16 IEEE ENGINEERING MANAGEMENT REVIEW, VOL. 51, NO.

1, FIRST QUARTER, MARCH 2023

Early Detection of Alzheimer’s

Disease Using Cognitive Features:
A Voting-Based Ensemble Machine
Learning Approach
—MUHAMMAD IRFAN Abstract—Early detection of Alzheimer’s disease (AD) is vital for adequate
Western Sydney University, Penrith South, control. Machine learning techniques have gained much attraction due to their
NSW 2751, Australia efficiency in predicting AD using cognitive tests. Ensemble machine learning
models are helpful in improving the robustness of the learning system via
—SEYED SHAHRESTANI combining multiple machine learning models. This article proposes a novel
Western Sydney University, Penrith South, ensemble machine learning technique for the early detection of AD. First, a
NSW 2751, Australia
novel feature selection technique referred to as Neighborhood Component
Analysis and Correlation-based Filtration (NCA-F) is proposed to select the
—MAHMOUD ELKHODR vital cognitive features from a given dataset. Second, various machine
School of Engineering and Technology, Central
learning classifiers were trained using the proposed NCA-F method. The top
Queensland University, Sydney, NSW 2700,
Australia classifiers were selected for voting based on the performance results. The
voting is performed using an adaptive weight matrix process. The output label
of a model is multiplied by the F1 score and represented as weight. The results
revealed an accuracy of 93.92% when using adaptive voting, which is better
(Corresponding author: Muhammad Irfan.) than the accuracy of 90.53% observed when using the traditional artificial
neural network method. The proposed technique improved the accuracy of
detecting AD at an early stage. Furthermore, the results against a recent study
IEEE DOI 10.1109/EMR.2022.3230820 using same features also revealed an improvement of 12.12% in accuracy.

Key words: Adaptive voting, Alzheimer’s disease (AD), cognitive features,

machine learning (ML), Neighborhood Component Analysis and Correlation-
based Filtration (NCA-F)

I. INTRODUCTION widespread mainly in people over the

age of 65. The likelihood of the AD
A LZHEIMER’S disease (AD) is a and other types of dementia diseases
increases with age. For instance, AD
progressive disorder where the
symptoms gradually develop over the affects 1 in 14 persons over 65 and 1
years and increase in severity. It in 6 of those over 80 [1]. Therefore,
affects many of the brain functions. precise and early AD symptoms may
The primary sign of AD that can be enable early intervention.
observed in individuals is memory
Predicting AD is a complicated task
loss, such as failing to remember
as it is challenging to identify the
recent events, conversations, names,
dementia subtype [2]. Interestingly,
and objects. As the AD progresses, research shows that AD is
the memory problem develops into accountable for about two-thirds of all
more severe conditions, and dementia detection [3]. Given the
additional symptoms develop. These impact of AD on individuals, families,
include confusion, disorientation, and and healthcare systems, some
getting lost in familiar places. AD is studies are using mathematical
0360-8581 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See ht_tps://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: VTU Consortium. Downloaded on February 24,2024 at 06:56:59 UTC from IEEE Xplore. Restrictions apply.
EARLY DETECTION OF ALZHEIMER'S DISEASE USING COGNITIVE FEATURES: A VOTING-BASED ENSEMBLE MACHINE LEARNING APPROACH 17

