Neuropsychological features evaluation of data related to Alzheimer’s disease progression using feature selection

Abstract

BACKGROUND:

Alzheimer’s Disease (AD) is normally assessed in clinical settings using neuropsychological tests and medical procedures such as neuroimaging techniques: Magnetic Resonance Imaging (MRI) and Positron Emission Tomography (PET) among others. The latter procedures are expensive and unavailable in most nations, so early diagnosis of AD does not occur, which heavily increases the subsequent treatment costs for the patients.

AIMS:

This research aims to evaluate cognitive features related to dementia progression based on neuropsychological tests’ data that are related to the Alzheimer’s Disease Neuroimaging Initiative (ADNI) We utilise data related to two neuropsychological tests including the Clinical Dementia Rating Scale Sum of Boxes (CDR-SOB), and Mini-Mental State Examination (MMSE), to assess advancement of the AD.

METHODS:

To achieve the aim, we develop a data process called Neuropsychological Feature Assessment via Feature Selection (NFAFS) to identify impactful features using Information Gain (IG) and Pearson Correlation to assess class-feature and feature-feature correlations Later we will model a minimal subset of neuropsychological features using machine learning techniques to derive classification models.

RESULTS AND IMPLICATIONS:

Results obtained show key cognitive features of the MMSE are Time Orientation, Recall and Complex Attention, since they correlate with the progression class being ranked high in results of the feature selection techniques. For the CDR-SOB features, and aside from the memory feature it was difficult to identify other specific features that are signs of the dementia progression Clinicians can use specific features in a digital knowledge base to pay more attention to specific cognitive deficits related to Recall, Orientation and Complex Attention during dementia clinical evaluations in order to seek possible signs of the disease progression early.

Keywords

Alzheimer’s disease neuropsychological features computational intelligence dementia feature selection

1. Introduction

Worldwide, approximately 50 million people have dementia, with 10 million new cases diagnosed every year (World Health Organization, 2020). The most common form of dementia, accounting for 60–80% of cases, is Alzheimer’s Disease (AD) [5] with approximately 200,000 Americans under 65 affected [3]. AD affects memory, thinking and behaviour – sometimes its symptoms are mistaken for stress or aging. AD a progressive condition that evolves over time is classified in three stages: Early, Middle and Late [75]. In its early stages, memory loss is mild, but with late-stage AD, individuals lose the ability to carry on a conversation and respond to their environment [59]. A common early symptom of AD is difficulty remembering newlylearned information due to changes in the part of the brain that affects learning [4]. As AD progresses symptoms become increasingly severe including disorientation, mood and behaviour changes, confusion about events, time and place, unfounded suspicions about family, friends and caregivers, more serious memory loss, and difficulty speaking, swallowing and walking [52].

Neuropsychological evaluation using cognitive/functional tests such as the Alzheimer’s Disease Assessment Scale – Cognitive Subscale [ADAS-Cog] [46], Clinical Dementia Rating Scale [CDR] [39], Functional Activities Questionnaire [FAQ] [55], MMSE [27] and others have been used to diagnose Mild Cognitive Impairment (MCI) and dementia. These tests mainly evaluate different cognitive and memoryrelated areas by providing the patients with questions and activities within clinical or in some cases non-clinical environments. These tests are effective in the early screening of dementia; however, determining neuropsychological features or activities within these tests during dementia progression is a challenging task [66, 62, 11].

One of the promising approaches for screening of AD is the use of Artificial Intelligence (AI) particularly Machine Learning (ML). ML techniques extract hidden patterns from historical data using search methods that can aid diagnosticians in AD space between screening related such as the progression stage of patients [72, 10, 26, 74]. Recent research works show that using ML techniques can improve the screening and diagnosis of AD, i.e., Hinrichs et al. [36], So et al. [65], Samper-González et al. [63], Grassi et al. [30], Das et al. [22], Thabtah et al. [67], Thabtah et al. [68, 69], and AlShboul et al. [1]. However, primary issues that recent studies have not addressed thoroughly are which cognitive features can be symptoms of dementia progression, and the primary cognitive areas that these features belong to according to the Diagnostic and Statistical Manual (DSM-5) [6]. To fill this gap, this research investigates cognitive features that impact on the progression of AD in dissimilar neuropsychological assessments (MMSE, CDR-SOB) using large real data observations from the ADNI data project [47, 2]. We would like to answer the below research question:

How can we determine cognitive features from the MMSE and CDR-SOB tests using feature selection techniques?

We propose an experimental that are signs of the dementia progression methodology in a data process called Neuropsychological Feature Assessment via Feature Selection (NFAFS) to assess different indicators in neuropsychological assessments from real data given in ADNI (ADNI-Merge, MMSE-sheet, CDR-SOB-sheet) using feature selection. The proposed methodology firstly models the data to present the AD progression problem (creating the progression class label), balances the data with respect to the class label, processes the data observations, and more importantly evaluates each neuropsychological test’s features using feature selection techniques. We model a new class variable that we call Diagnostic Change (DX_change) which holds information of the progression of dementia at any stage. More details on NFAFS are given in Section 3.

The paper is structured as follows: Section 2 highlights the background on the neuropsychological tests used for AD screening and critically analyses relevant research works. In Section 3, we present the NFAFS and discusses the data used and features. Section 4 is devoted to experimental analysis, and finally, the conclusion and future works are given in Section 5.

2. Literature review

2.1 Background on dementia and cognitive assessments

The current standard for the diagnosis of dementia conditions is the DSM-5 framework of which there are six cognitive domains associated with dementia: Complex Attention, Executive Function, Learning and Memory, Language, Perceptual-Motor and Social Cognition [6]. Major neurocognitive disorder (Major ND) necessitates that the patient exhibits a substantial decline over time in at least one of the six cognitive areas of the DSM-5. Minor neurocognitive disorder (Minor ND) requires a moderate decline in cognitive areas over time and that the patient’s independence during everyday activities is not affected. Both major ND and minor ND also require that the cognitive declines be not only observed when the patient is delirious, and the cognitive defects are not better explained by another mental disorder.

Normally, when AD diagnosis is utilised for research purposes, rigorous diagnosis criteria are needed. In the ADNI project [2], AD subjects have been diagnosed as having probable AD according to the Neurological and Communicative Disorders and Stroke (NINCDS) and the Alzheimer’s Disease and Related Disorders Association (NINCDS-ADRDA) criteria [45]. This involves a MMSE score between 20 and 26 and an original CDR score of 0.5 or 1.0. The NINCDS-ADRDA uses eight cognitive domains: Attention, Problem Solving, Orientation, Functional Abilities, Constructive Abilities, Memory, Language, and Perceptual Skills. While a definite AD diagnosis is possible through a biopsy or autopsy, a probable diagnosis can be given when cognitive impairments are present in two or more domains and have shown to progressively decline along with the age of onset between 40 and 90 years and no other diseases are present that could account for the symptoms [45].

