Classification of Alzheimer’s Disease, Mild Cognitive Impairment,
and Normal Controls with Multilayer Perceptron Neural Network
and Neuropsychological Test Data
Ibrahim Almubark
1
, Samah Alsegehy
2
, Xiong Jiang
3
and Lin-Ching Chang
1,
*
1
Department of Electrical Engineering and Computer Science, Catholic University of America, Washington, DC, U.S.A.
2
Department of Computer Science and Engineering, Penn State University, University Park, PA, U.S.A.
3
Department of Neuroscience, Georgetown University Medical Center, Washington, DC, U.S.A.
Keywords: Multilayer Perceptron Neural Network, Alzheimer’s Disease, Mild Cognitive Impairment,
Neuropsychological Test.
Abstract: Recent advances in machine learning have shown outstanding performances in biological and medical data
analysis to assist for early detection, diagnosis, and treatment of diseases. Alzheimer's disease (AD) is a
neurodegenerative disease and the most common cause of dementia in older adults. In this study, multilayer
perceptron (MLP) neural networks are developed to classify AD, Mild Cognitive Impairment (MCI), and
Cognitive Normal (CN) subjects based upon the data from standard neuropsychological tests. Three
neuropsychological tests from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database,
Alzheimer’s Disease Assessment Scale-Cognitive subscale (ADAS-Cog), Mini-Mental State Examination
(MMSE), and Functional Activities Questionnaire (FAQ), were used to train MLP neural networks. We first
build three MLP models that can classify AD vs. CN, AD vs. MCI, and MCI vs. CN. We then construct a 3-
way MLP classifier to classify AD vs. MCI vs. CN. Finally, we propose a cascade 3-way classification method
to further improve the model performance. Using the neuropsychological test data from ADNI database, our
result shows the pairwise MLP models (i.e., AD vs. CN, AD vs. MCI, and MCI vs. CN) have the accuracy of
99.760.48, 89.643.94, and 90.812.91, respectively. The multi-class MLP model has an average accuracy
of 84.283.66, and the proposed cascaded MLP approach further improves the performance of the multi-class
classification with an average accuracy of 86.263.15.
1 INTRODUCTION
Alzheimer’s disease (AD) is a neurodegenerative
disease and the most common cause of dementia in
older adults. AD pathologies often start 5, 10, or even
20 years before symptoms appear (Alzheimer’s
Association, 2020). Symptoms usually start with
difficulty remembering new information. Since this
condition is also common with the normal aging
process, distinguishing between early AD and normal
aging can be difficult (Holtzman et al., 2011). In
advanced stages, symptoms include confusion, mood
and behavior changes, and inability to care for one’s
self and perform basic life tasks. AD is ultimately
fatal (Taylor et al., 2017). While significant progress
has been made, there are yet no proven effective
treatments for AD. As a result, there is increasing
*
For the Alzheimer’s Disease Neuroimaging Initiative
pressure to develop techniques to assist in the
diagnosis of early AD, as early intervention may be
most effective in treating or slowing disease progress.
In addition, early diagnosis may provide useful
information for the development of more effective
treatments (Fiandaca et al., 2014).
Three groups of subjects are included in this
machine learning classification study: Cognitively
Normal (CN) older adults, Mild Cognitive
Impairment (MCI) due to AD, and AD. A CN subject
has no signs of cognitive impairment other than age-
related normal decline. A subject converts from CN
to MCI when symptoms become mild yet noticeable
to the patient or close family members/friends. MCI
is a transitional stage between CN and AD, and is the
earliest clinically detectable stage of progression
towards dementia or AD (Sperling et al., 2011).
Approximately 15-20% of seniors age 65 or older
Almubark, I., Alsegehy, S., Jiang, X. and Chang, L.
Classification of Alzheimer’s Disease, Mild Cognitive Impairment, and Normal Controls with Multilayer Perceptron Neural Network and Neuropsychological Test Data.
DOI: 10.5220/0010143304390446
In Proceedings of the 12th International Joint Conference on Computational Intelligence (IJCCI 2020), pages 439-446
ISBN: 978-989-758-475-6
Copyright
c
2020 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
439
have MCI (Alzheimer’s Association, 2020). Patients
diagnosed with MCI are of higher risk for developing
AD or other types of dementia, and are therefore
given special attention (Ansart et al., 2020). AD is the
last stage in this progression.
In recent years, the world has seen many major
breakthroughs in the field of healthcare because of the
rapid proliferation of large biomedical datasets,
concurrent with advances in machine learning,
especially in deep learning (Esteva et al., 2017).