modeling to predict the trend and machine (SVM), AdaBoost, and the presents the proposed ML approach.
growth of AD. These studies consider convolutional neural networks [12] Section IV presents the results.
some contributors or factors such as are also used to predict dementia. Section V presents the performance
increased life expectancy, alteration Generally, ML approaches learn the evaluation of the model. Finally,
in mortality, and cardiovascular in-depth neuropsychological patterns Section VI presents the concluding
diseases. Unfortunately, these of AD patients in training samples remarks.
studies suggest that an increasing after utilizing these learned samples
percentage of people will be affected to predict AD in other patients. These
II. RELATED WORK
by AD [4]. For example, it is projected ML algorithms are usually blind during
in 2030; there will be between 400 the training phase. They are trained ML is a potential and growing
000 and 459 000 Australians on all patients in a given dataset technology for detecting cognitive
diagnosed with dementia [5]. except for those used to make the diseases in its initial stages [13], [17],
Similarly, there will be a 57% rise in prediction. On the other hand, ML [18], [19]. ML used screening tests to
the number of people with AD in techniques used to predict AD predict AD in [14]. Other studies, such
England and Wales between 2016 resemble clinical situations. In clinical as the one reported in [15], proposed
and 2040, with more than 1.2 million situations, the primary pathology of a using the K-nearest neighbor (KNN)
people diagnosed with dementia by new patient is predicted using the network to predict and classify AD. In
2040, even after accounting for the neuropsychological tests' results. ML [16], an ML-based decision-making
expected improved life expectancy models employ these same tests to computer-aided design (CAD) tool
[4]. In the United States, estimates predict dementia [13]. was also presented to predict the
indicate that the American population early stages of AD. In fact, KNN, an
with AD or severe cognitive To this end, this article proposes ML approach, has been used to
impairment will reach 15 million by using an adaptive voting-based ML predict and categorize AD. To detect
2060, a significant increase from ensemble model to predict AD in the AD, ML-based decision-making
2017’s figure of 6.08 million cases [6]. early stages. Several ML classifiers systems that rely on CAD methods
were trained on the cognitive features were also suggested [16]. A study
On the other hand, recent data show filtered and supplied by the novel [19] suggested using an ensemble
that AD is declining in some countries proposed Neighborhood Component classification technique to identify
with advanced healthcare systems, Analysis and Correlation-based people with high and low AD risks.
such as the United Kingdom [7], [8]. Filtration (NCA-F) method [45]. The When only paralinguistic features
For example, a comparison of two proposed method incorporates all were used to train the model, it
decades regarding people aged 65 classifier’s advantages for various achieved a classification accuracy of
years and older from three types of disease detection. The 94.7%. Conversely, integrating
geographical regions of England model produced comparatively better paralinguistic and episodic memory
shows a decrease in the age-specific results using the ensemble learning features increased model
occurrence of AD [7]. In addition, the classifier. The ensemble learning performance by 2.5% (attained 97.2%
age-standardized mortality rate combines the performance of accuracy while trained by applying
(AMSR) for deaths because of the predictions, rather than discrete combined features). The ensemble of
dementia and AD in England and labels, of all the base classifiers to 13 ML models has been developed to
Wales in 2019 was 115.1 per 100 000 improve the generalization process predict the progression of mild
people, which was statistically and the robustness over a single cognitive impairment to AD [20].
significantly lower than ASMR in estimator. The results demonstrate According to the analysis, the
2018—123.8 per 100 000 people (69 significant improvements in the ensemble had an area under curve
478 deaths) [7]. Recent studies have performance of AD prediction by (AUC) of 0.88, a specificity of 79.9%,
demonstrated that early detection and using the ensemble feature selection and a sensitivity of 77.7%,
intervention plans can reduce the risk approach. Furthermore, the results respectively. A study in [21]
of dementia or slow down the demonstrate that by integrating the suggested an early AD detection
disease’s progression [8]. proposed new approach, which uses utilizing Cambridge cognitive
cognitive features only, into the ML features, which test a variety of
Machine learning (ML) methods use models, a better prediction of AD at cognitive processes. One new
neuropsychological tests to predict early stages was achieved. criterion for detecting AD is
AD [9]. Other emergent ML presented, based on the weighted
approaches such as the stacked The remainder of this article is final score from multiple tests [22].
autoencoder [10], deep belief organized as follows. Section II The experiments are carried out using
networks [11], support vector presents the related work. Section III principal components analysis and

Authorized licensed use limited to: VTU Consortium. Downloaded on February 24,2024 at 06:56:59 UTC from IEEE Xplore. Restrictions apply.
18 IEEE ENGINEERING MANAGEMENT REVIEW, VOL. 51, NO. 1, FIRST QUARTER, MARCH 2023