The DSM-5 classifies MCI as a mild cognitive decline that does not yet deprive individuals of leading an independent lifestyle or ability to perform complex daily activities such as driving a car. Major cognitive impairment is diagnosed when there is interference with daily living and there are no other culpable disorders [7]. While the DSM-5 does not name any specific diagnostic tests, there are a number of means that clinicians can use to provide a diagnosis of possible or probable AD. Since an AD diagnosis is reliant on the absence of other disorders, MRI and PET scans are useful for screening out other possible diagnoses, while neuropsychological tests such as MMSE CDR, and CDR-SOB are important tools for measuring impairments in the cognitive domains. Common neuropsychological tests that have been developed and used over the years by clinicians for dementia screening are shown in Table 1. This research scope is limited to CDR-SOB and MMSE.

MMSE is a commonly used screening test for early screening of symptoms related to dementia [27]. The test consists of multiple questions over eleven sections where patients score either a 1 or 0. The maximum score is 30 points and a screening is given based on where the patient scores on a table: A score of 20–24 suggests mild dementia, 13–20 suggests moderate dementia, and less than 12 indicates severe dementia

Table 1
Common neuropsychological tests

Test name	Abbreviation	Sensitivity %	Specificity %	Reference
Functional activities questionnaire	FAQ	85	81	[55]
Blessed orientation memory and concentration exam	BOMC	83	95	[40]
Alzheimer’s disease assessment scale – cognitive subscale	ADAS-Cog	89	88.5	[46]
Clinical dementia rating scale	CDR	79	96	[75]
Mini-mental state examination	MMSE	71–92	56–96	[27]
Montreal cognitive assessment	MOCA	90–100	87	[48]

The CDR-SOB is a clinical assessment for ranking dementia severity [9]. This test is performed by a clinician trained in an administering and interpreting the test results. In the CRD assessment, the patient is evaluated in six areas: memory, orientation, judgment and problem solving, community affairs, home and hobbies, and personal care. Scoring is 0: None, 0.5: Questionable, 1: Mild, 2: Moderate, 3: Severe. The Sum of Boxes scoring assessment (CDR-SOB) generates scores between 0 to 18 after adding up the CDR domain scores. Recommended cut-off scores are: 0.5–4: Questionable 4.5–9: Mild 9.5–15.5: Moderate to Severe, 1618 [77]. We used in this study CDR-SB data from ADNI study.

2.2 Related works

One of the earliest researches works to apply ML techniques for the prediction of AD is that by Datta et al. [23]. Looking to improve accuracy over BOMC and FAQ, they used data collected from 578 patients from the University of California’s Alzheimer’s Disease Research Centre and created three datasets using BOMC FAQ, and a combination of BOMC and FAQ. Features were selected manually from the BOMC and FAQ answers as well as age, sex, job, and education level. Of the ML techniques used the authors found that they were able to achieve the best classification accuracy (88.4%) using the Naïve Bayes classifier.

Devanand et al. [24] conducted research with 148 MCI patients over three years to predict the progression of MCI to AD. The authors used cognitive and other clinical tests in their dataset. Using backward and stepwise logistic regression [12], they achieved 90% specificity with the five features identified: informant report of functioning from FAQ, olfactory identification from the University of Pennsylvania Smell Identification Test [UPSIT] [25], immediate recall (verbal memory) from the Selective Reminding Test [SRT] [14], and MRI hippocampal volume as well as MRI entorhinal cortex volume. However, they also found that combining age, MMSE, SRT immediate recall, FAQ, and UPSIT provided 81.3% sensitivity.

Using data from 1,489 patients in the ADNI-1 dataset, Zhu et al. [79] aimed to predict changes in MMSE scores over 24 months. Using manually selected features (age, gender, education, MMSE evaluation, and APOE genotype), the authors developed their own prediction method – COMPASS (Computational Model to Predict the Development of Alzheimer’s Disease Spectrum), and compared this with linear regression [29], RBF [Radial Bias Function] Network [13], SMOreg [sequential minimal optimization – regression] [64], decision tree [56] and Gaussian Process Regression (Rasmussen, 2006). Their method outperformed these, and it achieved 80% accuracy for the top 20% of predictions but derived poor accuracy for MCI classification.

So et al. [65] investigated the problem of predicting dementia using ML with feature selection on two common datasets: MMSE-KC and CERAD-K [42]. The authors used Chi-squared testing [31] and Information Gain (IG) [56] techniques and 13 features. The top five features selected by both of the techniques were similar: Orientation to Place, Orientation to Time, Three-stage Commands, Recall, and Attention”. Simultaneously, the authors used multiple dissimilar classifiers to process the two datasets considered such as Support Vector Machine (SVM) [19], Naive Bayes, Multilayer Perceptron (MLP) [51] and others, and the results showed that the models derived by MLP and SVM are accurate to predict dementia at least on the two datasets considered.

Vinutha et al. [74] replaced the missing value using the imputation method with ML on the neuropsychological scores and the FAQ from the NACC database (National Alzheimer’s Coordinating Center, n.d) to improve the accuracy of classification. The authors conducted the analysis using two patient visits from datasets: Visit 1 and visit 4, this because approximately 50% of the progression from MCI to AD is concentrated in 24–36 months after the first visit. The authors used the Genetic Algorithm (GA) [76] and Logistic Regression to process the datasets after balancing the data with the SMOTE [17] method The results showed that paying bills, remembering appointments, and meal preparation in visit 1 and tax-paying records, remembering appointments, and paying attention in visit 4 are the main progression features.

Hsiao et al. [38] analysed longitudinal data related to the FAQ, focusing on MCI subjects and considering two possible scenarios: whether stable or progressive. Their previous studies suggesting that Instrumental Activities of Daily Living (IADLs) [43] may be helpful in distinguishing MCI from dementia, but not adequate to clarify mapping between longitudinal changes in IADLs and conversion of MCI to dementia. The authors assessed IADL performance using the FAQ and these assessments were split into 10 different categories according to their criteria. The authors focused on the recent longitudinal intervals. Changes in FAQ scores were calculated between the last two visits for stable MCI and between the point of a diagnosis of dementia conditions and two immediately preceding visits for progressive MCI. These two groups were compared employing unpaired $t$ tests and chi-square, to reveal that the progressive MCI group had greater longitudinal changes in FAQ scores. Their findings using Binomial Logistic Regression [20] revealed that significant increases in FAQ scores were spotted in progressive MCI subjects within 14.5 months after their initial visits and stressed that the difficulties in identifying determinative IADLs and specific cut-off points on FAQ indices in dementia.

