Early Diagnosis of Alzheimer’s Disease using Machine Learning

Techniques

A Review Paper

Aunsia Khan and Muhammad Usman

Dept. of Computing, Shaheed Zulfikar Ali Bhutto Institute of Science and Technology (SZABIST), Islamabad, Pakistan

Keywords: Alzheimer’s Disease, Machine Learning, Computer Aided Diagnosis, Pathologically Proven Data, Early

Diagnosis, Class Imbalance.

Abstract: Alzheimer’s, an irreparable brain disease, impairs thinking and memory while the aggregate mind size shrinks

which at last prompts demise. Early diagnosis of AD is essential for the progress of more prevailing

treatments. Machine learning (ML), a branch of artificial intelligence, employs a variety of probabilistic and

optimization techniques that permits PCs to gain from vast and complex datasets. As a result, researchers

focus on using machine learning frequently for diagnosis of early stages of AD. This paper presents a review,

analysis and critical evaluation of the recent work done for the early detection of AD using ML techniques.

Several methods achieved promising prediction accuracies, however they were evaluated on different

pathologically unproven data sets from different imaging modalities making it difficult to make a fair

comparison among them. Moreover, many other factors such as pre-processing, the number of important

attributes for feature selection, class imbalance distinctively affect the assessment of the prediction accuracy.

To overcome these limitations, a model is proposed which comprise of initial pre-processing step followed

by imperative attributes selection and classification is achieved using association rule mining. Furthermore,

this proposed model based approach gives the right direction for research in early diagnosis of AD and has

the potential to distinguish AD from healthy controls.

1 INTRODUCTION

Alzheimer’s disease (AD), a type of dementia, is

characterized by progressive problems with thinking

and behavior that starts in the middle or old age. The

pathologic characteristics are the presence of neuritic

plaques in the brain and degeneration of explicit brain

cells. The symptoms usually develop slowly and get

serious enough to interfere in daily life. Although the

paramount risk factor is oldness but AD is not just an

old age disease. In its early stages, the memory loss is

mild while in the later stages, the patient’s

conversation and their ability to respond degrades

dramatically. The current treatments cannot stop

Alzheimer’s disease (AD) from developing but early

diagnosis can aid in precluding the severity of the

disease and help the patients to improve the quality

life. It has been reported that the number of individuals

effected with AD will double in next 20 years (Zhang,

2011), while in 2050, 1 out of 85 individuals will be

effected (Ron Brookmeyer, 2007). Thus the accurate

diagnosis especially for the early stages of AD is very

important.

Machine learning is used to interpret and analyze

data. Furthermore it can classify patterns and model

data. It permits decisions to be made that couldn’t be

made generally utilizing routine systems while sparing

time (Mitchell T, 1997) and endeavors (Duda RO,

2001). Machine learning methodologies have been

extensively used for computer aided diagnosis in

medical image formation mining (Supekar, 2008) and

retrieval (Bookheimer, 2000) with wide variety of

other applications (Cruz, 2006) especially in detection

and classifications of brain disease using CRT images

(Cruz, 2006) and x-rays (Petricoin, 2004) It has just

been generally late that AD specialists have

endeavored to apply machine learning towards AD

prediction. As a consequence, the literature in the field

of Alzheimer’s disease prediction and machine

learning is relatively small. However, today’s imaging

technologies and high throughput diagnostics have

lead us overwhelmed with large number (even

hundreds) of cellular, clinical and molecular

parameters. In current circumstances, the standard

measurements and human instinct don’t frequently

work. That is the reason we must depend on

380

Khan, A. and Usman, M..

Early Diagnosis of Alzheimer’s Disease using Machine Learning Techniques - A Review Paper.

In Proceedings of the 7th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2015) - Volume 1: KDIR, pages 380-387

ISBN: 978-989-758-158-8

intensively computational and non-traditional

approaches such as machine learning. The custom of

using machine learning as a part of disease prediction

and visualization is a fragment of an expanding shift

towards prescient (Weston, 2004) and customized

prescription (Cruz, 2006). This drift is important, not

only for the patients in increasing their quality of life

and life style, but for physicians in making treatment

decisions and also for health economists.