These advances have opened new avenues for the
development of diagnostic tools to assist early
detection of AD. Recently, several studies have
focused on the detection of different cognitive groups
by utilizing various types of biomedical data
including brain imaging data (Pellegrini et al., 2018),
cerebrospinal fluid (CSF) specimens (Jack et al.,
2018; Shaffer et al., 2013), and behavioral data from
speech (Fraser et al., 2016; Nagumo et al., 2020),
body movement (Khan & Jacobs, 2020), and
neuropsychological tests (Grassi et al., 2019; Kang et
al., 2019; Lee et al., 2019).
Standard neuropsychological tests are typically
used in the diagnosis of cognitive impairment in
individuals with MCI, AD, or other neurological
conditions (Seo, 2018). These tests are less
expensive, easy to conduct, and widely available
compared to medical imaging. The scores from these
tests can measure normal and abnormal cognitive and
behavioral functions and provide useful features to
machine learning methods for the early detection of
AD (Anastasi & Urbina, 1997). Repeated assessment
with these tests is frequently used to evaluate changes
in a treated person’s condition over time (Harvey,
2012). The existence of multiple cognitive deficits
indicates that a combination of neuropsychological
tests from different domains may improve clinical
diagnosis accuracy (Harvey, 2012; Storey & Kinsella,
2007; Yeatts et al., 2018).
In this paper, we present fully connected
multilayer perceptron (MLP) networks to perform
binary classification between different cognitive
groups (i.e., AD vs. CN, AD vs. MCI, and MCI vs.
CN) using the baseline visit data from three
neuropsychological tests in the Alzheimer’s Disease
Neuroimaging Initiative (ADNI) database
(http://adni.loni.usc.edu/). A direct MLP based 3-way
classification (i.e., AD vs. MCI vs. CN) is also
developed. Additionally, we propose a MLP based
cascading approach to further improve the multi-class
classification performance.
The paper is organized as follows. Section 2
includes descriptions of data used in this study, data
pre-processing, experimental design, and proposed
methods. Section 3 includes results from several
multilayer perceptron models. Section 4 includes a
summary of results, discussion, and future research
directions.
2 MATERIALS AND METHODS
2.1 Data
2.1.1 ADNI Database
The data used in this study was obtained from the
ADNI database. ADNI is a longitudinal multicenter
study designed to develop clinical, imaging, genetic,
and biochemical biomarkers for the early detection
and tracking of AD progression. The ADNI project
draws on a broad range of academic institutions and
private corporations, with subjects recruited from
over 50 sites across the U.S. and Canada. The project
began in 2003 and has been extended to different
phases. The first phase of ADNI (ADNI-1) was
completed in 2010 and has been followed by ADNI-
GO, ADNI-2, and ADNI-3. These four protocols
have recruited over 1900 adults, with ages from 55 to
90, and consist of elderly CN controls, people with
MCI, and people with AD. The follow-up duration of
each group is described in the protocols for ADNI-1,
ADNI-GO, ADNI-2, and ADNI-3. For detailed
information, please see (www.adni-info.org).
2.1.2 Subjects
In this study, we used the baseline visit data from a
total of 808 subjects at the initial project period
(ADNI-1), including 188 AD, 391 MCI, and 229 CN.
The enrolled subjects were between 55-90 (inclusive)
years of age, in good general health, having a partner
who is able to provide an independent assessment of
the subject’s functioning, having at least 6 grades of
education or work history, and were fluent in English
or Spanish. All subjects and their study partners
completed the informed consent process, and study
protocols were reviewed and approved by the
Institutional Review Board at each ADNI data
collection site (Petersen et al., 2010). Table 1 shows
the characteristics of the AD, MCI, and CN subjects
included in this study. The mean test score was
computed by averaging the scores from all the
questions in one test.
NCTA 2020 - 12th International Conference on Neural Computation Theory and Applications
440
Table 1: Characteristics of subjects at their baseline visit recruited during ADNI-1.
Characteristic AD (n=188) MCI (n=391) CN (n=229) p-value
Age, years
74.97.4 74.4
7.3 75.4
5.0
0.167
Education, years
14.73.1 15.6
3.0 16.1
2.8 3.9810
-5
Sex, male/female 97/91 255/136 119/110
5.4310
-4
ADAS-Cog score
28.97.6 18.6
6.3 9.5
4.2 4.9510
-145
MMSE score
23.42.0 27.0
1.8 29.1
0.9 3.0810
-164
FAQ score
13.06.8 3.8
4.5 0.1
0.6 4.0710
-28
Values are shown as mean standard deviation or gender ratios. The p-values for differences between AD, MCI, and CN
are based on t-test. ADAS-Cog = Alzheimer's Disease Assessment Scale-Cognitive subscale; MMSE = Mini-Mental State
Examination; FAQ = Functional Activities Questionnaire.