Niyas et al. [50] suggested a feature selection algorithm based on fusion of the Fisher Score and the greedy searching heuristic approach. The authors applied SVM and k-NN techniques for the dementia screening and used the ADNI-TADPOLE (1737 patients) and ABIL (1000, Australian cohort) datasets. The authors selected a cross-sectional data at the baseline stage in the experiment. The selected features based on the algorithm were evaluated using Leave One Out Cross Validation (LOOCV) and stratified 10-fold cross-validation. The suggested algorithm provided good sensitivity (84%) and specificity (82.5%). Based on their experiment CDRSB, MMSE, ADAS-13, and AV45 from ADNI-TADPOLE were ranked highly important features to discriminate normal vs mild cognitive Alzheimer cases. In the ABIL dataset, the features with high importance are: LDELTOTAL (which assesses logical memory), CDGLOBAL (Clinical Dementia Rating Global), and MMSCORE. However, a recommendation by the authors was that the classifiers’ performance improved when combining highly important features with features of low importance. Furthermore, the researchers highlighted the potential role of cognitive tests in discriminating the MCI and AD cases. Examples of these tests are the Clinical Dementia Rating Global test, Memory Immediate Recall test (LIMMTOTAL) Partial Score of Logical Memory (LDELTOTAL), and MMSCORE which are promising features in classifying MCI vs AD.

Thabtah et al. [68, 69] assessed cognitive and functional features for two common medical questionnaires to identify how well machine learning technology can detect AD advancement. The authors used data related to the ADAS-13 and FAQ dementia assessments from the ADNI data project, and a few machine learning algorithms with feature selection. The results pinpointed that cognitive features are more strongly associated with the disease progression.

Recently, AlShbout et al. [1] evaluated the score of the CDR-SOB assessment and how well it correlates with the diagnosis of AD. The authors used several machine learning algorithms to process the ADNI-Merge dataset focusing on the baseline-demographic data with CDR-SOB score. The results indicated competitive performance when using machine learning algorithms in detecting MCI, Dementia or cognitively normal (CN) individuals. However, there was some variation in the evaluation metrics used showing some machine algorithms are better suited to dementia screening problems based on the baseline-demographic data.

Table 2 illustrates the summary of related works.

3. Methodology

3.1 Data, features, and methods

Figure 1 illustrates the methodology used in this research to consider the research problems raised earlier. In NFAFS, three datasets from the ADNI data repository: ADNI-Merge, MMSE data sheets (MMSE-feature), and CDR-SOB data sheets (CDR-SOB- feature), are processed using dissimilar feature selection techniques to identify key neuropsychological indicators. Each of the cognitive test’s data, i.e. CDR-SOB, MMSE, was integrated with ADNI-Merge to capture the diagnostic class using RID and visit code attributes. Some of the visits for a given patient (RID) were ignored when not appearing in the cognitive test data sheets. Since the problem under consideration is to assess the progression of dementia at any stage, this necessitates the creation of a new class variable for each dataset that we call DX_change. This class variable is created to record any progression from CN to MCI or MCI to dementia (AD) using the following rules:

1)
The data instances in each dataset are sorted based on the patient ID (RID) and visit date.

Table 2
Summary of related works

Feature selection technique Dataset Cohort size Reference

$\chi 2$ test, Kruskal-Wallis test, stepwise, and backward LR Clinical-based dataset collected by New York State Psychiatric Institute, Columbia University 148 Devanand et al. [24]

$-$ ADC dataset 578 Datta et al.[23]

$t$ -test, $\chi$ 2 test, and LR NACC UDS 3,183 (original), 2426 after removing reverted cases) Hsiao et al.[38]

Bayesian network to select genetic features, Pearson (Spearman’s) correlations, ADNI-1 dataset 489 Zhu et al.,[79]

$\chi^{2}$ test, and IG Patients of Gangbuk-Gu Center for Dementia in Seoul tested using MMSE-KC and CERAD-K Korea, Phase1: 14,000 patients; 9799 NC, 4201 CD Phase 2: 1236; 663 MCI, 573 AD So et al.[65]

Miss Forest Imputation, SMOTE and Genetic Algorithm (GA), LR Neuropsychological scores and FAQ from the NACC database 5281 Vinutha et al.[74]

Fusion of Fisher Score, and greedy searching heuristic approach ADNI-TADPOLE dataset ABIL Dataset ADNI-TADPOLE dataset (1737 patients) ABIL (1000, Australian cohort) Niyas et al.[50]

CSF, ReliefF, and two wrappers ADNI-Merge ADNI-Merge- CDR-SOB (8904) AlShbout et al.[1]

$\chi 2$ test, ReliefF, and IG ADNI-Merge $+$ ADAS-sheets ADNI-Merge $+$ ADAS-sheets a total of 6,770 of cognitive data observations Thabtah et al.[68]

$\chi 2$ test, PC, IG ADNI-Merge $+$ FAQ -sheets ADNI-Merge $+$ FAQ -sheets a total of 10,905 of functional data observations Thabtah et al. [69]

Figure 1.
Methodology Followed (NIAFE).

2)
For any two consecutive visits for the same patient (RID) if the original diagnostic class (DX) changes progressively (from CN to MCI or MCI to AD) then we record in DX_change ‘1’. However, if DX changes in a regressive manner, the DX_change is assigned $-$ 1. Otherwise if no change of DX then DX_change is assigned 0. We removed the regression instances, i.e. DX_change $=-$ 1.
3)
Whenever all visits of the current RID are considered the DX_change of the first visit for the next patient is automatically assigned 0 and we repeat steps 1–2 on the remaining visits for that patient until all of its visits are fulfilled.

The above steps are applied on the level of data observations (sorted visits for a patient) and for each test dataset.

Once each test’s data are integrated with ADNI-Merge, data instances that are linked with missing class, i.e. DX, are removed; we also exclude instances with DX_change $-$ 1 or null. More details on the data preprocessing phase can be found in Section 3.2. As soon as the data processing is completed, NFAFS implements three feature selection methods: Pearson Correlation (PC), IG [56], and Minimum Redundancy Maximum Relevance mRMR with Correlation Feature Set (CFS) [54, 33], to measure the rank of the available cognitive features in the neuropsychological tests’ data. These methods have been chosen for the following reasons:

1)
They adopt different mathematical models in the way they compute feature relevance
2)
They have been utilized previously in medical screening applications, i.e. Remeseiro & Bolon-Canedo [60]; Liu et al. [44]; Zhang et al. [78]
3)
They are highly efficient in the way they produce the feature relevance (time taken)
4)
They are available on different ML platforms such as Waikato Environment for Knowledge Analysis (WEKA) [28]. Rapid Miner (RapidMiner, n.d), R (R Foundation, n.d), and Python (Python Software Foundation, n.d).