In evaluating and analyzing the existing studies, a

number of common trends and gaps has been

identified. The most evident trends include a rapid

growth in the AD detection and prognosis using

machine learning methods. Among the major gaps

was an imbalance of events with attributes (few

instances and too many attributes), the use of

pathologically unproven data set (which cause

uncertainty in results), class imbalance (too few

instances in one class and too many instances in other

class), overtraining and lack of external testing or

validation. Nevertheless, the better designed and

validates studies made it clear that machine learning

methods, in comparison to standard statistical

methods, could improve the accuracy of AD

prediction. Besides, machine learning play an

important role in AD prediction and prognosis.

To overcome these limitation, a model is proposed

for effective diagnosis of onset of AD. While

considering the pathologically proven data set, the

proposed model involve a pre-processing step for

eliminating the class imbalance issue. Important

attributes selection using machine learning method

help avoiding the problem of too few instances and too

many attributes, known as curse of dimensionality

(Cruz, 2006). The model divides the dataset into

training and testing data. Training data on a limited

testing data leads to a phenomenon of over-training

(Chaves, 2010). Thus, training data should be selected

to span a representative fragment of the actual data.

The model presents classification using association

rule mining with minimum support and minimum

confidence. The paper is organized in a manner that

Section II describes the different machine learning

techniques. Section III and IV describes the literature

review and critical evaluation. Proposed model is

explained in Section V. Finally conclusions are drawn

in Section VI.

1.1 Machine Learning Methods

Before starting the detailed analysis of machine

leaning methods, it is significant to have a better

understanding of what actually machine learning is

and what machine learning techniques are commonly

used AD prognosis. Machine learning comes under

the umbrella of artificial intelligence and has variety

of tools to make statistical, probabilistic decisions

based on previous learning. It uses past learning

(training) to classify new event and predict new

patterns. Machine learning is very powerful as

compared to standard statistical tools. In machine

learning, a good understanding of a problem and

limitations of the algorithms are needed to be

understood well to get effective results. Therefore, it

has a good chance for success if an experimentation is

properly conducted and training is carefully and

correctly employed and results are vigorously

validated. Furthermore, all the algorithms and

methods in machine learning are somewhat made

different. For instance, few methods are designed on

the basis of certain assumptions or for certain type of

data which make it inapplicable for other type of data.

That is why it is crucial to apply more than one

machine learning method on given training data.

Machine learning generally have three types of

learning algorithms: 1. Supervised learning 2.

Unsupervised learning (Duda RO, 2001) 3.

Reinforcement learning (Mitchell T, 1997). In

supervised learning, a training data is given whereas

the program tries to learn it and learns how to draw the

input to the required output. The unsupervised

learning algorithms employs self-learning based on

unclassified and unlabeled data. Interestingly, the

algorithms used in AD prognosis and diagnosis are

almost all supervised learning algorithms including

Artificial Neural Networks, Decision Trees, genetic

algorithms and linear discriminant analysis.

Other techniques which are generally in use are

SVM, AR mining, and Ensemble methods. In

comparison to the above, SVM or support vector

machine is somewhat newer technique (Duda RO,

2001) and is world known machine learning technique

now but it is almost unidentified in AD prognosis

field. The other methods such as KNN (K-Nearest

Neighbors) and DTs (decision trees), are not widely

used in AD predictions. Although, many high quality

papers were studied for this review. However, almost

all of them lacked a valid proven dataset for AD,

lacked external or internal validation, were using too

many attributes (causing over training) and no well-

defined standard was made with which results were

compared. These issues are further discussed in

Section IV.