discriminant analysis. A research AUC of 0.79 using 4 cognitive increases the generalization and
attempted to examine the capabilities components. Weakley et al. [29] robustness over a single classifier.
of ML in classifying the AD procedure considered the capability of the ML This study selects seven ML
detection and optimizing or reducing regarding conventional statistical classifiers such as AdaBoost, artificial
the number of cognitive tests [13]. methods in completely automating neural network (ANN), SVM, and NB
the detection procedure by € Bayes) to detect the early AD.
(Naõve
Filtering, wrapper, and embedding minimizing the time-overwhelming Based on the comparative test
techniques are common features and subjective physical analysis and scores, five top-performing classifiers
extraction methodologies in the producing consistent information on were selected. Afterward, various
literature [23]. The features are the connection between inputs parameters of the model were
chosen based on the results of (cognitive measures) and response adjusted to improve the detection
various statistical analyses and the variables without the need of defining performance of the model. Finally, the
resulting are derived by correlating assumptions on data. adaptive voting model with different
with a target variable. The correlation weights is used to acquire the optimal
coefficient, which specifies a detection results. The adaptive
statistical relationship between the
III. PROPOSED ML METHOD ensemble learning model is depicted
variables, is applied to calculate the Ensemble models are ML techniques in Figure 1, which shows the various
final scores. Other procedures, such that combine several base models to processes. Yi shows the output, and
as Pearson’s correlation, linear construct one optimal model [38]. Wci represents the weight of model i
discriminant analysis, analysis of Several classifiers are trained using for the respective class (normal or
variance, and the Chi-square the entire feature space, and the final diseased).
approach [24], employ this correlation classifier is selected with the highest
coefficient as a feature identifier. weight contribution [39], [40], [41], A novel proposal, the NCA-F method,
Zhou et al. [25] suggested a unique [42]. Ensembling multiple classifiers is used to select the essential
dementia detection method based on is also widely used to improve the cognitive features for AD detection.
a three-stage deep feature learning detection accuracy of ML models. However, redundant features can
and fusion methodology. Yang et al. The adaptive ensembling of the also be the reason for the overfitted
[26] suggested a unique feature classifiers has gained much attention, ML method. This concern has been
weighting approach based on the which adaptively selects the classifier resolved by using a selection of
closest neighbor approach that during the inference process [43], independent features. In the first step,
outperforms previous benchmark [44]. This study proposes an adaptive feature filtration is carried out using
methods in terms of classification ensemble learning model that Pearson’s correlation method. A
accuracy. Other studies, such as incorporates the benefits of various single feature with a correlation value
those described in [27], focused on ML classifiers for detection and greater than 0.9 is retained in the
cognitive features to predict AD. Ford achieving optimal results using data, whereas the remaining features
et al. predicted dementia using a set ensemble learning. Since the data or are discarded. In the second step,
of 18 cognitive features. The area base ML models are changing, the features are ranked, and the best F
under the receiver operating weight matrix (used for voting) is also features are selected from all filtrated
characteristic (ROC) curve was 0.74 changing accordingly; hence the features in the first steps. After
[28]. Gill et al. [27] utilized cognitive voting is adaptive depending on the feature selection, normalization of the
features to detect AD. The study performance of base ML models. The data is achieved using a standard
obtained 81.8% accuracy and an adaptive ensemble learning scalar formula, resulting in good
prediction performance of the models.

Seven different ML classifiers are

trained on the cognitive features
resultant from the NCA-F method.
Out of seven classifiers, the top five
classifiers are selected based on the
weight matrices which are generated
from the F1 scores, including the
Decision Tree (DeciTree), ANN, SVM,
Random Forest (RanForest), and
AdaBoost ensemble, respectively. A
Figure 1. Adaptive Ensemble learning model. weight matrix for five selected

Authorized licensed use limited to: VTU Consortium. Downloaded on February 24,2024 at 06:56:59 UTC from IEEE Xplore. Restrictions apply.
EARLY DETECTION OF ALZHEIMER'S DISEASE USING COGNITIVE FEATURES: A VOTING-BASED ENSEMBLE MACHINE LEARNING APPROACH 19

classifiers is presented in Table 1. This ADNI1 [30]. The ADN1 dataset (NCA)-based wrapper technique [35]
weight matrix should be of size C
M includes the cognitive test scores and is used to assign weights to the
where C is the number of classes and values for 5013 records selected cognitive features. The
M is the ML chosen model. The voting corresponding to 819 different AD irrelevant features were identified first
process is completed according to patients. During the clinical trials, and discarded from the dataset by
the weight matrices. In Figure 1, the various patients used to visit the clinic defining a threshold. The highly
output label of model 1 is multiplied multiple times. Each time an AD correlated features, which have a
with the F1 score, as weight, of the patient visits the clinic to do the test correlation value greater than 0.9,
same model. The weight matrix’s as part of the trial, a new cognitive were filtered. Finally, a correlation
corresponding values associated with test score is generated and saved as heat map was created using
the predicted class are appended to a new record in the dataset since the Pearson’s correlation. The correlation
the predicted labels of all classifiers. patients repeat the cognitive tests in coefficient matrix (CRM) is defined as
The new result (Res) for each record all their visits. There are 1643 a matrix for all pairwise variable
(rec x) belonging to a class is Cognitive Normal (CN) and 3370 AD combinations, as given in
calculated by records in the ADN1 dataset.  
However, the dataset has missing rða; aÞ rða; bÞ
CRM ¼ (2)
X
C X
M
values and contains irrelevant rðb; aÞ rðb; bÞ
Res ðrec xÞ ¼ max Wðm;cÞ (1) features. Consequently, cleaning and
c ¼ 1m ¼ 1
data preprocessing were performed. where rða; bÞ indicates the correlation
The dataset used in this study has coefficient between features a and b
where Wðm;cÞ is the F1 score (weight)
many missing values in different and is computed as given in
of model m for class c. The maximum
value for a record is considered a new cognitive features. Therefore, the COV ða; bÞ
class for the record. Finally, the new cognitive features with missing values r ða; bÞ ¼ (3)
sa sb
predicted labels for all the records are of more than 40% were discarded
considered, and performance from the data. The remaining missing where s a and s b indicate mean and
measures are calculated. The records values are retained using the iterative standard deviation, whereas
here indicate the patient data. The imputer method [33], filling them COV ða; bÞ indicates the covariance
methodology of the proposed ML based on the residual values. This function, respectively After filtration,
method to predict early AD follows the technique imputes the missing values the selected cognitive features were
following processes, as shown in of the features in a round-robin further processed by the NCA. In the
Figure 2. method. In addition, the dataset proposed NCA-F feature selection, the
contained some features, which had NCA with stochastic gradient descent
several values. Therefore, the [36] optimizer is used to assign
IV. DATASET AND FEATURES
dataset’s dynamic range was weights to the selected cognitive
DETECTION normalized between 0 and 1 using features. The NCA-F adopted a one-
The AD Neuroimaging Initiative the min–max method [34]. Primarily, nearest neighbor classifier [37] to
(ADNI) [30], [31], [32] dataset is used the cognitive features were filtered study several cognitive feature
in this study, composed of three using Pearson’s correlation combinations and assign weights to
stages. This study has focused on coefficient. After that, the all the selected features. The selected
data from the first stage known as neighborhood component analysis weighted cognitive features are then