The considered feature selection technique evaluates the associations between the progression class and the cognitive features in pairs. For instance, PC measures the independent relationship between two features and identifies the strength and signs of the relationship based on a computed correlation coefficient, i.e. $\pm$ 1. If the features are independent, the correlation coefficient is zero. The linear relationship for a pair of features (A, B) can be measured with $r$ which denotes the PC coefficient by:

$\displaystyle r=\frac{\sum\left(A_{i}-\bar{A_{i}}\right)(B_{i}-\bar{B_{i}})}{% \sqrt{(A_{i}-\bar{A_{i}})^{2}}\sqrt{(B_{i}-\bar{B_{i}})^{2}}}$ (1)

where $A$ and $B$ are the two numeric features. The magnitude of $r$ gives the strength of the relationship between the two features in which $r$ can be positive or negative.

IG measures the amount of information each available feature embraces to distinguish between the class variable values [56]. It has been used successfully in constructing decision tree classification systems by ML algorithms as a way to evaluate data split in reference to the class variable in the dataset. In doing so, IG relies on uncertainty reduction metrics such as Shannon entropy to estimate the worthiness of the data split using each feature in the training dataset by evaluating the information gained before the data split and after the data split. IG mathematical notations are shown below:

$\displaystyle\textit{IG}(D,A)=E(D)-\sum((|D_{a}|/|D|)*E(D_{a}))$ (2)

where

$\displaystyle E(T)=-\sum P_{r}\log_{2}P_{r}$ (3)

where $P_{v}=$ the likelihood that $D$ belongs to response variable $r$ .

$D_{a}$ $=$ subset of $D$ with $A$ equals $a$

$|D_{a}|$ $=$ the frequency of data observation in subset $D_{a}$ , and

$|D|$ $=$ the frequency of data observation inside $D$ .

mRMR utilises the maximum relevancy of the available features in the training dataset to the class variable and the minimum redundancy among the features themselves [54]. In other words, this technique tries to not only capture the relationship between the features and the class as a conventional filtering technique, but also minimises the similarities among these features. mRMR employs the F-Statistics metric to compute the features-class relevancy, and PC to compute any redundancy among the features using the below equations based on [54].

$\displaystyle f^{\textit{mRmR}}(A_{i})=\textit{MI}\left({R,X_{i}}\right)-\frac% {1}{\left|s\right|^{2}}\sum_{A\varepsilon S}\textit{MI}\left({A_{s},A_{i}}\right)$ (4)

Where,

$S$ $=$ A subset of chosen features

$\left|S\right|=$ The size of the chosen features

MI $=$ Mutual Information

$R$ $=$ Response variable

$A_{i}=$ Features that are not chosen

$A_{s}\varepsilon S=$ One feature out of the feature set ‘S’

In CFS, the feature goodness is identified without calculating weight/score per feature and only relevant features are selected using mutual information metrics.

In the feature selection phase of NFAFS, features that are a) highly correlated with the class variable and b) dissimilar in reference to the neurodegenerative domains they belong to in the DSM-5, are identified. This will provide diagnosticians with valuable knowledge of which neuropsychological features are more crucial in the screening process of AD progression. Since each of the feature selection techniques may provide a different range or scale when calculating the feature relevance we normalised the scores between 0–1 for the results of both PC and IG. In Section 5, a demonstration of the feature assessment results will be provided, illustrating how each feature subset was chosen In brief we analysed the results produced by the feature selection techniques in terms of common features, features rank, nonoverlapping features with reference to the DSM-5 cognitive domains among others.
3.2 Data preprocessing

Feature selection technique	Dataset	Cohort size	Reference
$\chi 2$ test, Kruskal-Wallis test, stepwise, and backward LR	Clinical-based dataset collected by New York State Psychiatric Institute, Columbia University	148	Devanand et al. [24]
$-$	ADC dataset	578	Datta et al.[23]
$t$ -test, $\chi$ 2 test, and LR	NACC UDS	3,183 (original), 2426 after removing reverted cases)	Hsiao et al.[38]
Bayesian network to select genetic features, Pearson (Spearman’s) correlations,	ADNI-1 dataset	489	Zhu et al.,[79]
$\chi^{2}$ test, and IG	Patients of Gangbuk-Gu Center for Dementia in Seoul tested using MMSE-KC and CERAD-K	Korea, Phase1: 14,000 patients; 9799 NC, 4201 CD Phase 2: 1236; 663 MCI, 573 AD	So et al.[65]
Miss Forest Imputation, SMOTE and Genetic Algorithm (GA), LR	Neuropsychological scores and FAQ from the NACC database	5281	Vinutha et al.[74]
Fusion of Fisher Score, and greedy searching heuristic approach	ADNI-TADPOLE dataset ABIL Dataset	ADNI-TADPOLE dataset (1737 patients) ABIL (1000, Australian cohort)	Niyas et al.[50]
CSF, ReliefF, and two wrappers	ADNI-Merge	ADNI-Merge- CDR-SOB (8904)	AlShbout et al.[1]
$\chi 2$ test, ReliefF, and IG	ADNI-Merge $+$ ADAS-sheets	ADNI-Merge $+$ ADAS-sheets a total of 6,770 of cognitive data observations	Thabtah et al.[68]
$\chi 2$ test, PC, IG	ADNI-Merge $+$ FAQ -sheets	ADNI-Merge $+$ FAQ -sheets a total of 10,905 of functional data observations	Thabtah et al. [69]

Data used in this research has been obtained from ADNI.2

²
Data used in the preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). The 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). For up-to-date information, see www.adni-info.org.

ADNI is a research project that started in 2004 in multiple sites in the United States (US) and Canada to longitudinally study the progression of dementia [2]. The ADNI project enables researchers to share and collaborate to create cost effective and innovative technology solutions for detecting AD. ADNI researchers gather, validate and apply various types of data, including biological markers (MRI, PET images, etc), neuropsychological, and demographics as predictors of the AD. One of the main purposes of ADNI is to develop clinical, imaging, genetic, and biomarkers systems for the early detection and tracking of AD. To date, there have been four phases of the ADNI project: ADNI-1, ADNI-Go, ADNI-2, ADNI-3 with data collected from all phases, (ADNI-Merge) on 2260 unique patients diagnosed as CN, MCI or demented.

The experiment is based on three datasets related to ADNI including ADNI-Merge, MMSE- feature, and CDR-SOB-feature. The latter two datasets contain specific cognitive features (questions/activities) that are related to the MMSE, and CDR-SOB cognitive tests, and were recorded on any particular visit by a patient/control who had participated in the ADNI project. For example, there are 30 attributes related to the original MMSE sheet in the ADNI data repository many assessing memory, learning, communication, etc. ADNI-Merge is a comprehensive dataset that contains data from ADNI 1/2/ GO clinical data, demographics and numeric summaries of pathological indicators of AD. We revised the datasets by merging ADNI-Merge with each cognitive test’s data so we can access the original diagnostic class that will be used to create the progression class variable. The revised datasets’ MMSE features, CDRSOB features include the cognitive features, the diagnostic class (DX) and the new modelled class variable (DX_change).

