,

2020) provides a new method for detecting AD based

on spectrogram characteristics collected from voice

data. This can assist families in better understand-

ing the progression of a patient’s sickness at an early

stage.

The following studies serve as the foundation

for our study. They are listed in ascending order.

The ﬁrst, (Qiu et al., 2018), explains how MRI data

can improve the accuracy of diagnoses for the Mini-

Mental State Examination and logical memory tests.

It accesses model correctness via Multilayer Precep-

tor. The second, (Grassi et al., 2019), shows how

clinically translatable strategies for conversion can

be predicted. It also detects high-risk people who

are converted. Then, (Haaksma et al., 2018) address

the link between Alzheimer’s disease and its pre-

dictors. It included some Alzheimer’s disease cases

that have had at least one examination following di-

agnosis. To determine whether there are any latent

classes of Mini Mental State Examination and Clini-

cal Dementia Rating sum of boxes routes across time.

To ﬁnd baseline predictors of class membership, re-

searchers utilised bias-corrected multinomial logistic

regression. A multimodal data (Shikalgar and Sona-

vane, 2020) classiﬁer that employs a hybrid deep neu-

ral network classiﬁer. It is based on a set of MRI pic-

tures as well as EEG inputs. The goal is to improve

the learning process by incorporating the weight com-

ponent of DNN into CNN. Then it explains how the

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

690

accuracy of hybrid classiﬁers is determined.

To ﬁnd correlations between brain areas and

genes, use the appropriate correlation analysis ap-

proach at the conclusion. (Bi et al., 2020) was

proposed via a cluster evolutionary random forest

(CERF). It adds the concept of clustering evolution

to increase the random forest’s generalisation perfor-

mance. (Farouk and Rady, 2020) investigate the use

of unsupervised clustering methods for the early iden-

tiﬁcation of Alzheimer’s disease. This research de-

veloped a two-stage technique for effectively predict-

ing Alzheimer’s Disease (AD)(Soliman et al., 2021).

Using an upgraded sparse autoencoder (SAE), an un-

supervised neural network, the initial stage includes

ﬁnding the optimal representation of the training data.

Based on the learned records and brain MRI scan, the

second stage involves employing a 3D-Convolutional

Neural Network (3D-CNN) to discern between health

and ill status.

3 METHODS AND

PREPROCESSING

3.1 Data Preparation

Data used in this article’s preparation is obtained

from the Alzheimer’s Disease Neuroimaging Initia-

tive (ADNI) database (adni.loni.usc.edu). The ADNI

was launched in 2003 as a public-private partner-

ship. ADNI’s primary goal is developing clinical,

imaging, genetic, and biochemical biomarkers for

the early detection and tracking of Alzheimer’s dis-

ease. It contains nine classes cognitive normal nor-

mal aging (NL), Mild Cognitive Impairment (MCI),

Alzheimer’s disease (AD), early mild cognitive im-

pairment (EMCI), and late mild cognitive impairment

(LMCI) to dementia or AD. The subjects recruited in

over 50 different US and Canada centers, with follow

up assessments performed every six months. The pro-

posed work uses only two types of data: 1)Magnetic

resonance imaging (MRI), 2)Assessment data. Due to

the numbers of records in each classes the only ones

that contains multiple data, only three classes is se-

lected(AD, MCI, and NL).

3.2 Subjects

Our prediction model was trained and tested on data

extracted from the ADNI; 18-month longitudinal tra-

jectories of 900 cases on each class (MCI, NL, AD).

It covers a total of 2700 instances and 90 attributes.

Each patient proﬁle consisted of multiple data sources

(24 tests and 7 image ﬁles with personal records).

Data is classiﬁed as ordinal, continuous, or image.

Patient trajectories described the time evolution of all

variables in 3-month intervals. The following subsec-

tion describes data processing steps in more detail.

The subject inclusion criteria employed in this study

are: 1- Age ranges from 55 to 90; 2-Education lev-

els range from primary to graduate; 3-All colors and

ethnicities included. The proposed work uses differ-

ent types of data: 1-Neurological test (neuropsycholo-

gist), and 2-Brain image technology (MRI only). Per-

sonal Information is excluded from the study, where

the main objective of this work is to predict the next

stage of ad over 3-month interval values.