Table 1. Weight-Matrix for Five Selected Classifiers.

ADB_f1_sc ANN_f1_sc svm_f1_sc RF_f1_sc DT_f1_sc
AD 0.90 0.90 0.89 0.93 0.92
CN 0.78 0.78 0.78 0.87 0.84

Figure 2. Overview of the working flow of the proposed approach.

Authorized licensed use limited to: VTU Consortium. Downloaded on February 24,2024 at 06:56:59 UTC from IEEE Xplore. Restrictions apply.
20 IEEE ENGINEERING MANAGEMENT REVIEW, VOL. 51, NO. 1, FIRST QUARTER, MARCH 2023

arranged in descending order and where measures. The better results are
processed by ML models to predict  true positive (TP) is the positive achieved with the adaptive voting
AD. The resultant cognitive features predicted value (required); technique that outscored the
are given in Table 2. The accuracy,  true negative (TN) is the positive competing ML models. Results on the
precision, recall, and F1 scores are predicted value (required); selected features are better than all
used as performance measures to  false positive (FP) is the negative the features because of irrelevant and
assess the ML models. The relevant predicted value (false alarm); repetitive features removal using the
equations are as follows:  false negative (FN) is the proposed NCA-F. The trained model
negative predicted value (the achieved the followings performance
TP þ TN worst), respectively. metrics: accuracy>93%, precision
Accuracy ¼ (4Þ
TP þ TN þ FP þ FN >93%, and recall>93%, respectively.
TP The lowest performance was
SensitivityðRecallÞ ¼ (5Þ
TP þ FN V. RESULTS AND DISCUSSIONS observed with LR in terms of
TP accuracy and F1 score. For example,
F1 score ¼ (6) After selecting the optimized cognitive
TP þ 2 ðFP þ FNÞ
1 adaptive voting has improved the
feature combinations required for the
accuracy from 82.05% with LR to
early detection of AD, seven different
93.92% and achieved an 11.87%
ML models were evaluated, which
Table 2. List of Selected Cognitive Features
performance gain. Similarly, the recall
include DeciTree, RanForest, KNN,
After NCA-F Scheme. is improved from 82.05% with LR to
LR, SVM, ANN, AdaBoost, and
S. No Feature Weight 93.92% with adaptive voting. The
adaptive voting. The performance of
1 EcogSPDivatt 0.0033972
accuracy is improved from 85.68%
the selected models is evaluated for
2 EcogPtVisspat 1.07E-07 with SVM to 93.92% and achieved an
3 RAVLTpercforgetting 1.51E-12
the selected cognitive features in
8.08% performance gain. Moreover,
4 EcogPtPlan 7.03E-16 terms of AD detection. The
the proposed approach achieved a
5 RAVLTlearning 8.71E-24 experiments are validated with a
6 RAVLTforgetting 9.52E-47 7.28% accuracy gain over AdaBoost,
7 LDELTOTAL 1.1009
fivefold cross-validation set. A ratio of
which can be considered a significant
8 EcogPtLang 1.0243 80:20 training and the testing dataset
9 FAQ 0.91021
improvement.
is considered in the experiments. The
10 MOCA 0.79935
11 EcogPtTotal 0.75844
top eighteen (18) weighted cognitive
12 EcogPtDivatt 0.72765 features selected by the proposed In addition, F1-scores have improved
13 ADAS11 2.0056 NCA-F are used to examine and from 88.93% with DeciTree and
14 RAVLTimmediate 1.9877 90.33% with RanForest to 93.89%
15 ADASQ4 1.7356
detect early AD. Table 3 provides the
16 MMSE 1.565 results achieved by various ML with adaptive voting. Thus, it
17 EcogPtOrgan 1.5628 models, which were trained using the achieved a considerable performance
18 EcogSPVisspat 1.5223
18 selected cognitive features gain of 4.93% and 3.56%,
19 EcogSPMem 0.33403
20 EcogPtMem 0.14011 compared to all the available features respectively. The lowest errors (in
21 TRABSCOR 0.024946 in the dataset. The results are based percentage) were achieved with the
22 AGE 15.288 adaptive voting model, which is
23 EcogSPPlan 3.0912
on the Accuracy, Precision, Recall,
24 PTGENDER 2.9801 F1-Score, and errors of the models. It 6.08%. The errors have significantly
25 CDRSB 2.6719 is comprehensible that the proposed improved over other ML models; for
26 EcogSPOrgan 2.345 example, the most significant errors
27 EcogSPLang 2.18
approach shows significant
improvement in performance were observed with LR (17.95%),