Table 3A
General statistics of the datasets before preprocessing
Dataset name	# of features	# of patients	# of data observations (visits)	Class distribution (DX) (CN:MCI:AD)	# of missing DX instances
MMSE-features	30	3,655	12,028	10,524	1,504
CDR-SOB features	27	3,322	11,924	10,528	1,396
ADNI-Merge	113	2,260	14,627
Table 3B
General statistics of the datasets after data modelling and preprocessing
Dataset name	# of features	# of patients	# of data observations (visits)	Class distribution before data balancing (DX_change) (1:0:-1)	Class Distribution after data balancing (DX_change) (1:0) using SMOTE
MMSE-features	11	1,730	8,923	8,923	16,102
CDR-SOB features	6	2,256	10,412	10,412	18,640

Table 3A shows the datasets’ characteristics for instance, the MMSE-feature dataset contains many features (questions, answers, scores, total_score, demographics, etc) Table 3B displays the data observations after we modelled and preprocessed the datasets. The processed dataset contains the cognitive features plus the DX_change class variable; all other attributes such as demographics, questions, total_score, etc were excluded. In addition, and to simplify the process of feature selection some of the features within the datasets have been merged when they belong to the same cognitive domain within the same cognitive test. We call this process ‘data modelling and mapping’. For the MMSE-feature dataset, the first column in Table 4 shows the derived columns after data modelling and to which features are mapped within the original MMSE cognitive test. For example, the scores of the first five features which are related to orientation in the MMSE cognitive test have been merged into one feature that we call ‘Q1 OrientoTime’. This new feature will have a score between 0–5. Similarly, the scores of features 6–10 in the MMSE test have been merged to form a new feature that we call ‘Q2 OrientoPlace’. Overall, the revised MMSE-feature dataset consists of 10 cognitive features, and the class (DX_change). All other features have been excluded.

For the CDR-SOB feature dataset, we validated the total scores per data observation based on Table 4B by mapping the recorded value for each CDR-SOB feature in ADNI to its corresponding value within the original CDR test. This is since the recorded features’ values in ADNI for the CDR- feature dataset do not match the corresponding values within the original CDR cognitive test. For CDR-SOB value of ‘0’ participants are mapped to ‘CN’ class, for those with scores between 0.5–4.0 participants are considered with ‘MCI’, and any score above 4, participants are considered ‘demented’. We have not considered the dementia level in the data representation prior training the ML algorithms.

Table 4A

The modelled MMSE-features data

New column name	Score range	Original question in the MMSE test	Name of the columns in dataset	Cognitive function/domain-DSM-5 category
Q1 OrientoTime	0–5	Orientation to Time 1. What is today’s date? 2. What is the year? 3. What is the month? 4. What day of the week is today? 5. What season is it?	MMDATE MMYEAR MMMONTH MMDAY MMSEASON	Complex Attention; Learning and Memory [34]
Q2 OrientoPlace	0–5	Orientation to Place 6. What is the name of this hospital (clinic, place)? 7. What floor are we on? 8. What town or city are we in? 9. What county (district, borough, area) are we in? 10. What state are we in?	MMHOSPIT MMFLOOR MMCITY MMAREA MMSTATE	Complex Attention; Learning and Memory [34]
Q3 ImmRecall	0–3	Immediate Recall 11. Ball (alt: Apple) 12. Flag (alt: Penny) 13. Tree (alt: Table)	MMBALL MMFLAG MMTREE	Complex Attention; Learning and Memory [34]
Q4 Attention	0–5	Attention 14. D 15. L 16. R 17. O 18. W	MMD MML MMR MMO MMW	Complex Attention; Learning and Memory [34]
Q5 DeVerbalRecall	0–3	Delayed Verbal Recall 19. Ball (alt: Apple) 20. Flag (alt: Penny) 21. Tree (alt: Table)	MMBALLDL MMFLAGDL MMTREEDL	Complex Attention; Learning and Memory [34]
Q6 Naming	0–2	Naming 22. Show the participant a wrist watch and ask, ‘What is this?’ 23. Repeat for pencil.	MMWATCH MMPENCIL	Language; Learning and Memory [34]
Q7 Repetition	0–1	Repetition 24. Say, ‘Repeat after me: no ifs, ands, or buts’	MMREPEAT	Language/ Learning and Memory [34] Memory
Q8 3StageCmd	0–3	3-Stage Command 25. Takes paper in right hand. 26. Folds paper in half 27. Puts paper on floor	MMHAND MMFOLD MMONFLR	Language; Learning and Memory [34]; Complex Attention [73]
Q9 Reading	0–1	Reading 28. Present the piece of paper which reads, ‘CLOSE YOUR EYES,’ and say: ‘Read this and do what it says’	MMREAD	Language/Verbal skills; Attention and Concentration (two stimuli)
Q10 Writing	0–1	Writing 29. Give the participant a blank piece of paper and say: ‘Write a sentence’	MMWRITE	Language; Executive Function; Visuospatial Cognition. Motor Skills and Construction
Q11 Copying	0–1	Copying 30. Present the participant with the Construction Stimulus page. Say, ‘Copy this design’	MMDRAW	Complex Attention; Learning and Memory; Visuospatial Cognition; Praxis Functions [34, 15]

Table 4B

CDR-SOB features validation

CDR feature data entry score (ADNI CDR-SOB sheet)	Representation	CDR test corresponding score	Action taken
$-$ 1	Missing or incomplete data	n/a	Remove
0	None	0	Keep
0.50–4.0	Questionable cognitive impairment	0.5	Keep
4.50–9.0	Mild dementia	1	Keep
9.50–15.50	Moderate dementia	2	Keep
16.0–18.0	Severe dementia	3	Keep

4. Experimental analysis

4.1 Settings

We experimented on three datasets related to the ADNI project: ADNI-Merge, MMSE-feature and CDR-SOBfeature. We integrated ADNI-Merge with each cognitive tests’ dataset respectively to ensure we capture the DX class variable as discussed in Section 3. The experiments were conducted using a Rapid Miner Studio version 9.8 tool. RapidMiner is an open-source software package that provides solutions for data preparation, ML, text mining and predictive analytics [37]. We also used WEKA (Waikato Environment for Knowledge Analysis) to run IG and PC feature assessment experiments WEKA is an open-source tool of ML that was developed at the University of Waikato, New Zealand.

All experiments have been conducted using ten-fold cross validation technique. Cross-validation is mainly applied in ML studies to estimate the predictive validity and to perform model selection – a widespread technique due to its simplicity and its universality [8]. The procedure is also known as k-fold cross-validation with one parameter ‘k’ that represents the number of groups resulting from a given dataset’s split. The processing machine applied to perform the experiments is an Intel(R) Core (TM) i5 4210U CPU @ 1.7GHZ 2.4GHz with 8 GB RAM and working on Windows 7, 64-bit operating system.