3.3 Data Preprocessing

ADNI data (per subject) was captured multiple times

over a maximum period of 120 months. The chal-

lenge for the prediction is that it encounters many

real-world data problems, such as the following three:

• Incomplete data: Some subjects lack physical ex-

amination data at speciﬁc time points. The num-

ber of values in a set is known, but the values

themselves are unknown, resulting in incomplete

data from missing data.

• Missing data: Some subjects lack several values

of data. As a result, they are deleting the missing

values exceeding 60%. Table 1 views the rest of

the data it shows the percentage of missing data.

Other missing data are replaced by mean or vari-

ance depending on the data type.

• Time-frequency data: The different aspects of the

subjects over speciﬁc time value. However, not all

the subjects had the same time-frequency. Some

of the subjects may have only one type of data

model.

3.4 Methods

Alzheimer’s disease is a progressive neurological ill-

ness that begins gradually and worsens with time.

A primary diagnostic system, comprised of a ran-

dom forest selection and multiple illness prediction

models, is constructed to prevent disease progression.

Different methods exhibit different changes as the

disease progresses and identify relevant components.

The methodology is applied to different inputs sepa-

rately to uncover relevant biomarkers for each type.

It is critical to identify the most signiﬁcant disease-

related risk factors (Remeseiro and Bolon-Canedo,

2019; Feng et al., 2021). In recent years, most au-

thors have concentrated on hybrid approaches to fea-

ture selection (Hancer et al., 2020). Variable selection

A New Neural Network Model for Prediction Next Stage of Alzheimer’s Disease

691

Table 1: The model contains cognitive function variables, as well as MRI images. The percentage of missing data for each

feature is indicated in the missing percentage column.

Name Type Mean Missing %

CDRSB continuous 2.425 1

MMSE continuous 25.862 0

ADAS11 continuous 13.009 0

ADAS13 continuous 19.939 2

RAVLT immediate continuous 31.223 1

RAVLT learning continuous 3.547 1

RAVLT forgetting continuous 4.067 1

RAVLT perc forgetting continuous 63.86 3

FAQ continuous 6.769 0

Ventricles image – 2

Hippocampus image – 2

WholeBrain image – 1

Entorhinal image – 2

Fusiform image – 2

MidTemp image – 2

ICV image – 1

approaches include ﬁlter methods, wrapper methods,

ensemble methods, and embedding methods.

Feature selection becomes more critical in data

sets with many variables and features. Random For-

est (RF) has proven to be a practical feature selection

approach, even when dealing with many variables.

It will eliminate insigniﬁcant variables and increase

classiﬁcation accuracy and performance. The fact that

it is simple to calculate the relevance of each variable

on the tree decision contributes to its interpretabil-

ity(Hameed et al., 2015; Khaire and Dhanalakshmi,

2019). It falls under the area of embedded techniques.

Embedded methods combine ﬁlter and wrapper meth-

ods to implement algorithms with built-in feature se-

lection methods used. It indicates critical advantages

over other methodologies regarding handling highly

non-linearly correlated data, robustness to noise, tun-

ing simplicity, and opportunity for efﬁcient parallel

processing (Dimitriadis et al., 2018). Moreover, It

presents another essential characteristic: an intrinsic

feature selection step, applied before the classiﬁca-

tion task, to reduce the variables’ space by giving an

importance value to each feature (Zhong et al., 2021;

Helal et al., 2015).

LSTM networks are recurrent neural networks

that may learn order dependence in sequence predic-

tion challenges. They are a problematic area of deep

learning to master. It is required in various compli-

cated issue domains, including machine translation,

speech recognition, and others (Alom et al., 2019).

Developing and selecting accurate time series models

is a difﬁcult task because it requires training several

different models and selecting the best among them.

At the same time, it needs extensive feature engineer-

ing to derive informative features and ﬁnd optimal

time lags, which are commonly used input features

for time series models. For the processing of tempo-

ral data, LSTM models are prominent. The majority

of articles that utilize LSTM models do so with mod-

est variations. The following section describe it in

more details. Almost all of the gates in this model

contain the concept of peepholes.