Early Diagnosis of Alzheimer’s Disease using Machine Learning Techniques - A Review Paper

381

2 LITERATURE REVIEW

A detailed study on classification and diagnosis of AD

has been proposed by many researchers. This section

includes a brief review of the related work.

2.1 Single Modality Approach

The computer aided diagnosis of AD at the early stage

of dementia is more challenging that lead R. Chaves

et al., (2010) to introduce a classification method for

effective and early diagnosis of Alzheimer's disease.

Using association rule mining, they found out the

associations between attributes of the pre-processed

data sets. The proposed method was based on the tri-

dimensional activated brain regions of interests

(ROIs). These ROIs were obtained through a series of

steps such as voxels of each image were considered as

features (VAF) and the activation estimation using a

certain threshold. For this purpose, a SPECT dataset

of 97 instances was used out of which 43 were normal

controls and remaining 54 were AD patients. The

authors made comparisons with other techniques like

VAF, PCA-SVM and GMM-SVM, and results

revealed a classification accuracy of 95.87% (100%

sensitivity, 92.86 specificity) with a claim of reducing

the computational cost. This results show negligible

difference in the accuracies with better efficiency in

terms of computational time. The author claim it to be

an “Effective” approach rather than efficient diagnosis

of AD.

Distinguishing the early stage of the disease in AD

patients using clinical conventions remained a

diagnostic challenge. R. Chaves et al. (2011), later on,

continued with his work by finding the associations

among attributes while characterizing the perfusion

patterns in SPECT images of normal subjects. For this

purpose, complete image database was evaluated to

reproduce the knowledge of medical experts. The

pathologically unproven dataset from ADNI of 97

participants was used, of which 41 were labeled as

healthy controls and 56 were labeled as AD patients

by expert physicians. Comparisons were made with

other techniques like PCA-SVM, GMM-SVM, output

revealed the classification accuracy of 94.87% with

91.07% sensitivity and 100% specificity. The class

imbalance was minimized as possible while the results

were based on pathologically unproven data with no

discussion about missing values.

The pathological unproven data sets of AD, made

it applicable to different imaging technologies, as

well, to diagnose other neuro-degenerative diseases.

To address this, R. Chaves et al. (2012) introduced a

mining technique using association rule mining

defined over discriminant regions using pre-processed

SPECT and PET imaging modalities. 97 participants

contributed for the datasets, 42 were labeled as healthy

controls and 55 were labeled as AD patients by expert

physicians. The proposed method was compared with

other techniques like PCA-SVM, VAF-SVM and

results of this paper out proved them with accuracy of

92.78% with 87.5% sensitivity & 100% specificity for

SPECT and 91.33% accuracy with 82.67% sensitivity

& 100% specificity for PET. With no discussion about

the missing values, the class imbalance have been

reduced.

The study by Veeramuthu et al. (2014) developed

a CAD tool for decision making about the presences

of abnormalities in human brain. The author suggested

preprocessing of PET dataset for instance, spatial

normalization and intensity normalization. Fisher

Discriminants ratio (FDR) was used for feature

extraction to get ROIs. The instances were classified

to normal if the extracted number of verified rules

were above the final threshold otherwise image was

classified as AD. The authors claimed 91.33%

accuracy with 82.67% sensitivity & 100% specificity

in comparison with other methods as VAF,

PCA+SVM, and NFM+ SVM. It is observed that the

authors did not mention the number of instances used

in dataset. The methods adopted for dealing the

missing data and class imbalance are also ignored. The

dataset taken for the proposed study is not

pathologically proven. Support and confidence,

effective parameters of AR mining, are not discussed

as well as no method for validation has been

mentioned by the authors.