4.2 Results analysis

4.2.1 MMSE features

Table 5 depicts the features along with their computed weights which have been derived by the feature selection techniques against the MMSE data-feature. We normalised scores per feature for easier comparison and simplicity because each feature selection technique provides different score ranges. The normalised scores are obtained per Eq. (5). By providing normalised scores we are able to easily contrast between the feature selection results obtained. Min-max normalisation is used in the project using the following equation:

$\displaystyle x_{\textit{normalized}}=\frac{x-x\min}{x\max-x\min}$ (5)

No normalisation was done for the mRMR results since this technique only pinpoints whether the feature is relevant or not without calculating the weight of the feature; mRMR adopted in this research employed Correlation Feature Set (CSF) [33].

In Table 5, the top ranked features offered are ‘Q5 DeVerbalRecall’ and ‘Q1 OrientoTime’, which relate to the memory and learning cognitive domains in respect to DSM-5 neurodegenerative criteria for dementia. The findings are also consistent with the NINCDS-ADRDA Criteria for AD in which memory worsening is considered a sign in AD probable and possible diagnosis. The ‘Q1 OrientoTime’ feature in the dataset comprises of five sub-questions all related to recent memory (recall) which involve asking the participant about date, year, month, season, and day; these have been combined into one feature to simplify data modelling. Whereas ‘Q5 DeVerbalRecall’ involves repeating the names of three objects.

Since ‘Q5 DeVerbalRecall’ and ‘Q1 OrientoTime’ cognitive features overlap in terms of the cognitive area they cover, we therefore assess their similarity using a correlation matrix as shown in Fig. 2. The matrix was modelled using PC on all available features in the dataset excluding the class variable. Based on Fig. 2, ‘Q5 DeVerbalRecall’ and ‘Q1 OrientoTime’ have good correlation with each other with computed correlation above 0.41, signalling their similarity as they measure similar cognitive domains.

Another important feature that has been shown on the top ranked results of both IG and PC feature selection techniques is ‘Q4 Attention’, which is linked with complex attention and memory. However, this feature was not chosen by mRMR coupled with CSF feature selection method. For example, the stage-command feature covers a similar cognitive domain which is complex attention, similar to the attention feature. Also, there is some overlapping albeit low with remaining features in terms of the cognitive domain. Further, the ranking results of the feature selection techniques show that ‘Q8 3StageCmd’ and ‘Q10 Writing’ cognitive features have been chosen by the three techniques having ranking positions of (4, 5) and (5, 6) by the PC and IG, respectively. The ‘Q8 3StageCmd’ and ‘Q10 Writing’ as per Han et al. [34] require more complex cognitive functions, that are related Language, Complex Attention, Learning & Memory, and according to [73] Executive Function. When these are considered with the top ranked cognitive features (‘Q5 DeVerbalRecall’ and ‘Q1 OrientoTime’) more cognitive domains related to dementia conditions can be covered in the DSM-5 framework including.

The feature selection results against the MMSE data showed that ‘Copying’ has little impact with dementia at least when using the dataset and the feature selection considered since it was not chosen by mRMR with CFS, and also ranked last in IG and PC. This indicates that constructional and visuoperceptual skills are not as important as memory and complex attention when assessing individuals using the MMSE at least on the data and techniques we consider

Table 5

Feature selection results on the MMSE data

MMSE features ranking
PC			mRMR with CSF	IG
Feature	Un- normalised weight	Normalised			Un- normalised weight	Normalised
Q5 DeVerbalRecall	0.32519	1.00	Q5 DeVerbalRecall	Q5 DeVerbalRecall	0.16264199	1.00
Q1 OrientoTime	0.20509	0.63	Q1 OrientoTime	Q1 OrientoTime	0.08599731	0.53
Q4 Attention	0.15280	0.47	Q7 Repetition	$-$	0.03067758	0.19
Q8 3StageCmd	0.12626	0.38	Q8 3StageCmd	Q2 OrientoPlace	0.02659339	0.16
Q10 Writing	0.11414	0.35	Q10 Writing	Q8 3StageCmd	0.01431345	0.09
Q3 ImmRecall	0.11151	0.34		Q10 Writing	0.01163374	0.07
Q2 OrientoPlace	0.10249	0.31		Q3 ImmRecall	0.01080052	0.07
Q9 Reading	0.08302	0.25		Q9 Reading	0.00646211	0.04
Q7 Repetition	0.08223	0.25		Q7 Repetition	0.00493100	0.03
Q6 Naming	0.06989	0.21		Q6 Naming	0.00428655	0.03
Q11 Copying	0.00291	0.00		Q11 Copying	0.00000610	0.00

This can be explained because the memory is the first cognitive feature affected in the early stages of AD [34]. In addition, attention and memory are closely related in detecting prodromal dementia and should be assessed first by the medical practitioner [41].

4.2.2 CDR-SOB features

The feature-feature results using the PC is depicted in Fig. 3. Further, the feature-class (feature ranking) results of CDR-SOB are displayed in Table 6. The feature-feature results show high similarity among features since the computed correlations are high. For instance, the correlation between memory and orientation is close to 79%, and the correlation between memory and judgment is 75%. The feature class correlations are also high which make it difficult to distinguish between features of the CDR-SOB test at least when using the ADNI data repository. One principal reason for such results is the fact that the CDR-SOB test was used originally to assign the diagnostic class in the ADNI project and therefore the results of the feature selection techniques could favour its features as shown in Table 6 and thus will be difficult to differentiate between these features in a clinical cognitive assessment Nevertheless, memory dominates as the main factor that correlates with AD in all feature selection techniques when assessing the CDR – SOB feature dataset, which is consistent with the results obtained against the MMSE-feature dataset.

For the CDR-SOB features and as expected, the memory feature had been ranked first by all the feature selection techniques This is consistent with result of other researchers due to the important role of memory in defining the AD [16, 18]. However, the judgment & problem solving feature was ranked second as per the IG and mRMR-CSF result. The personal care feature was ranked last by the correlation and IG; however, it had been ranked in the third position by the mRMR-CSF. This can be explained due to overlapping of the personal care feature with other CDR-SOB features that had been selected by the two techniques (IG and correlation) such as Orientation [53, 34].

Table 6
Feature selection results on the CDR-SOB data

CDR features ranking
Correlation			mRMR (with CSF)	Information gain
Attribute	Weight	Normalised	Attribute	Attribute	Weight	Normalised
CDMEMORY	0.286	1.00	CDMEMORY	CDMEMORY	0.2395	1.00
CDCOMMUN	0.286	1.00	CDJUDGE	CDJUDGE	0.1120	0.44
CDORIENT	0.261	0.85	CDCARE	CDORIENT	0.1116	0.43
CDHOME	0.249	0.78		CDCOMMUN	0.1094	0.42
CDJUDGE	0.237	0.71		CDHOME	0.1024	0.39
CDCARE	0.118	0.00		CDCARE	0.0133	0.00

Figure 2.