2.1.3 Neuropsychological Data
The itemized scores of three neuropsychological tests
were used, which include Alzheimers Disease
Assessment Scale-Cognitive subscale (ADAS-Cog)
(Rosen et al., 1984), Mini-Mental State Examination
(MMSE) (Folstein et al., 1975), and Functional
Activities Questionnaire (FAQ) (Pfeffer et al., 1982).
Table 2 details the cognitive/daily functions
associated with individual tests. In total, there are 13,
30, and 10 questions from ADAS-Cog (note that Q13
was not used by default), MMSE, and FAQ,
respectively. The score of each individual question is
treated as a feature in our machine learning task,
resulting in a total of 13, 30, and 10 features for the
ADAS-Cog, MMSE, and FAQ datasets. The three
neuropsychological tests are widely used to assist
cognitive impairment in AD. The most prominent
feature of AD is memory impairment. Therefore,
ADAS-Cog and MMSE tests were used in this study.
ADAS-Cog and MMSE also test global cognitive
function as well as several domains other than
memory (Casanova et al., 2013). Information on
function from the FAQ test was also included as
functional changes begin to appear earlier in the
dementia process (John et al., 2016).
Table 2: Neuropsychological tests used in this study.
Neuropsychological Tests
ADAS-Cog Registration (3)
Q1. Word recall Attention and calculation (5)
Q2. Word recognition Recall (3)
Q3. Object naming Language (8)
Q4. Recall test instructions Visual construction (1)
Q5. Orientation FAQ
Q6. Commands Q1. Manage finances
Q7. Clarity of language Q2. Complete forms
Q8. Comprehension Q3. Shop
Q9. Word finding Q4. Perform games of skill or hobbies
Q10. Ideational praxis Q5. Prepare hot beverages
Q11. Constructional praxis Q6. Prepare a balanced meal
Q12. Delayed word recall Q7. Follow current events
Q14. Number cancellation Q8. Attend to TV, books, or magazines
MMSE Q9. Remember appointments
Orientation (10) Q10. Travel out of the neighborhood
2.2 Experimental Design for
Classification
Our multilayer perceptron (MLP) neural networks
were developed using (1) the original set of features
from each neuropsychological test (i.e., 13 from
ADAS-Cog, 30 from MMSE, and 10 from FAQ), and
(2) the combined-test of 53 features from three
neuropsychological tests. We first trained three
binary MLP models to perform binary classification
between different cognitive groups for both the
individual tests and the combined-test: (Case 1) AD
vs. CN, (Case 2) AD vs. MCI, and (Case 3) MCI vs.
CN. We also trained a MLP network with multi-class
classification network, i.e., AD vs. MCI vs. CN (case
4). Furthermore, we proposed a MLP based cascading
approach to further improve the multi-class
classification performance (case 5).
The proposed cascade MLP method is composed
of 2 steps. In the first step, we classified CN vs. (AD
+ MCI) with a MLP network. Then we trained
another MLP network with the predicted AD or MCI
samples from the first step to classify AD vs. MCI
(step 2). Note that the true CN samples misclassified
in step 1 as (AD + MCI) will participate in the second
step and they will be counted as misclassified CN
regardless of their predicted results in step 2. On the
other hand, the true MCI or AD samples misclassified
as CN will not participate in the training in step 2, but
will be counted in the final 3×3 confusion matrix
along with the correctly classified CN samples after
step 1.
The implementation was carried out using Python
and related libraries including Scikit-learn, Pandas,
Numpy, TensorFlow, and Keras (Chollet, 2015;
Pedregosa et al., 2011).
2.3 Data Pre-processing
The original baseline visit dataset in ADNI-1 consists
of 200 AD, 400 MCI, and 229 CN. We excluded 12
subjects from AD and 9 subjects from MCI before our
data analysis since answers to some questions were
Classification of Alzheimer’s Disease, Mild Cognitive Impairment, and Normal Controls with Multilayer Perceptron Neural Network and
Neuropsychological Test Data
441
not recorded (i.e., missing values in some features).