R. Chaves et al. (2012), impressed from the

findings of PET data, tried to improve the prediction

accuracy of the AD especially in early stage which has

been of most concern to the researchers. The aim was

the improvement in diagnosis of AD using Apriori AR

progression and to develop new treatments and

monitor their effectiveness while reducing the

computational time and cost of clinical trials. The

authors have introduced a method for analyzing of

Alzheimer’s disease by incorporating more detailed

PET for instance, FDG-PET and PiB-PET. The data

set comprised of 103 participants where 19 were

control (CTRL), 19 were AD patients and 65 were

with Mild cognitive impairment (MCI). The authors

came with good results for PiB PET having

classification accuracy of 97.37% and in combination

with FDG it achieved the classification accuracy of

94.74% while FDG PET alone received 92.11%

accuracy. The proposed method worked with a very

small sized pathologically unproven data set with a

class imbalance problem which produces uncertainty

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

382

in the acquired accuracies.

Similarly, Liu, Zhang et al. (2012) also contributed

for early diagnosis of AD by implementing ensemble

sparse method for the classification. The study

revealed that noise and small sample size is very

challenging to achieve good classification accuracy.

As cited by the author, the high feature dimensionality

will probably reduce the classification capability with

standard classifier models, such as linear discriminant

analysis, SVM and decision trees. The proposed study

used Sparse representation-based classifier (SRC) to

generate local patch based classifiers which are fused

later on to give more robust and accurate

classification. The authors found out that individual

sub classifier can be easily trained thus dimensionality

to subject ratio can be substantially improved. The

study revealed that if the patches are from AD regions

then classification accuracy will be high otherwise it

will be low. Furthermore, pathologically unproven

data set and class imbalance will demonstrate

uncertainty in results.

Chaves et al., (2013) elaborated the early diagnosis

later on by discretizing the continuous attributes of

feature selection. Mean of control images were used

to obtain a mask using histogram segmentation. AR

mining used those RIOs as input and Control subject

images were used to fully characterize the normal

pattern of the image. The data of 97 participants was

collected for SPECT, of which 41 were normal

controls while 56 were AD. Moreover for PET, data

of 150 participants was collected which comprised of

75 AD and 75 healthy controls. The results revealed

96.61% accuracy for SPECT while 92% accuracy for

PET while comparison was made with VAF-SVM and

PCA-SVM. To the best of our knowledge, it is the

highest accuracy achieved for SPECT so far.

However, the missing values have not been considered

and mentioned in this study while the data used for this

study was unproven which may degrade the results.

Klöppel et al, (2008) used the structural MRI to

distinguish Alzheimer’s disease from healthy controls

at early stage. The authors applied SVM to MRI for

the reliable detection of disease while distinguishing it

from normal aging. This research was based on the

pathologically proven data sets, collected from

different centers as an input for effective

classification. Finally, proposed method was

implemented using normalized datasets from 67 AD

and 91 healthy controls from different scans. Using

the whole brain images, 96% of AD patients, who

were pathologically verified, were correctly classified

using leave one out cross validation. The proposed

research showed generalization by combining data

from different centers however, the data set is too

small for fair comparisons with other methods.

2.2 Multimodal Approach

Although the use of different single biomarkers yield

promising results but they are designed to characterize

group differences and are not for individual

classification. D. Zhang et al. (2011) came up with a

method of combing all the three biomarkers for

Alzheimer’s disease diagnosis i.e. MRI, PET, CSF etc.

to discriminate between healthy and AD participants.

The authors made use of baseline data set with total

202 instances, out of which 51 were AD, 99 were MCI

and 52 were healthy controls. Different tests were

conducted for MRI, PET and CSF and the

combination of these using 10 fold cross validation.

The classification accuracy of 93.2% with 93%

sensitivity and 93.3% specificity was achieved with

combination of these modalities while individual test

yielded highest accuracy of 86.5%. Authors claimed

that multimodal classification method (using all MRI,

PET, and CSF) achieves consistent improvement and

is more robust over those using individual modality,

for any number of brain regions selected.. These

results directed that CSF and PET have the highest

complementary information, while MRI and PET

have the highest similar information for classification.

Furthermore, it is noted that the availability of data of

individual subject on all the modalities is too small for

reasonable classification. The knowledge of missing

values and how they are handled are not mentioned in

this study. Class imbalance is another prominent

limitation in this paper.