Correlation matrix computed for the features of MMSE data.

Figure 3.

Correlation matrix computed for the features of CDR-SOB data.

Based on the presented results earlier, features related to the Memory cognitive domain were considered as the strongest discriminator for dementia progression based on the IG, PC and mRMR results of both tests (MMSE and CDR-SOB). This is in agreement with previous research that recommended to include Memory as well as Complex Attention in the assessment of the dementia progression [16, 18, 34]. The second group of features is related to Orientation In addition, Cognitive Executive Functions is associated with cognitive decline in many research studies and found to be a good indicator for dementia advancement [32].

The MMSE scope do not assess ‘Executive Function’ explicitly, however, some features suggested by mRMR-CSF method such as writing [21] can be argued that it may cover this cognitive area. As per Tinklenberg [71], writing is related to language domain. However, [34] found out that ‘writing’ can be associated with tasks such as making a phone call, shopping and laundry. This is coherent with [21] suggestion that writing can be associated with cognitive domains like linguistics, visual spatial and executive function. Another item of MMSE that can be an indicator to different domains is the ‘3-stage command’, which is used to measure the language skills cognitive domain. However, [34] found out that it can be related to working memory domain. They found out that it can be associated with activities such as toileting and dressing. In spite of the simplicity of ‘3-command stage’, it can provide indicators that it can be related to managing different simple cognitive domains such as language, processing speed and planning to the execution of the provided instructions. The mapping between activities of cognitive tests and cognitive area requires more in-depth investigation to highlight the overlapping, which is out of the scope of this research.

5. Conclusion

Dementia is a group of conditions that affects humans, of which AD accounts for over 60% of cases. Ten million individuals are diagnosed with AD every year globally, particularly the elderly, and millions of dollars are spent to diagnose and manage the disease imposing huge pressures on governments’ limited healthcare resources. The process of diagnosing dementia by a specialised clinician primarily requires assessing cognitive factors that measure the patient’s decline in multiple cognitive areas within a neuropsychological test besides the medical history of the patient and his/her family. The problem becomes more difficult when the clinician attempts to identify the disease progression using a specified timeframe since dementia sublevels and dementia precursors may overlap in signs of cognitive decline. Hence, it is essential to develop ways to identify impactful cognitive features of prodromal dementia for clinicians to exploit within digital information sheets during the clinical assessments based on cognitive features. In this research, we investigated the problem of identifying cognitive features that are symptoms of dementia progression using real data of two neuropsychological assessments (MMSE, CDR-SOB) within the ADNI data repository. We have proposed an experimental data process: NFAFS that consists of feature selection techniques in which the correlations between the features and the progression class variable besides correlations between pairs of feature-feature were investigated.

Experimental evaluation using real data observations of cases and controls that have been undertaken the MMSE and CDR-SOB tests of ADNI have been processed using different feature selection techniques including IG, PC and mRMR with CFS. The derived results from the MMSE-feature data showed that Time Orientation and Recall features albeit having correlation have high influence on the progression class variable being ranked the top two features of the PC and IG feature selection results. This result if limited pinpoints that these features signs of a change during the disease progression. Moreover, the results demonstrated that Complex Attention is also a cognitive domain that clinicians need to pay more attention to during the clinical assessment since the MMSE’s attention feature showed correlation with the class progression variable, and was apparent in all considered feature selection techniques’ results. Whereas features related to perceptual motor skills such as MMSE’s copying shapes has low ranking and was not detected by the mRMR with CFS feature selection results.

The results derived from the CDR-SOB features reveal that it is challenging to differentiate among the features, which makes it difficult to spot the ones that are symptoms of the disease advancement. However, CDR-SOB’s feature related to memory was the highest sign of AD progression, which supports previous research findings.

In the near future, we will employ the cognitive features besides biological markers to design and implement an AI algorithm based on deep learning to build classification systems for AD progression. In addition, we would like to determine whether the cognitive features will change based on the level of dementia.

Footnotes

Funding

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 (www.fnih.org). 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.

References

AlShboul

Thabtah

Walter Scott

Wang

. The Application of Intelligent Data Models for Dementia Classification. Applied Sciences. 2023; 13(6): 3612.

Alzheimer’s Disease Neuroimaging Initiative (ADNI) (2017). Available from: http://adniloni.usc.edu/about/.

Alzheimer’s Association. (2020a). Dementia vs. Alzheimer’s Disease: What Is the Difference? Retrieved November 29, 2021, from https//www.alz.org/alzheimers-dementia/difference-between-dementia-and-alzheimer-s.

Alzheimer’s Association. (2020b). What is Alzheimer’s Disease? Symptoms & Causes. Retrieved November 29, 2021, from https//www.alz.org/alzheimers-dementia/what-is-alzheimers.

Alzheimer’s Association. (2020c). What is Dementia? Symptoms, Causes & Treatment. Retrieved November 29, 2021, from https//www.alz.org/alzheimers-dementia/what-is-dementia.

American Psychiatric Association (2013). Diagnostic and Statistical Manual of Mental Disorders. In Pediatria Integral (Vol. 17). doi: 10.1176/appi.books.9780890425596.

Apostolova

. (2016). Alzheimer Disease. Retrieved from www.ContinuumJournal.com.

Arlot

Celisse

. A survey of cross-validation procedures for model selection. Statistics Surveys. 2010; 4: 40-79.

Balsis

Benge

Lowe

Geraci

Doody

. How do scores on the ADAS-cog, MMSE, and CDR-SOB correspond? The Clinical Neuropsychologist. 2015; 29(7): 1002-1009.

10.

Battista

Salvatore

Castiglioni

. Optimizing Neuropsychological Assessments for Cognitive, Behavioral, and Functional Impairment Classification: A Machine Learning Study. Behavioural Neurology. 2017; 1-19. doi: 10.1155/2017/1850909.

11.

Bergeron

Flynn

Verret

Poulin

Bouchard

Bocti

, … Laforce

. Multicenter Validation of an MMSE-MoCA Conversion Table. Journal of the American Geriatrics Society. 2017; 65(5): 1067-1072. doi: 10.1111/jgs.14779.

12.

Berkson

. Application of the Logistic Function to Bio-Assay. Journal of the American Statistical Association. 1944; 39(227): 357-365. doi: 10.1080/01621459.1944.10500699.

13.

Broomhead

Lowe,

. Radial basis functions, multi-variable functional interpolation and adaptive networks. ROYAL SIGNALS AND RADAR ESTABLISHMENT MALVERN (UNITED KINGDOM), RSRE-MEMO-4148, 1988.

14.

Buschke

. Selective Reminding for analysis of memory and learning. Journal of Verbal Learning and Verbal Behavior. 1973; 12: 543-550. doi: 10.1016/S0022-5371(73)80034-9.

15.