Table 3. Performance Comparison of Selected and All Features.

Algorithm All features results (in %) Selected features from NCA-F results (in %)
Accuracy Precision Recall F1 Error Accuracy Precision Recall F1 Error
DeciTree 86.43 85.36 84.9 85.12 13.57 88.93 88.93 88.93 88.93 11.07
RanForest 88.53 87.8 87.13 87.45 11.47 90.33 90.30 90.33 90.31 9.67
KNN 84.85 83.09 83 83.04 15.15 85.14 85.40 85.14 85.24 14.86
LR 84.05 82.6 81.17 81.79 15.95 82.05 81.83 82.05 81.45 17.95
SVM 85.84 84.79 83.03 83.79 14.16 85.84 85.68 85.84 85.71 14.16
DNN 87.04 87.08 83.42 84.83 12.96 90.53 90.67 90.53 90.59 9.47
AdaBoost 85.24 85.09 81.86 83.14 14.76 86.64 86.61 86.64 86.31 13.36
Adaptive voting 90.23 90.25 90.23 90.04 9.77 93.92 93.89 93.92 93.89 6.08

The bold values indicate better performance in terms of results.

Authorized licensed use limited to: VTU Consortium. Downloaded on February 24,2024 at 06:56:59 UTC from IEEE Xplore. Restrictions apply.
EARLY DETECTION OF ALZHEIMER'S DISEASE USING COGNITIVE FEATURES: A VOTING-BASED ENSEMBLE MACHINE LEARNING APPROACH 21

which is 11.87% more than the AdaBoost (13.36%) were 8.08% and voting with the lowest model are
adaptive voting. Similarly, the errors 7.28% more than the adaptive voting. depicted in Figure 3. The LR is the
observed with SVM (14.16%) and The overall improvements of adaptive reference lowest performance ML
model.

The correlation matrix of the 18

weighted cognitive features is
provided in Figure 4. A high
correlation can be observed among
the cognitive features in pair
variables. The correlation matrix
shows that there is a strong positive
correlation between EcogPtDivatt
and EcogPtTotal cognitive features.
Figure 3. Percentage improvements in accuracy, precision, recall, and F1-score.
Also, there is a moderate positive
correlation between EcogPtOrgan
and EcogPtVisspat cognitive
features. A strong correlation can
be observed between the features
having greater absolute values.

The confusion matrix heat-map for

the proposed adaptive voting is
provided in Figure 5. The AUC of
various ML models is provided in
Table 4, which shows that the
proposed method with adaptive
voting technique achieved the
highest AUC, 99.3%. Finally, the
ROC curves for various ML models
are depicted in Figure 6, which
shows the outperformance of the
proposed adaptive voting model.

A. Comparison With State of the

Art This section compares the
proposed method with some existing
filtrations methods, which also used
the ADNI dataset. The comparison
aims to examine and benchmark the
proposed approach against the
existing methods under the same
experimental settings and using the
Figure 4. Correlation matrix for selected 18 cognitive features. same dataset. First, the performance
of the proposed method is compared
with a current state-of-the-art method
proposed in [27]. The experimental
results obtained with the proposed
method and the existing filtration
method on feature configurations are
presented in Table 5.