Figure 1 shows the model preprocessing pipeline;

It uses two different data types. One is an MRI image,

while the second one is neurological test results. It

contains 24 data attributes (9 neurological tests) and 8

MRI that views the participant’s brain. The ﬁrst stage

in the proposed model got the patients data as input

values. In this stage the missing values and outliers

data is dealt with. Second come the feature impor-

tance where the best feature is selected. To include all

this type of data costs a high amount of money and

processing power. As a result, the proposed model

chooses the best features that will increase perfor-

mance. The output of this is the attributes that best

descripe patients next progression time. Prediction

phase organize the data with speciﬁc time and each

feature with corresponding disease stage. Finally, the

future prediction of next test value appears in sequen-

tial data preprocess.

4 AD PREDICTION MODEL

The proposed model is a combination between LSTM

and RF. This paper proposed a new hybrid model

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

692

Figure 1: Propsed Model Pipeline.

called Random Forest Least Short Term Memory Pre-

diction Model(RFLSTM-PM). It is a recurrent neu-

ral network that links the previous state to the current

one. A network with fully connected and activation

layers is created to encode the temporal relationship

between features and the next stage of AD (Hochreiter

and Schmidhuber, 1997). It is supposed to avoid the

long-term dependency problem by using a sequence

of repeating neural network modules.

Figure 2 illustrates the proposed architecture. It

involves two layers: Fully Connected Layer and Out-

Fully Connected Layer. A fully connected layer is

used to ﬁnd the correlation between these selected

features and Alzheimer’s Disease. It is connected

with the exponential linear unit (ELU) as the activa-

tion function predicts AD by feeding these features to

the layer. This model uses two types of data (MRI and

Neurological test), as stated earlier. The Out-Fully

Connected layer consists of a sigmoid layer fully con-

nected. It consist of four layers that communicate

with one another, such as:

• forget gate: The decision of whether the informa-

tion is thrown away from the cell state is made by

the forget gate, shown in Equation1

• input gate: Equation 2 and 3 show the input gate

that decides which values to update with sigmoid

and tanh layers.

• update gate: The update gate in Equation4. up-

dates the old cell state with the value from the in-

put gate.

• output gate: The output gate in Equation5 and 6

decides which value is to be output from the layer.

= σ(W

t−1

, x

] + b

) (1)

where W

is the weight matrix, b

is the bias vector,

and f

is a number between 0 and 1, where 0 repre-

sents the forget and 1 represents the keep.

= σ(W

t−1

, x

] + b

) (2)

= tanh(W

t−1

, x

] + +b

) (3)

where W

and W

are the weight matrices; b

and b

are the bias vectors; and i

are the outputs of these

two equations.

= f

∗ C

t−1

+ i

∗

(4)

where f

decides which information is to be forgotten,

and i

∗

chooses the updated values for the cell.

= σ(W

t−1

, x

] + b

) (5)

= o

∗ tanh(C

) (6)

where the value of i

in Equation 5 decides which part

of the cell state will be the output. The new cell state

multiplied by o

, and function tanh have selected,

to obtain h

in Equation 6, which is the output of the

parts t

The proposed work based on one assumption is

that each time-dependent variable in a patient’s clini-

cal record is stochastic. It is effective at differentiat-

ing between different aspects of data. RFLSTM-PM

represents the underlying time-dependent probability

distribution of value. It sampled from a range of val-

ues rather than taking on a single deterministic value.

During model training, preprocess sequential data

with time steps feds into the model, and the model

predicts the status of the following year. The model

predicts the stage of the 18-th month, one year fol-

lowing the ﬁnal month, as AD, MCI, or NI. During

model testing patient’s 18th and 24-th month features

data are entered into the model, the output is a forecast

of his state in the next year. The model selects MMSE

and CDR-SB from the test, the Ventricles view from

the MRI image. Mean Absolute Error , Absolute Per-

centage Error, and Root Mean Square Error are three

of the most critical evaluation metrics for evaluating

the forecasting model’s performance. The following

section describes the experiments and test results in

more details.

A New Neural Network Model for Prediction Next Stage of Alzheimer’s Disease

693

Figure 2: Proposed Model architecture.

5 EXPERIMENTS

Studies show that symptoms associated with AD may

begin several years before it already appears. In this

work, the prediction stated of AD after two years for

MRI image and neurological tests. As a result, the

states of disease are labeled as prediction status. The

proposed model can predict the next state with an er-

ror near 0.04. Also, the development of AD is pre-

dicted by putting into consideration multiple factors.