Therefore, the dataset (n=808, AD=188, MCI=391,
CN=229) used in our study does not have any missing
values. In statistics it is a common practice to drop
cases with missing values as long as the sample size
is sufficiently large and the number of dropped cases
does not exceed 5% of the overall sample. In this
study, the number of missing value subjects was
relatively small (12 out of 200 for AD and 9 out of
400 for MCI). Therefore, we elected to drop without
replacement the twenty-one samples with missing
value. Feature normalization was performed by
standard scaling with a zero mean and standard
deviation equal to one.
2.4 Data Partitioning
Cross-validation (CV) with stratified K-Fold was
used to evaluate the predictive model. Data was
divided into 5 disjoint subsets with consistent ratios
between classes in each fold. Eighty percent of the
data was used in training and 20% of the data was
used for testing in each fold.
2.5 Multilayer Perceptron (MLP)
Neural Network
Multilayer perceptron (MLP) is a feed-forward
artificial neural networks that uses back-propagation
to update weights (Marsland, 2015). The neurons are
connected to later layers in a way that pushes
information from the input, through hidden layer(s),
to the output layer. MLP leverages a layered
architecture of stacked perceptrons to solve complex,
often supervised, problems. MLP can approximate
non-linear functions for both classification and
regression (Joshi, 2020).
In this paper, we developed multiple fully
connected MLP networks to classify different
cognitive groups (AD, MCI, and CN) using data from
three neuropsychological tests. Figure 1 shows a 3-
layer MLP network that can be used to classify AD
and CN subjects. As an example, if the combined-test
data was used, the resulted MLP model will have 53
nodes in the input layer, 6 nodes in the hidden layer,
and 1 node in the output layer. The Rectified Linear
Unit (ReLU) was selected as the activation function
for the input and hidden layers (Xu et al., 2015). The
sigmoid function and the binary cross-entropy loss
function were used for the binary classifications. The
softmax function and the multi-class cross-entropy
were used for the multi-class classification (Sharma,
2017). The Adaptive Movement Estimation (Adam)
was used as our optimizer to tune the network during
the training (Kingma & Ba, 2014).
The cross-entropy loss function defined in
equation (1) is used to quantify MLP model errors:
𝑤
𝑝
𝑐|𝑥
log𝑞
𝑐|𝑥

𝜆
||𝑤
,
||
(1)
where c, i, k are indices for classes, samples, and
layers, respectively; w
c
is the class weight for class c;
p
𝑐|𝑥
is the true probability for sample x
i
to be
assigned to class c; q
𝑐|𝑥
is the predicted
probability for sample x
i
to be assigned to class c;
𝜆
is the regularization strength for layer k; and w
k,k-1
are the weights between the (k-1)
th
and the k
th
layer.
A larger class weight on class c will penalize more if
samples in class c are misclassified. We also
considered the probability threshold as another hyper-
parameter. For example, a sample can be predicted to
have a probability of 0.45 to be class 1 and 0.55 to be
class 0, if we set this probability at default 0.5, this
sample will be classified to class 0. However, one
could set the probability threshold to 0.4 instead, the
sample will then be classified to class 1. In our
training, we also included the class weights and
probability thresholds as hyper-parameters to avoid
the imbalance issue between the model sensitivity and
specificity.
Figure 1: A 3-layer MLP network. Circular nodes represent
artificial neurons. Arrows represent connection from a
neuron output to a neuron input.
NCTA 2020 - 12th International Conference on Neural Computation Theory and Applications
442
To avoid overfitting, we have employed L2
regularization in the hidden layers as shown in
equation (1). The regularization strength for each
layer λ
was also tuned as a hyper-parameter. We also
applied the EarlyStopping function in Keras by
observing the loss function on the validation set. If the
loss function reduction was less than (110

) for
5 consecutive epochs, the training will be stopped.
Learning rate (or shrinkage factor) was adjusted
based on the reduction rate of loss function to train
the network more efficiently, and the
ReduceLROnPlateau function in Keras was used to
observe the loss function reduction. If the loss
function reduction was less than (110

) for 10
consecutive epochs, the learning rate would be
reduced to one tenth of the previous learning rate or
the minimum learning rate. The initial learning rate
was set to 0.01 and the minimum value is (510

).
The early stopping and the learning rate shrinkage
helped our training process with a positive impact in
the classification performance.
2.6 Performance Evaluation
To evaluate the performance of the classifiers,
sensitivity, specificity, and accuracy were calculated
for each model. Sensitivity measures the ratio of
actual positive subjects to the total numbers of
subjects identified by the test as being positive, i.e.,
true positive rate. Specificity measures the ratio of
actual negative subjects to the total number of
subjects testing negative, i.e., true negative rate.