The experimental results are

presented in terms of selected
Figure 5. Confusion matrix for adaptive voting. features, the number of records, and

Authorized licensed use limited to: VTU Consortium. Downloaded on February 24,2024 at 06:56:59 UTC from IEEE Xplore. Restrictions apply.
22 IEEE ENGINEERING MANAGEMENT REVIEW, VOL. 51, NO. 1, FIRST QUARTER, MARCH 2023

accuracy. It is noticeable from Table 5 Similarly, the AUC of the ensemble comparison between the complete
that the proposed adaptive ensemble model has increased by 20%, that is, set of features versus selected set of
model has outperformed the from 79% with Gill et al. [27] to 99% features from the NCA-F method,
performance of the model proposed by with the proposed method with proposed. The selected set of
Gill et al. [27] and all the results adaptive voting technique. The features resulted in 94.72% of
reported from all the 30 cognitive results also show that cognitive accuracy while complete features
features are available in the dataset. features (clinical tests) play a set’s accuracy is 90.23%. The reason
All the results are evaluated using a significant role in the early detection for better performance for the
fivefold cross-validation set. The of AD. Table 3 indicates the additional reduced/selected set of features is
accuracy of the ensemble model using experimental results where we have that irrelevant features. Such features
the selected cognitive features on the compared the performance of feature are considered noise for ML models
ADNI1 dataset produced the better selection. We have compared the because anything repetitive, extra, or
performance results (93.92%). In results achieved with the full features irrelevant is considered noise. When
addition, there is a 12.12% set and the selected features. The we removed the irrelevant features
improvement in terms of accuracy results clearly indicate the superiority from the complete set of features,
when compared with the competing of feature selection over the entire then the model training and fitting are
method. features set. Table 3 shows the more optimized toward the
classification.
Table 4. Performance Evaluation of Models Using Selected Cognitive Features in Terms of
AUC.
VI. CONCLUSION
Voting-based ensemble AdaBoost Ensemble Neural network SVM RanForest
99.3% 96.19% 96.2% 93.52 97.94
The NCA-F-based novel feature
selection technique is proposed to
select vital cognitive features, and
various ML classifiers are trained
based on selected features to obtain
novel results in a short period for
early AD detection. The adaptive
voting technique with different
weights is used to get the optimal
detection results in accuracy,
precision, recall, F1-score, AUC, and
ROC. The experimental results
indicate that the proposed approach
has achieved better results for
various performance measures. The
key idea of the proposed method is
to exploit the ensemble learning to
get the benefits of various ML
models. Ensemble learning is helpful
to improve the early detection of AD.
The accuracy of the ensemble
models with adaptive voting
technique is 93.92%, which is
improved compared to other models.
Figure 6. ROC curves of various ML models for selected cognitive features.
Similarly, the AUC is increased from
79% to 99%. The experiments
Table 5. Performance Comparison of the Proposed Method.
indicate that the cognitive features
Features Model Dataset Selected Number Accuracy (%) AUC (clinical tests) are essential to detect
features of records
the early AD using ensemble ML
Cognitive features Gill et al. [27] ADNI1 4 600 81.80 79.0
Adaptive voting ADNI1 18 5013 93.92 99.3
models. Furthermore, ML has
Adaptive voting ADNI1 All (30) 5013 90.23 95.9 apparent advantages in AD detection
since it is less time-consuming than
The bold values indicate better performance in terms of results.
clinical trials.

Authorized licensed use limited to: VTU Consortium. Downloaded on February 24,2024 at 06:56:59 UTC from IEEE Xplore. Restrictions apply.
EARLY DETECTION OF ALZHEIMER'S DISEASE USING COGNITIVE FEATURES: A VOTING-BASED ENSEMBLE MACHINE LEARNING APPROACH 23