In order to evaluate model performance over time,

they are computing three distinct measurements and

comparing the anticipated value to the actual value.

First, Mean Absolute Error (MAE) is the most basic

measure of forecast performance. Second, compare

forecasts of different series in different scales using

the Mean Absolute Percentage Error (MAPE). Be-

cause both of these strategies are dependent on the

mean error, the impact of signiﬁcant errors is underes-

timated. The uncommon error will catch us off guard

if it focuses too much on the mean. That is the reason

why the Root Mean Square Error (RMSE) adjusts for

huge, infrequent faults.

A recurrent neural network (RNN) is a type of

Artiﬁcial Neural Network used to execute prediction

operations on sequential or time-series-based data.

These Deep learning layers are widely employed for

ordinal or temporal problems like Natural Language

Processing, Neural Machine Translation, and auto-

mated picture captioning. The LSTM is a type of

RNN that can learn long-term sequences. It is created

to avoid long-term dependency issues. Its method

of operation is to remember large sequences for an

extended period. A Gated Recurrent Unit (GRU)

workﬂow is similar to that of an RNN, except for

the operation and gates associated with each GRU

unit. GRU integrates two gate operating techniques

named Update gate and Reset gate to solve the prob-

lem presented by ordinary RNN. Table 2 shows the

model performance against the previously described

models. It shows good prediction responses over a

speciﬁc time sequence; Lower RMSE values ﬁnd a

better match. If the model objective is prediction,

then RMSE is a good indicator of how accurately the

model predicts the response, and it is the most impor-

tant criterion for ﬁt.

Table 2: Performance Metrices.

Model MAE MAPE RMSE

Proposed Model 0.0298 2.9795 0.0358

RNN 0.0363 3.8431 0.0448

GRU 0.03889 4.0859 0.0471

LSTM 0.0394 4.1171 0.0470

Furthermore, To validate the resulting model and

analyze the summary results, cross-validation based

on prediction is used. It is a statistical tool for deter-

mining how well machine learning models perform. It

is used to safeguard our predictive model from over-

ﬁtting, especially when the amount of data available

is minimal. Following the discovery of the best lag

length and the number of layers, the following are the

tested meta-parameters for the RFLSTM-PM:

1. batch size as 50, the learning rate as 0.001

2. the number of pre-fully connected cells as 128

3. the number of post-fully connected cells as 3 and

keep probability as 0.8

4. the number of fully connected cells is 128, and the

number of layers is 2

5. Sigmoid, hyperbolic tangent (tanh), exponential

linear unit (ELU) activation functions in hidden

layers. The exponential linear unit helps reach the

best values.

At the same time, the proposed model tuned with

the following parameters to discover the optimal pa-

rameters: the number of fully connected cells, the

number of cells, and the number of layers. Experi-

ments on AD vs. MCI vs. NC predict that the model

obtains the best MAE, MAPE, and RMSE. Our model

achieves the best prediction for the MMSE exam,

CDR-SB neurological test and Ventricles. Figure 3

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

694

,ﬁgure 4 and ﬁgure 5 shows a good ﬁt and stable pre-

diction for the medium term horizon of next stages in

symmetric order.

Figure 3: MMSE Prediction Results.

Figure 4: CDRSB Prediction Results.

Figure 5: Ventricles Prediction Results.

6 CONCLUSION

This research introduces a deep learning model for

predicting the next stage of Alzheimer’s disease. Be-

cause the condition is essentially progressive, the

model considers the timing data gathered from the

cases. In contrast to previous methodologies, our

model can predict the disease’s future condition rather

than classify the state of a current diagnosis. Exper-

iments have shown that our model outperforms the

vast majority of existing techniques. Improving the

model’s performance will require further research in

future studies. Ventricles’ prediction also needs to

be improved. Personal data could improve the ac-

curacy and efﬁciency of AD prediction at an earlier

stage. Furthermore, the proposed methodology will

be tested using actual data.

REFERENCES

Alom, M. Z., Taha, T. M., Yakopcic, C., Westberg, S.,

Sidike, P., , Nasrin, M. S., Hasan, M., Essen, B. C. V.,

Awwal, A. A. S., and Asari, V. K. (2019). A state-of-

the-art survey on deep learning theory and architec-

tures. Electronics, 8(3).