Accuracy is the ratio of correctly classified subjects
to the entire set of subjects. In other words,
sensitivity, specificity, and accuracy are described in
terms of TP (True Positives), TN (True Negatives),
FN (False Negatives), and FP (False Positives), and
defined in equations (2), (3), and (4), respectively.
Sensitivit
y
TP
TP  FN
(2)
Specificit
y
TN
TN  FP
(3)
Accurac
y
TP  TN
TP  TN  FP  FN
(4)
To illustrate the diagnostic ability of a classifier,
we also calculated the Area Under the Curve (AUC)
from the Receiver Operating Characteristic (ROC)
curve. The AUC was also used as the target metric
during hyper-parameter tuning. An algorithm with an
AUC closer to 1, indicating a near perfect
performance, is considered as the more reliable
predictive model.
3 RESULTS AND DISCUSSION
Table 3 summarizes the performance of binary and 3-
way classification using each individual
neuropsychological test (i.e., ADAS-Cog, MMSE,
and FAQ) as well as a combination of these three tests
to discriminate different cognitive groups. The
default class weights was set to 1:1 for binary
classifier or 1:1:1 for 3-way classifier, and the default
probability threshold was set to 0.5. As shown in
Table 3, the model using the combined features from
three tests outperformed the models using each single
test. The classification of AD vs. CN subjects had
very high accuracies (98%~100%) in all models. This
indicates the proposed MLP method using
neuropsychological test data is very effective in
classifying AD and CN. The classification accuracy
of MCI vs. CN was 77%~82% when using a single
test, but reached 90% when using the combined tests.
Classification between AD and MCI was our most
challenging task. Although the overall accuracy was
acceptable (80%~84%), the sensitivity was very poor
(47%~69%) when using a single neuropsychological
test. However, the sensitivity significantly improved,
to 81.38%, when using the combined-test.
In Table 4, we demonstrated that class weights
and probability thresholds could be used to improve
the model performance and obtain a more balanced
sensitivity and specificity ratio. By tuning the class
weights and probability thresholds in training of MLP
networks, the sensitivity as well as the accuracy can
be further improved. For example, the sensitivity of
AD vs. MCI improved from 46.81% to 79.26% with
ADAS-Cog test (data not shown) and from 81.38% to
91.49% with the combined-test. The accuracy of AD
vs. MCI vs. CN was improved from 82.43% to
84.28%. While this MLP 3-way classification
accuracy is notably lower than the binary
classifications, it outperformed other existing
methods. For example, its accuracy was 22% higher
than Lee’s model (Lee et al., 2019).
Table 5 shows the performance of the MLP model
with the cascade approach using the combined-test.
The class weights and probability thresholds were
tuned to obtain optimal model performance. This new
model further improved the results compared to the
direct 3-way classification (Tables 3 and 4) in terms
of sensitivity, specificity, and accuracy.
Standard neuropsychological tests are often
incorporated into regular physical examinations for
Classification of Alzheimer’s Disease, Mild Cognitive Impairment, and Normal Controls with Multilayer Perceptron Neural Network and
Neuropsychological Test Data
443
seniors, this study demonstrated that early screening
of AD is possible when using these tests with neural
networks. The proposed methods are effective to
classify different cognitive groups, and do not require
the medical procedures that are presently more
expensive, invasive, or not offered in many clinical
settings. These medical procedures include
neuroimaging, cerebrospinal fluid (CSF), and genetic
testing. This study also showed that a combination of
a variety of neuropsychological tests and assessments
used for AD diagnosis improved the accuracy of
clinical diagnosis. Lastly, this study points to the
potential for MLP neural network enabled classifiers
in discriminating between AD progression classes.
Table 3: The classification performance of binary and 3-way multilayer perceptron (MLP) networks using data from a single
neuropsychological test and the combined-test. The default class weights (1:1 or 1:1:1) and probability thresholds (0.5) were
used in each model.