REFERENCES
[1] Alzheimer’s Association. (2016). Alzheimer’s disease facts and figures.
Alzheimer’s & Dementia, 12(4), 459–509.
[2] Tarawneh, R. and Holtzman, D. M. (2012). The clinical problem of symptomatic
Alzheimer’s disease and mild cognitive impairment. Cold Spring Harbor
Perspectives in Medicine, 2(5), a006148.
[3] World Health Organization. Dementia: Fact sheet, 2017. Accessed: Jan. 4,
2021. [Online]. Available: https://www.who.int/news-room/fact-sheets/detail/
dementia
[4] Ahmadi-Abhari, S. et al (2017). Temporal trend in dementia incidence since
2002 and projections for prevalence in England and Wales to 2040: Modelling
study. BMJ, 358, j2856.
[5] Australian Institute of Health. (2012). Australia’s Health 2012: The Thirteenth
Biennial Health Report of the Australian Institute of Health and Welfare.
Sydney, Australia: AIHW.
[6] Brookmeyer, R., Abdalla, N., Kawas, C. H., and Corrada, M. M. (2018).
Forecasting the prevalence of preclinical and clinical Alzheimer’s disease in the
United States. Alzheimer’s & Dementia, 14(2), 121–129.
[7] Matthews, F. E. et al (2016). A two decade dementia incidence comparison
from the cognitive function and ageing studies I and II. Nature
Communications, 7(1), 1–8.
[8] Sullivan, K. J. et al (2019). Declining incident dementia rates across four
population-based birth cohorts. The Journals of Gerontology: Series A, 74(9),
1439–1445.
[9] Gurevich, P., Stuke, H., Kastrup, A., Stuke, H., and Hildebrandt, H. (2017).
Neuropsychological testing and machine learning distinguish Alzheimer’s
disease from other causes for cognitive impairment. Frontiers in Aging
Neuroscience, 9, 114.
[10] Martinez-Murcia, F. J., Ortiz, A., Gorriz, J. M., Ramirez, J., and
Castillo-Barnes, D. (Jan. 2020). Studying the manifold structure of Alzheimer’s
disease: A deep learning approach using convolutional autoencoders. IEEE
Journal of Biomedical and Health Informatics, 24(1), 17–26.
[11] Shen, T. et al (2019). Predicting Alzheimer disease from mild cognitive
impairment with a deep belief network based on 18F-FDG-PET Images.
Molecular Imaging, 18, 1536012119877285.
[12] Feng, C. et al (2019). Deep learning framework for Alzheimer’s disease
diagnosis via 3D-CNN and FSBi-LSTM. IEEE Access, 7, 63605–63618.
[13] Battista, P., Salvatore, C., and Castiglioni, I. (2017). Optimizing neuropsychological
assessments for cognitive, behavioral, and functional impairment classification: A
machine learning study. Behavioural Neurology, 2017, 1850909.
[14] Youn, Y. C. et al (2018). Detection of cognitive impairment using a machine-
learning algorithm. Neuropsychiatric Disease and Treatment, 14, 2939.
[15] Kruthika, K. R., Maheshappa, H. D., and Alzheimer’s Disease Neuroimaging
Initiative. (2019). Multistage classifier-based approach for Alzheimer’s disease
prediction and retrieval. Informatics in Medicine Unlocked, 14, 34–42.
[16] Veeramuthu, A., Meenakshi, S., and Manjusha, P. S. (2014). A new approach
for Alzheimer’s disease diagnosis by using association rule over pet images.
International Journal of Computer Applications, 91(9), 9–14.
[17] Kang, M. J. et al (2019). Prediction of cognitive impairment via deep learning
trained with multi-center neuropsychological test data. BMC Medical
Informatics and Decision Making, 19(1), 1–9.

Authorized licensed use limited to: VTU Consortium. Downloaded on February 24,2024 at 06:56:59 UTC from IEEE Xplore. Restrictions apply.
24 IEEE ENGINEERING MANAGEMENT REVIEW, VOL. 51, NO. 1, FIRST QUARTER, MARCH 2023