In support to the above, Westman, Muehlboeck et

al. (2012) studied the combination of baseline MRI

and CSF data to enhance the classification of AD

while making comparison to individual modality. The

data from 369 participants was collected to study

regional subcortical volumes and cortical thickness

measures. The data set comprised of 96 AD and 273

healthy controls, labeled by expert physicians. As

cited by the author, FDG-PET can be expensive and it

would have been interesting to see how the method of

Zhang et al. performed with just the combination of

MRI and CSF, but this data was not presented.

Orthogonal partial least squares to latent structures

(OPLS) multivariate analysis was used for 60

variables (57 from MRI and 3 from CSF). The

proposed method resulted in classification accuracies

of 91.8% for combined MRI and CSF which is slightly

lower than those of (Cuingnet R1, 2011). The study

also revealed that SVM and LDA have previously

been utilized by others while OPLS showed more

Early Diagnosis of Alzheimer’s Disease using Machine Learning Techniques - A Review Paper

383

Table 1: Summary and Critical Evaluation of techniques and limitations of different machine learning based AD studies.

Modality Technique Data Set

Details

Pathologically

proven Data

set

Accuracy Limitation

Validation

performed

(No. of Folds )

(Chaves,

Ramírez et al.

2013)

SPECT

PET

Apriori- AR

mining

SPECT:

AD = 56

CTRL = 41

PET:

AD = 75

CTRL = 75

No SPECT:

96.91%

PET: 92%

Pathologically

unproven data with no

justification about

missing values

Leave one out

Cross

Validation

(Klöppel,

Stonnington et

al. 2008)

MRI Linear SVM 3-groups

AD= 67

CTRL= 91

Yes

96%

Sample size is too small

with no justification of

missing values.

Leave one out

Cross

Validation

(Chaves,

Ramírez et al.

2010)

SPECT Apriori- AR

mining

AD = 54

CTRL = 43

95.87%

Did not mention the

how they limited the

effect of missing values

Leave one out

Cross

Validation

(Chaves, Górri

et al. 2011)

SPECT Apriori- AR

mining

AD = 56

CTRL = 41

94.87%

The data may contain

Missing values which

will cause uncertainty

Leave one out

Cross

Validation

(Chaves,

Ramirez et al.

2012)

FDG- PET

PiB-PET

Apriori- AR

mining

AD = 19

CTRL = 84

94.74%

Unproven data with

missing values

Leave one out

Cross

Validation

(Zhang, Wang

et al. 2011)

MRI+ FDG-

PET + CSF

SVM

AD = 51

CTRL = 151

93.2%

Class Imbalance and

missing values

10-fold Cross

Validation

(Chaves,

Ramírez et al.

2012)

SPECT

PET

Apriori- AR

mining

SPECT:

AD = 55

CTRL = 42

PET:

AD = 75

CTRL = 75

92.78%

Unproven data with

missing values

Leave one out

Cross

Validation

(Westman et

al., 2012)

CSF

MRI

Apriori- AR

mining+ SVM

AD = 96

CTRL = 273

91.8%

Class Imbalance and

missing values

7-fold Cross

Validation

(Chaves,

Ramírez et al.

2012)

SPECT

PET

Apriori- AR

mining for

feature

selection PCA,

SVM

SPECT:

AD = 56

CTRL = 41

PET:

AD = 75

CTRL = 75

No 91.75% Unproven data with

missing values

Leave one out

Cross

Validation

A. Veeramuth

et al. (2014)

PET AR mining Not Given No 91.33%

No dataset details,

missing values or any

preprocessing steps

highlighted

Robi Polikar et

al. (2010)

EEG + MRI

+ PET

Ensemble

based decision

fusion

AD = 37

CTRL = 36

No 85.55% Unproven data with

missing values

5-fold Cross

Validation

similarities with SVM except for the ability to separate

structured noise from the correlated variation

modeling. Previous studies like (O. Kohannim, 2010)

has shown that the combination of MRI and CSF

significantly improves classification accuracy.