Association, A. (2019). 2019 alzheimer’s disease facts and

ﬁgures. Alzheimer’s & Dementia, 15(3):321–387.

Basher, A., Kim, B. C., Lee, K. H., and Jung, H. Y.

(2021). Volumetric feature-based alzheimer’s disease

diagnosis from smri data using a convolutional neu-

ral network and a deep neural network. IEEE Access,

9:29870–29882.

Bi, X., Hu, X., Wu, H., and Wang, Y. (2020). Multimodal

data analysis of alzheimer’s disease based on clus-

tering evolutionary random forest. IEEE Journal of

Biomedical and Health Informatics.

Bi, X., Li, S., Xiao, B., Li, Y., Wang, G., and Ma, X. (2019).

Computer aided alzheimer’s disease diagnosis by an

unsupervised deep learning technology. Neurocom-

puting.

Bui, C., Pham, N., Vo, A., Tran, A., Nguyen, A., and Le,

T. (2018). Time series forecasting for healthcare di-

agnosis and prognostics with the focus on cardiovas-

cular diseases. In Vo Van, T., Nguyen Le, T. A.,

and Nguyen Duc, T., editors, 6th International Con-

ference on the Development of Biomedical Engineer-

ing in Vietnam (BME6), pages 809–818, Singapore.

Springer Singapore.

Cai, J., Hu, L., Liu, Z., Zhou, K., and Zhang, H. (2020).

An embedded feature selection and multi-class clas-

siﬁcation method for detection of the progression

from mild cognitive impairment to alzheimer’s dis-

ease. Journal of Medical Imaging and Health Infor-

matics, 10(2):370–379.

Davda, N. and Corkill, R. (2020). Biomarkers in the diag-

nosis and prognosis of alzheimer’s disease. Journal of

Neurology, 267:2475–2477.

Dimitriadis, S. I., Liparas, D., Tsolaki, M. N., Initiative, A.

D. N., et al. (2018). Random forest feature selection,

fusion and ensemble strategy: Combining multiple

morphological mri measures to discriminate among

healhy elderly, mci, cmci and alzheimer’s disease pa-

tients: From the alzheimer’s disease neuroimaging

initiative (adni) database. Journal of neuroscience

methods, 15(302):14–23.

A New Neural Network Model for Prediction Next Stage of Alzheimer’s Disease

695

Edwards III, G. A., Gamez, N., Escobedo Jr, G., Calderon,

O., and Moreno-Gonzalez, I. (2019). Modiﬁable risk

factors for alzheimer’s disease. Frontiers in aging neu-

roscience, 11(164).

El-Sappagh, S., Abuhmed, T., Islam, S. R., and Kwak, K. S.

(2020). Multimodal multitask deep learning model

for alzheimer’s disease progression detection based on

time series data. Neurocomputing, 412:197–215.

Farouk, Y. and Rady, S. (2020). Early diagnosis of

alzheimer’s disease using unsupervised clustering. In-

ternational Journal of Intelligent Computing and In-

formation Sciences, 20(2):112–124.

Feng, J., Zhang, S.-W., Chen, L., and Xia, J.

(2021). Alzheimer’s disease classiﬁcation using

features extracted from nonsubsampled contourlet

subband-based individual networks. Neurocomputing,

421:260–272.

Grassi, M., Rouleaux, N., Caldirola, D., Loewenstein, D.,

Schruers, K., Perna, G., Dumontier, M., Initiative,

A. D. N., et al. (2019). A novel ensemble-based

machine learning algorithm to predict the conversion

from mild cognitive impairment to alzheimer’s disease

using socio-demographic characteristics, clinical in-

formation, and neuropsychological measures. Front

Neurology, 10(756).

Haaksma, M. L., Calder

on-Larra

naga, A., Rikkert, M.

G. O., Melis, R. J., and Leoutsakos, J. S. (2018). Cog-

nitive and functional progression in alzheimer disease:

A prediction model of latent classes. International

journal of geriatric psychiatry, 33(8).

Hameed, B., Elfetouh, A., and Elkheir, M. A. (2015).

Data cleaningtool: Usageoffuzzyroughsettheory as

machine learningpre-processing. International Jour-

nal of Intelligent Computing and Information Sci-

ences, 15:41–54.