[18] Zhu, F. et al (2020). Machine learning for the preliminary diagnosis of dementia.
Scientific Programming, 2020, 5629090.
[19] You, Y., Ahmed, B., Barr, P., Ballard, K., and Valenzuela, M. (2012). Predicting
dementia risk using paralinguistic and memory test features with machine
learning models. in Proceeding IEEE Healthcare Innovations and Point of Care
Technologies, pp. 56–59.
[20] Grassi, M. 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 information, and
neuropsychological measures. Frontiers in Neurology, 10, 756.
[21] Schmand, B., Walstra, G., Lindeboom, J., Teunisse, S., and Jonker, C. (2000).
Early detection of Alzheimer’s disease using the Cambridge Cognitive
Examination (CAMCOG). Psychological Medicine, 30(3), 619–627.
[22] Chapman, R. M. et al (2010). Diagnosis of Alzheimer’s disease using
neuropsychological testing improved by multivariate analyses. Journal of
Clinical and Experimental Neuropsychology, 32(8), 793–808.
[23] Guyon, I., Gunn, S., Nikravesh, M., and Zadeh, L. A., Eds. (2008). Feature
extraction: Foundations and Applications, vol. 207. Berlin, Germany: Springer-
Verlag.
[24] Fisher, R. A. (1992). Statistical methods for research workers. in Breakthroughs
in Statistics. Berlin, Germany: Springer-Verlag, pp. 66–70.
[25] Zhou, T., Thung, K. H., Zhu, X., and Shen, D. (2019). Effective feature learning
and fusion of multimodality data using stage-wise deep neural network for
dementia diagnosis. Human Brain Mapping, 40(3), 1001–1016.
[26] Yang, W., Wang, K., and Zuo, W. (2012). Neighborhood component feature
selection for high-dimensional data. Journal Computational, 7(1), 161–168.
[27] Gill, S. et al. (2020). Using machine learning to predict dementia from
neuropsychiatric symptom and neuroimaging data. Journal of Alzheimer’s
Disease, 75(1), 277–288.
[28] Ford, E. et al (2019). Identifying undetected dementia in UK primary care
patients: A retrospective case-control study comparing machine-learning and
standard epidemiological approaches. BMC Medical Informatics and Decision
Making, 19(1), 1–9.
[29] Weakley, A., Williams, J. A., Schmitter-Edgecombe, M., and Cook, D. J. (2015).
Neuropsychological test selection for cognitive impairment classification: A
machine learning approach. Journal of Clinical and Experimental
Neuropsychology, 37(9), 899–916.
[30] Weiner, M. W. et al. (2010). The Alzheimer’s disease neuroimaging initiative:
Progress report and future plans. Alzheimer’s & Dementia, 6(3), 202–211.
[31] Petersen, R. and Weiner, M. W. (2014). Alzheimer’s disease neuroimaging
initiative 2 (ADNI2) protocol (ADC-039). [Online]. Available: https://adni.loni.
usc.edu/wp-content/themes/freshnews-dev-v2/documents/clinical/ADNI-
2_Protocol.pdf.
[32] Weiner, M. W. et al. (2017). The Alzheimer’s disease neuroimaging initiative 3:
Continued innovation for clinical trial improvement. Alzheimer’s & Dementia, 13
(5), 561–571.
[33] Van Buuren, S. and Groothuis-Oudshoorn, K. (2011) mice: Multivariate
imputation by chained equations in R. Journal of Statistical Software, 45, 1–67.
[34] Zang, I. (1980). A smoothing-out technique for min–max optimization.
Mathematical Programming, 19(1), 61–77.
[35] Yang, W., Wang, K., and Zuo, W. (2012). Fast neighborhood component
analysis. Neurocomputing, 83, 31–37.
[36] Ketkar, N. (2017). Stochastic gradient descent. in Deep Learning With Python.
Berkeley, CA, USA: Apress, pp. 113–132.

Authorized licensed use limited to: VTU Consortium. Downloaded on February 24,2024 at 06:56:59 UTC from IEEE Xplore. Restrictions apply.
EARLY DETECTION OF ALZHEIMER'S DISEASE USING COGNITIVE FEATURES: A VOTING-BASED ENSEMBLE MACHINE LEARNING APPROACH 25

[37] Kontorovich, A. and Weiss, R. (2015). A Bayes consistent 1-NN classifier. in

Proceedings of the 18th International Conference on Artificial Intelligence and
Statistics, pp. 480–488.
[38] Zhang, C., and Ma, Y., Eds. (2012). Ensemble Machine Learning: Methods and
Applications. Berlin, Germany: Springer-Verlag.
[39] Jacobs, R. A., Jordan, M. I., Nowlan, S. J., and Hinton, G. E. (1991). Adaptive
mixtures of local experts. Neural Computation, 3(1), 79–87.
[40] Woods, K., Kegelmeyer, W. P., and Bowyer, K. (Apr. 1997). Combination of
multiple classifiers using local accuracy estimates. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 19(4), 405–410.
[41] Alpaydin, E. and Jordan, M. I. (May 1996). Local linear perceptrons for
classification. IEEE Transactions on Neural Networks, 7(3), 788–794.
[42] Giacinto, G. and Roli, F. (2001). An approach to the automatic design of multiple
classifier systems. Pattern Recognition Letters, 22(1), 25–33.
[43] Inoue, H. (2019). Adaptive ensemble prediction for deep neural networks based
on confidence level. in Proceeding 22nd International Conference on Artificial
Intelligence and Statistics, pp. 1284–1293.
[44] Wang, C., Deng, C., Yu, Z., Hui, D., Gong, X., and Luo, R. (2021). Adaptive
ensemble of classifiers with regularization for imbalanced data classification.
Information Fusion, 69, 81–102.
[45] Irfan, M., Shahrestani, S., and Elkhodr, M. (2021). Early detection of the
Alzheimer’s disease: A novel cognitive feature selection approach using
machine learning. in Proceeding International Conference on Information,
Communication Cybersecurity, pp. 383–392.

Authorized licensed use limited to: VTU Consortium. Downloaded on February 24,2024 at 06:56:59 UTC from IEEE Xplore. Restrictions apply.