However, CSF measures are highly invasive and could

cause distress for patients which may provide a basis

for combination of MRI and PET rather than MRI and

CSF. Furthermore, the data set is not pathologically

proven and author did not mention anything regarding

missing data which may decrease the overall accuracy

of the proposed method.

Polikar, Tilley et al. (2010) supported the use of

CSF biomarker for being most promising in early

diagnosis of AD. In contrast to that they revealed the

costly and highly invasive nature of CSF biomarker

along with its potentially painful lumber puncture. The

proposed study investigated the fusion of non-invasive

biomarkers such as PET, MRI as well as EEG to check

their feasibility for the early diagnosis of Alzheimer’s

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

384

disease. Using ensemble method, each classifier was

trained on each datasets from different sources.

Classifiers were then combined using an appropriate

combination rule (Parikh, D. and R. Polikar, 2007).

The Sum and simple majority voting (SMV) rules

were used to obtain the data fusion diagnostic

accuracies. Followed by the 5-fold cross validation,

the outcome indicated the classification accuracy of

85.55% which is 10% -20% improvement as

compared to fusion of any of two mentioned

modalities. The Ensemble method is promising

(Westman, 2012) however, the resultant accuracy is

below as compared to the accuracies achieved in

previous studies (Chaves, 2013). Although the authors

reduced the class imbalance effect but they did not

mentioned how they dealt with missing data.

3 CRITICAL EVALUATION

A detailed study on the early diagnosis and

classification of AD has been proposed by many

researchers. This segment contains a brief critical

review and analysis of the related work.

3.1 Limitations

These studies outlined in the previous section are just

few examples of how well machine learning

experiments should be conducted and obviously there

are other good and equally impressive studies with

good results. These studies exemplified how the

outcomes should be validated and described especially

in the prognosis and prediction of AD. However,

being able to identify the potential issues wither in the

input data, experimental design, validation or the

implementation is very critical especially for those

who evaluate different studies as well as for those

aiming to use machine learning.

Through the analysis of the above studies in this

review, the most common problems among them were

the input size, attributes and validation. It is easier to

get higher accuracies with smaller datasets, such

methods could not be used to represent larger

population of data. It has been noted that small sample

size is prone to overtraining and large data size ensures

several effects on robustness, accuracy and

reproducibility. It is impressive that 96.6% accuracy is

attained, but unproven data used as input and the given

size of the data put some doubt on the robustness of

the model. Most of the research is done using

pathologically unproven data which consequently

may introduce uncertainty in the results. While such

data can be obtained from specialist centers so no

reason of not using such data have been identified in

the related studies. The attributes to instance ratio also

effects the results. In the above studies, lack of

attention paid towards the number and general

information of attributes.

Data quality and important attribute selection is

also very important for effective results generation in

machine learning. Unfortunately, the authors rarely

described the methods used to ensure the data integrity

and quality. Feature selection is also too important as

data quality. However, the features chosen for some

clinical data, for instance histological assessments,

may not be applicable over time. Therefore, a

classifier must be able to update feature sets with

respect to time. Similarly, the details about the training

and testing data should be clearly mentioned. Most of

the algorithms focus more on the classification of

major class whereas misclassify or ignore the minority

class. Such class imbalance results in choosing the

dominant class with poor class prediction, damaging

the quality of classification.

4 PROPOSED MODEL

The proposed method consists of four steps as

presented in Fig. 1: 1. Pre-processing, 2. Attribute

selection, 3. Classification and 4. Class Threshold

For effective classification of the AD data, the first

step is preprocessing. The pathologically proven data

set is processed to avoid class imbalance and then it is

converted to readable data type. Machine learning

algorithms works very well when the number of

instances of one class are almost equal to the number

of instances of other class. Class imbalance damage

the classification result severely so to avoid class

imbalance, data is over sampled using machine

learning technique for instance, synthetic minority

oversampling technique (SMOTE). The input data

type is converted from numeric into nominal/numeric

to nominal values so that the algorithms which uses

said data type as input can be implemented.