Hancer, E., Xue, B., and Zhang, M. (2020). A survey on

feature selection approaches for clustering. Artiﬁcial

Intelligence Review volume, 53:4519–4545.

Helal, M., Elmogy, M., and Al-Awady, R. (2015). Using

rough set and boosting ensemble techniques to en-

hance classiﬁcation performance of hepatitis c virus.

International Journal of Intelligent Computing and

Information Sciences, 15:45–59.

Hochreiter, S. and Schmidhuber, J. (1997). Long Short-

Term Memory. Neural Computation, 9(8):1735–1780.

Hong, X., Lin, R., Yang, C., Zeng, N., Cai, C., Gou, J., and

Yang, J. (2019). Predicting alzheimer’s disease using

lstm. IEEE Access, 7:80893–80901.

Janghel, R. R. (2020). Deep-Learning-Based Classiﬁcation

and Diagnosis of Alzheimer’s Disease, chapter 76,

page 25. IGI Global.

Khaire, U. M. and Dhanalakshmi, R. (2019). Stability of

feature selection algorithm: A review. Journal of

King Saud University – Computer and Information

Sciences.

Liu, L., Zhao, S., Chen, H., and Wang, A. (2020). A new

machine learning method for identifying alzheimer’s

disease. Simulation Modelling Practice and Theory,

99.

Mart

ı-Juan, G., Sanroma-Guell, G., and Piellaa, G. (2020).

A survey on machine and statistical learning for lon-

gitudinal analysis of neuroimaging data in alzheimer’s

disease. Computer Methods and Programs in

Biomedicine.

Perneczky, R., editor (2018). Biomarkers for Preclinical

Alzheimer’s Disease, volume 137 of Neuromethods.

Humana Press, New York, NY.

Piccialli, F., Giampaolo, F., Prezioso, E., Camacho, D.,

and Acampora, G. (2021). Artiﬁcial intelligence and

healthcare: Forecasting of medical bookings through

multi-source time-series fusion. Information Fusion,

74:1–16.

Qiu, S., Chang, G. H., Panagia, M., Gopal, D. M., Au,

R., and Kolachalama, V. B. (2018). Fusion of deep

learning models of mri scans, mini-mental state ex-

amination, and logical memory test enhances diagno-

sis of mild cognitive impairment. Alzheimers Dement

(Amst)., 28(10):737–749.

Remeseiro, B. and Bolon-Canedo, V. (2019). A review

of feature selection methods in medical applications.

Computers in Biology and Medicine, 112.

Shikalgar, A. and Sonavane, S. (2020). Hybrid deep learn-

ing approach for classifying alzheimer disease based

on multimodal data. In Iyer, B., Deshpande, P. S.,

Sharma, S. C., and Shiurkar, U., editors, Computing

in Engineering and Technology, pages 511–520, Sin-

gapore. Springer Singapore.

Soliman, S. A., El-Dahshan, E.-S. A., and Salem, A.-B. M.

(2021). Diagnosis of alzheimer’s disease by three

dimensional convolutional neural network using un-

supervised feature learning method. International

Journal of Intelligent Computing and Information Sci-

ences.

Tanveer, M., Richhariya, B., Khan, R. U., Rashid, A. H.,

Khanna, P., Prasad, M., and Lin, C. T. (2020).

Machine learning techniques for the diagnosis of

alzheimer’s disease: A review. ACM Transactions on

Multimedia Computing, Communications, and Appli-

cations, 16(15).

Vemuri, P., Lesnick, T. G., Przybelski, S. A., Knopman,

D. S., Lowe, V. J., Graff-Radford, J., Roberts, R. O.,

Mielke, M. M., Machulda, M. M., Petersen, R. C.,

et al. (2017). Age, vascular health, and alzheimerdis-

ease biomarkers in an elderly sample. Annals of neu-

rology, 82:706–718.

Yang, C., Delcher, C., Shenkman, E., and Ranka, S. (2018).

Machine learning approaches for predicting high cost

high need patient expenditures in health care. BioMed-

ical Engineering OnLine, 17(113).

Zhong, Y., Yang, H., Zhang, Y., Li, P., and Ren, C. (2021).

Long short-term memory self-adapting online random

forests for evolving data stream regression. Neuro-

computing, 457:265–276.

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

696