Dataset Classification Case SEN% SPE% ACC% AUC
ADAS-Cog
(13)
AD vs. CN 93.62 98.69
96.04 1.71
0.994
AD vs. MCI 46.81 95.43
79.73 2.29
0.874
MCI vs. CN 82.23 80.35
81.54 5.24
0.905
AD vs. MCI vs. CN 72.75 80.90
72.75 2.82
0.887
MMSE
(30)
AD vs. CN 95.85 97.82
96.92 1.62
0.998
AD vs. MCI 69.43 92.19
84.75 3.75
0.914
MCI vs CN 78.34 75.98
77.48 3.43
0.857
AD vs. MCI vs. CN 70.21 80.28
70.21 5.89
0.873
FAQ
(10)
AD vs. CN 90.16 99.56
95.26 1.52
0.982
AD vs. MCI 54.40 92.89
80.24 3.50
0.868
MCI vs. CN 71.83 92.14
79.29 2.26
0.845
AD vs. MCI vs. CN 71.32 85.00
71.33 3.74
0.853
Combined-Test
(53)
AD vs. CN 98.04 100.00
99.28 0.59
1.0
AD vs. MCI 81.38 93.35
89.46 3.93
0.964
MCI vs. CN 90.79 89.08
90.16 3.55
0.960
AD vs. MCI vs. CN 82.43 88.62
82.43 3.92
0.946
Table 4: The classification performance of binary and 3-way multilayer perceptron (MLP) networks using data from the
combined three neuropsychological tests. The class weights and probability thresholds were tuned during the training to
obtain a balanced sensitivity and specificity ratio.
Classification Case Probability Threshold Class Weight SEN% SPE% ACC% AUC
AD vs. CN 0.5 1:1.5 99.47 100.00
99.76 0.48
1.0
AD vs. MCI 0.4 1:1.5 91.49 88.75
89.64 3.94
0.965
MCI vs. CN 0.5 1:1.5 92.07 88.65
90.81 2.91
0.964
AD vs. MCI vs. CN 0.5 1.5:1.5:1 84.28 90.36
84.28 3.66
0.954
Table 5: The multilayer perceptron (MLP) cascading classification performance. The classification performance of the tuned
class weights and probability thresholds are shown.
Classification steps Probability Threshold Class Weight SEN% SPE% ACC% AUC
Step1: CN vs. (AD + MCI) 0.5 1:1 93.27 92.14
92.95 2.33
0.973
Step2: AD vs. MCI using the (AD + MCI)
from step 1
0.6 1:1 86.26 91.15
86.26 3.15
0.957
NCTA 2020 - 12th International Conference on Neural Computation Theory and Applications
444
4 CONCLUSIONS AND FUTURE
WORK
Most previous studies on AD detection using machine
learning techniques have been focusing on utilizing
brain imaging data. With the rich availability and low
cost of standard neuropsychological tests, this paper
investigated the classification performance for
detecting different cognitive groups (AD, MCI, and
CN) using MLP neural networks. Several important
conclusions can be drawn from this study. First, using
a single neuropsychological test to classify AD and
MCI yielded a very poor sensitivity. Second, the
combination of three neuropsychological test data
with MLP networks showed good potential for early
AD detection. The MLP classifiers performed well on
all three binary cases with the combined-test as well
as for the 3-way classification. Finally, the proposed
cascade MLP approach can further improve the
performance of multi-class classification.
The proposed method is not only reliable but also
cost effective, and therefore it can support large-scale
cognitive screening. Our future work will include
identifying individuals with MCI who would be more
likely to develop AD within a defined period of time.
Additionally, we will investigate other artificial
neural networks on diagnostics and prediction of AD.
We also plan to study the combination of brain
imaging and behavioral data with both machine
learning and deep learning techniques that may offer
additional insights into the progression of various
stages of AD.
ACKNOWLEDGEMENTS
Data used in preparation for this article was obtained
from the Alzheimer's Disease Neuroimaging
Initiative (ADNI) database (adni.loni.usc.edu). As
such, the investigators within the ADNI contributed
to the design and implementation of ADNI and/or
provided data but did not participate in analysis or
writing of this report. A complete listing of ADNI
investigators can be found at:
http://adni.loni.usc.edu/wp-content/uploads/how to
apply/ADNI_Acknowledgement_List.pdf.
REFERENCES
Alzheimer’s Association. (2020). 2020 Alzheimer’s disease
facts and figures. Alzheimer’s & Dementia, 16(3), 391–
460.
Anastasi, A., & Urbina, S. (1997). Psychological testing
(7th ed.). Prentice Hall/Pearson Education.
Ansart, M., Epelbaum, S., Bassignana, G., Bône, A.,
Bottani, S., Cattai, T., Couronne, R., Faouzi, J., Koval,
I., & Louis, M. (2020). Predicting the Progression of
Mild Cognitive Impairment Using Machine Learning:
A Systematic, Quantitative and Critical Review.
MedRxiv.