Attribute selection involves searching through all

possible combinations of attributes in the data to find

which subset of attributes works best for prediction

and classification. It is helpful in the dimensionality

reduction and omitting improper attributes. For

classification tasks, it can lead to increased accuracy

or to reduced computational costs. The third step is

based on classification using AR mining with

minimum support and minimum confidence.

Classification is done using 10-fold cross validation

that is, data is divided into 10 parts. One part is used

as test and remaining 9 are used as training data and

Early Diagnosis of Alzheimer’s Disease using Machine Learning Techniques - A Review Paper

385

the process is repeated 10 times to validate the results.

The training set is used for classification in order to

identify the specific parameters. The association rules

results in unique associations among the attributes

which are exploited in next step. In the last step, a

certain threshold is used over the resultant rules to

classify the instances into one of the two classes such

as Control and AD.

Figure 1: A proposed model for early detection of AD.

5 CONCLUSIONS

This study is based on the comparison and evaluation

of recent work done in the prognosis and prediction of

Alzheimer’s disease using machine learning methods.

Explicitly, the recent trends with respect to machine

learning has been revealed including the types of data

being used and the performance of machine learning

methods in predicting early stages of Alzheimer’s. It

is obvious that machine learning tends to improve the

prediction accuracy especially when compared to

standard statistical tools. However, based on the

review, the clinical diagnosis were not 100% accurate

as pathological verification was not provided which

consequently introduce uncertainty in the predicted

results. The proposed method deals with

pathologically proven data and overcomes the class

imbalance and overtraining issues. Proposed model is

based on single modality to overcome the increased

cost of computing and combining different modalities.

We believe that pathologically proven data may

increase the accuracy and validity, while a balanced

class will help the classifiers to give accurate results.

This model is can help to improve the prediction

performance by physicians and cover the limitations

pointed out in the previous research.

REFERENCES

D, Zhang, 2011. Multimodal classification of Alzheimer's

disease and mild cognitive impairment. Neuroimage,

55(3), 856-867.

Brookmeyer, Ron , Elizabeth Johnson, Kathryn Ziegler-

Graham, and H. Michael Arrighi, 2007. Forecasting the

Global Burden of Alzheimer's Disease. Alzheimer's and

Dementia, 3.3, 186-191.

Brookmeyer, Duda RO, Hart PE, Stork DG. , 2001. Pattern

classification. Alzheimer'(2nd edition). New York:

Wiley.s and Dementia, 2nd edition.

T, Mitchell , 1997. Machine Learning. New York: McGraw

Hill, 2nd edition, 135.

SY, Bookheimer, Strojwas MH, Cohen MS, Saunders AM,

Pericak-Vance MA, Mazziotta JC, 2000. , et al. Patterns

of brain activation in people at risk of Alzheimer’s

disease. N Engl J Med, 6, 343:450.

Supekar, K, Menon V, Rubin D, Musen M, Greicius MD,

2008. Network analysis of intrinsic functional brain

connectivity in Alzheimer’s disease. . PLoS Comput

Biol , 4(6), 1–11.

Cruz, J.A. and D.S. Wishart, 2006. Applications of

Machine Learning in Cancer Prediction and

Prognosis.Cancer Informatics, 2, 59-77.

EF, Petricoin, Liotta LA. , 2004. SELDI-TOF-based serum

proteomic pattern diagnostics for early detection of

cancer. Curr Opin Biotechnol, 15, 24-30.

Bocchi, L, Coppini G, Nori J, Valli G, 2004. Detection of

single and clustered microcalcifications in

mammograms using fractals models and neural

networks. Med Eng Phys, 26, 303-12.