Casanova, R., Hsu, F.-C., Sink, K. M., Rapp, S. R.,
Williamson, J. D., Resnick, S. M., Espeland, M. A., &
Initiative, A. D. N. (2013). Alzheimer’s disease risk
assessment using large-scale machine learning
methods. PLoS One, 8(11), e77949.
Chollet, F. (2015). Keras: Deep learning library for theano
and tensorflow. URL: Https://Keras. Io/k, 7(8), T1.
Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M.,
Blau, H. M., & Thrun, S. (2017). Dermatologist-level
classification of skin cancer with deep neural networks.
Nature, 542(7639), 115–118.
https://doi.org/10.1038/nature21056
Fiandaca, M. S., Mapstone, M. E., Cheema, A. K., &
Federoff, H. J. (2014). The critical need for defining
preclinical biomarkers in Alzheimer’s disease.
Alzheimer’s & Dementia: The Journal of the
Alzheimer’s Association, 10(3 Suppl), S196-212.
https://doi.org/10.1016/j.jalz.2014.04.015
Folstein, M. F., Folstein, S. E., & McHugh, P. R. (1975).
“Mini-mental state”: A practical method for grading the
cognitive state of patients for the clinician. Journal of
Psychiatric Research, 12(3), 189–198.
Fraser, K. C., Meltzer, J. A., & Rudzicz, F. (2016).
Linguistic Features Identify Alzheimer’s Disease in
Narrative Speech. Journal of Alzheimer’s Disease:
JAD, 49(2), 407–422. https://doi.org/10.3233/JAD-
150520
Grassi, M., Rouleaux, N., Caldirola, D., Loewenstein, D.,
Schruers, K., Perna, G., & Dumontier, M. (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 Information And
Neuropsychological Measures. Frontiers in Neurology,
10, 756.
Harvey, P. D. (2012). Clinical applications of
neuropsychological assessment. Dialogues in Clinical
Neuroscience, 14(1), 91.
Holtzman, D. M., Morris, J. C., & Goate, A. M. (2011).
Alzheimer’s disease: The challenge of the second
century. Science Translational Medicine, 3(77), 77sr1–
77sr1.
Jack, C. R., Bennett, D. A., Blennow, K., Carrillo, M. C.,
Dunn, B., Haeberlein, S. B., Holtzman, D. M., Jagust,
W., Jessen, F., Karlawish, J., Liu, E., Molinuevo, J. L.,
Montine, T., Phelps, C., Rankin, K. P., Rowe, C. C.,
Scheltens, P., Siemers, E., Snyder, H. M.,
Contributors. (2018). NIA-AA Research Framework:
Toward a biological definition of Alzheimer’s disease.
Alzheimer’s & Dementia: The Journal of the
Alzheimer’s Association, 14(4), 535–562.
https://doi.org/10.1016/j.jalz.2018.02.018
Classification of Alzheimer’s Disease, Mild Cognitive Impairment, and Normal Controls with Multilayer Perceptron Neural Network and
Neuropsychological Test Data
445
John, S. E., Gurnani, A. S., Bussell, C., Saurman, J. L.,
Griffin, J. W., & Gavett, B. E. (2016). The effectiveness
and unique contribution of neuropsychological tests
and the δ latent phenotype in the differential diagnosis
of dementia in the uniform data set. Neuropsychology,
30(8), 946.
Joshi, A. V. (2020). Perceptron and Neural Networks. In
Machine Learning and Artificial Intelligence (pp. 43–
51). Springer, Cham.
Kang, M. J., Kim, S. Y., Na, D. L., Kim, B. C., Yang, D.
W., Kim, E.-J., Na, H. R., Han, H. J., Lee, J.-H., Kim,
J. H., Park, K. H., Park, K. W., Han, S.-H., Kim, S. Y.,
Yoon, S. J., Yoon, B., Seo, S. W., Moon, S. Y., Yang,
Y., … Youn, Y. C. (2019). Prediction of cognitive
impairment via deep learning trained with multi-center
neuropsychological test data. BMC Medical
Informatics and Decision Making, 19(1), 231.
https://doi.org/10.1186/s12911-019-0974-x
Khan, T., & Jacobs, P. (2020). Prediction of mild cognitive
impairment using movement complexity. IEEE Journal
of Biomedical and Health Informatics.
https://doi.org/10.1109/JBHI.2020.2985907
Kingma, D. P., & Ba, J. (2014). Adam: A method for
stochastic optimization. ArXiv Preprint
ArXiv:1412.6980.
Lee, G. G. C., Huang, P.-W., Xie, Y.-R., & Pai, M.-C.