Weston, AD, Hood L., 2004. Systems biology, proteomics,

and the future of health care: toward predictive,

preventative, and personalized medicine. . J Proteome

Res, 3, 179-96.

Bellman, R, 1961. Adaptive Control Processes: A Guided

Tour. Princeton University Press, 1, 45.

Rodvold, DM, McLeod DG, Brandt JM, 2001. Introduction

to artificial neural networks for physicians: taking the

lid off the black box. Prostate, 46, 39-44.

Chaves, R., J. Ramírez, et al., 2010. Effective Diagnosis of

Alzheimer’s Disease by Means of Association Rules.

Hybrid Artificial Intelligence Systems, Springer, 1, 452-

459.

Chaves, R., J. Górriz, et al. (2011). Efficient mining of

association rules for the early diagnosis of Alzheimer's

disease. Physics in medicine and biology 56(18): 6047.

Chaves, R., J. Ramírez, et al. (2012). Association rule-

based feature selection method for Alzheimer’s disease

diagnosis. Expert Systems with Applications 39(14):

11766-11774.

Chaves, R., J. Ramírez, et al. (2012). Functional brain

image classification using association rules defined

over discriminant regions. Pattern Recognition Letters

33(12): 1666-1672.

Veeramuthu, A., S. Meenakshi, et al. (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.

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

386

Chaves, R., J. Ramirez, et al. (2012). FDG and PIB

biomarker PET analysis for the Alzheimer's disease

detection using Association Rules. Nuclear Science

Symposium and Medical Imaging Conference

(NSS/MIC), IEEE.

Liu, M., D. Zhang, et al. (2012). Ensemble sparse

classification of Alzheimer's disease. Neuroimage

60(2), 1106-1116.

Hinrichs, C., V. Singh, et al. (2009). Spatially augmented

LPboosting for AD classification with evaluations on

the ADNI dataset. Neuroimage 48(1), 138-149.

Cuingnet R1, G. E., Tessieras J, Auzias G, Lehéricy S,

Habert MO, Chupin M, Benali H, Colliot O; (2011).

Automatic classification of patients with Alzheimer's

disease from structural MRI: a comparison of ten

methods using the ADNI database. Neuroimage 56,

766–781.

Chaves, R., J. Ramírez, et al. (2013). Integrating

discretization and association rule-based classification

for Alzheimer’s disease diagnosis. Expert Systems with

Applications 40(5), 1571-1578.

Klöppel, S., C. M. Stonnington, et al. (2008). Automatic

classification of MR scans in Alzheimer's disease.

Brain 131(3), 681-689.

Westman, E., J.-S. Muehlboeck, et al. (2012). Combining

MRI and CSF measures for classification of

Alzheimer's disease and prediction of mild cognitive

impairment conversion. Neuroimage 62(1),229-238.

Westman, Eric, et al. (2011). Multivariate analysis of MRI

data for Alzheimer's disease, mild cognitive impairment

and healthy controls. Neuroimage 54.2, 1178-1187.

O. Kohannim, X. Hua, D.P. Hibar, S. Lee, Y.Y. Chou, A.W.

Toga, C.R. Jack Jr., M.W. Weiner, P.M. Thompson

(2010) Alzheimer's Disease Neuroimaging Initiative

Boosting power for clinical trials using classifiers based

on multiple biomarkers, Neurobiology of Aging, 31 (8),

1429–1442.

Polikar, R., C. Tilley, et al. (2010). Multimodal EEG, MRI

and PET data fusion for Alzheimer's disease diagnosis.

Engineering in Medicine and Biology Society (EMBC),

2010 Annual International Conference of the IEEE,

IEEE.

Parikh, D. and R. Polikar (2007). An ensemble based

incremental learning approach to data fusion. Systems,

Man, and Cybernetics, Part B: Cybernetics, IEEE

Transactions, 37(2), 437-450.

Early Diagnosis of Alzheimer’s Disease using Machine Learning Techniques - A Review Paper

387