(2019). Classification of Alzheimer’s Disease, Mild
Cognitive Impairment, and Cognitively Normal Based
on Neuropsychological Data via Supervised Learning.
TENCON 2019-2019 IEEE Region 10 Conference
(TENCON), 1808–1812.
Marsland, S. (2015). Machine learning: An algorithmic
perspective (2nd ed). CRC press.
Nagumo, R., Zhang, Y., Ogawa, Y., Hosokawa, M., Abe,
K., Ukeda, T., Sumi, S., Kurita, S., Nakakubo, S., Lee,
S., Doi, T., & Shimada, H. (2020). Automatic Detection
of Cognitive Impairments through Acoustic Analysis of
Speech. Current Alzheimer Research, 17(1), 60–68.
https://doi.org/10.2174/1567205017666200213094513
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,
Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P.,
Weiss, R., & Dubourg, V. (2011). Scikit-learn:
Machine learning in Python. Journal of Machine
Learning Research, 12(Oct), 2825–2830.
Pellegrini, E., Ballerini, L., Hernandez, M. D. C. V.,
Chappell, F. M., González-Castro, V., Anblagan, D.,
Danso, S., Muñoz-Maniega, S., Job, D., Pernet, C.,
Mair, G., MacGillivray, T. J., Trucco, E., & Wardlaw,
J. M. (2018). Machine learning of neuroimaging for
assisted diagnosis of cognitive impairment and
dementia: A systematic review. Alzheimer’s &
Dementia (Amsterdam, Netherlands), 10, 519–535.
https://doi.org/10.1016/j.dadm.2018.07.004
Petersen, R. C., Aisen, P. S., Beckett, L. A., Donohue, M.
C., Gamst, A. C., & Harvey, D. J. (2010). Alzheimer’s
disease neuroimaging initiative (ADNI). Neurology,
74(3), 201–209.
Pfeffer, R. I., Kurosaki, T. T., Harrah Jr, C. H., Chance, J.
M., & Filos, S. (1982). Measurement of functional
activities in older adults in the community. Journal of
Gerontology, 37(3), 323–329.
Rosen, W., Mohs, R., & Davis, K. (1984). Alzheimer’s
Disease Assessment Scale—Cognitive and Non-
Cognitive Sections (ADAS-Cog, ADAS Non-Cog).
Journal of Psychiatry, 141, 1356–1364.
Seo, E. H. (2018). Neuropsychological assessment of
dementia and cognitive disorders. Journal of Korean
Neuropsychiatric Association, 57(1), 2–11.
Shaffer, J. L., Petrella, J. R., Sheldon, F. C., Choudhury, K.
R., Calhoun, V. D., Coleman, R. E., Doraiswamy, P.
M., & Initiative, A. D. N. (2013). Predicting cognitive
decline in subjects at risk for Alzheimer disease by
using combined cerebrospinal fluid, MR imaging, and
PET biomarkers. Radiology, 266(2), 583–591.
Sharma, S. (2017). Activation functions in neural networks.
Towards Data Science, 6(12), 310–316.
Sperling, R. A., Aisen, P. S., Beckett, L. A., Bennett, D. A.,
Craft, S., Fagan, A. M., Iwatsubo, T., Jack Jr, C. R.,
Kaye, J., & Montine, T. J. (2011). Toward defining the
preclinical stages of Alzheimer’s disease:
Recommendations from the National Institute on
Aging-Alzheimer’s Association workgroups on
diagnostic guidelines for Alzheimer’s disease.
Alzheimer’s & Dementia, 7(3), 280–292.
Storey, E., & Kinsella, G. J. (2007). Principles of
neuropsychometric assessment. In Neurology and
Clinical Neuroscience (pp. 22–30). Mosby.
Taylor, C. A., Greenlund, S. F., McGuire, L. C., Lu, H., &
Croft, J. B. (2017). Deaths from Alzheimer’s Disease—
United States, 1999–2014. MMWR. Morbidity and
Mortality Weekly Report, 66(20), 521.
Xu, B., Wang, N., Chen, T., & Li, M. (2015). Empirical
evaluation of rectified activations in convolutional
network. ArXiv Preprint ArXiv:1505.00853.
Yeatts, S. D., Palesch, Y. Y., & Temkin, N. (2018).
Biostatistical issues in TBI clinical trials. In Handbook
of Neuroemergency Clinical Trials (pp. 167–185).
Elsevier.
NCTA 2020 - 12th International Conference on Neural Computation Theory and Applications
446