Cameron

Worrall-Carter

Page

Stewart

Ski

. Screening for mild cognitive impairment in patients with heart failure: Montreal cognitive assessment versus mini mental state exam. Eur J Cardiovasc Nurs. 2013 Jun; 12(3): 252-60. doi: 10.1177/1474515111435606. Epub 2012 Apr 18. PMID: 22514141.

16.

Capucho

Brucki

. Judgment in Mild Cognitive Impairment and Alzheimer’s disease. Dementia & Neuropsychologia. 2011; 5(1): 297-302. doi: 10.1590/S1980-57642011DN05040007.

17.

Chawla

Bowyer

Hall

Kegelmeyer

. SMOTE: synthetic minority over-sampling technique. Journal of Artificial Intelligence Research. 2002; 16: 321-357.

18.

Choe

Lee

Choi

Suh

Lee

Kim

, Alzheimer’s Disease Neuroimaging Initiative. MMSE Subscale Scores as Useful Predictors of AD Conversion in Mild Cognitive Impairment. Neuropsychiatric Disease and Treatment. 2020; 16: 1767-1775. doi: 10.2147/NDT.S263702.

19.

Cortes

Vapnik

. Support-vector networks. Machine Learning. 1995; 20(3): 273-297.

20.

Cox

Snell

. Applied statistics-principles and examples (Vol. 2). CRC Press, 1981.

21.

Croisile

. Writing, aging and Alzheimer' s disease. Psychologie & amp; neuropsychiatrie du vieillissement. 2005; 3(3): 183-197.

22.

Das

Ito

Kadowaki

Tsuda

. An interpretable machine learning model for diagnosis of Alzheimer’s disease. Peer J. 2019; 7: e6543. doi: 10.7717/peerj.6543.

23.

Datta

Shankle,

Pazzani

. Applying machine learning to an Alzheimer’s database. Artificial Intelligence in Medicine. 1996; 26-30.

24.

Devanand

Liu

Tabert

Pradhaban

Cuasay

Bell

, … Pelton

. Combining Early Markers Strongly Predicts Conversion from Mild Cognitive Impairment to Alzheimer’s Disease. Biological Psychiatry. 2008; 64(10): 871-879. doi: 10.1016/j.biopsych.2008.06.020.

25.

Doty

Shaman

Dann

. Development of the university of pennsylvania smell identification test: A standardized microencapsulated test of olfactory function. Physiology & Behavior. 1984; 32(3): 489-502. doi: 10.1016/0031-9384(84)90269-5.

26.

Farooq

Anwar

Awais

Rehman

. A deep CNN based multi-class classification of Alzheimer’s disease using MRI. 2017; IEEE International Conference on Imaging Systems and Techniques (IST), (2017); 1-6. doi: 10.1109/IST.2017.8261460.

27.

Folstein

McHugh

. Mini-Mental State: A practical method for grading the cognitive state of patients for the clinician. Journal of Psychiatric Research. 1975; 12: 189-198.

28.

Frank

Hall

Ian

. The WEKA Workbench. Online Appendix for “Data Mining: Practical Machine Learning Tools and Techniques” (4

{}^{\text{th}}

; ed). Morgan Kaufmann, 2016.

29.

Galton

. Regression Towards Mediocrity in Hereditary Stature. The Journal of the Anthropological Institute of Great Britain and Ireland. 1886; 15: 246-263. doi: 10.2307/2841583.

30.

Grassi

Loewenstein

Caldirola

Schruers

Duara

Perna

. A clinically-translatable machine learning algorithm for the prediction of Alzheimer’s disease conversion: Further evidence of its accuracy via a transfer learning approach. International Psychogeriatrics. 2019; 31(7): 937-945. doi: 10.1017/S1041610218001618.

31.

Greenwood

Nikulin

. A guide to chi-squared testing (Vol. 280); John Wiley & Sons, 1996.

32.

Guarino

Favieri

Boncompagni

Agostini

Cantone

Casagrande

. Executive Functions in Alzheimer Disease: A Systematic Review. Frontiers in Aging Neuroscience. 2019; 10: 437. doi: 10.3389/fnagi.2018.00437.

33.

Hall

. Correlation-based feature selection for machine learning, 1999.

34.

Han

Maruta

Ikeda

Ishikawa

Tanaka

Koyama

Fukuhara

Boku

Takebayashi

Tabira

. Relationship between Performance on the Mini-Mental State Examination Sub-Items and Activities of Daily Living in Patients with Alzheimer’s Disease. Journal of Clinical Medicine. 2020; 9(5): 1537. doi: 10.3390/jcm9051537.

35.

Harvey

. Domains of cognition and their assessment? Dialogues in Clinical Neuroscience. 2019; 21(3): 227-237. doi: 10.31887/DCNS.2019.21.3/pharvey.

36.

Hinrichs

Singh

Johnson

. Predictive markers for AD in a multi-modality framework: An analysis of MCI progression in the ADNI population. NeuroImage. 2011; 55(2): 574-589. doi: 10.1016/j.neuroimage.2010.10.081.

37.

Hofmann

Klinkenberg

. eds. RapidMiner: Data mining use cases and business analytics applications. CRC Press, 2016.

38.

Hsiao

Grill

Teng

. Longitudinal declines in instrumental activities of daily living in stable and progressive mild cognitive impairment. Dementia and Geriatric Cognitive Disorders. 2015; 39(Mci): 12-24. doi: 10.1159/000365587.

39.

Hughes

Berg

Danziger

Coben

Martin

. A new clinical scale for the staging of dementia. British Journal of Psychiatry. 1982; 140(6): 566-572. doi: 10.1192/bjp.140.6.566.

40.

Katzman

Brown

Fuld

Peck

Schechter

Schimmel

. Validation of a short orientation-memory-concentration test of congestive impairment. American Journal of Psychiatry. 1983; 140(6): 734-739. doi: 10.1176/ajp.140.6.734.

41.

Kolanowski

Fick

Yevchak

Hill

Mulhall

McDowell

. Pay attention! The critical importance of assessing attention in older adults with dementia. J Gerontol Nurs. 2012 Nov; 38(11): 23-7. doi: 10.3928/00989134-20121003-05. Epub 2012 Oct 15. PMID: 23066682; PMCID: PMC3571625.

42.

Lee

Kim

Jhoo

Kim

Woo

. Development of the Korean Version of the Consortium to Establish a Registry for Alzheimer’s Disease Assessment Packet (CERAD-K) clinical and neuropsychological assessment batteries. The Journals of Gerontology Series B: Psychological Sciences and Social Sciences. 2002; 57(1): P47-P53.

43.

Lawton

Brody

. Assessment of older people: Self-maintaining and instrumental activities of daily living. Gerontologist. 1969; 9: 179-186.

44.

Liu

Wang

Di Dong

Fang

Zhou

Sun

Tian

. The applications of radiomics in precision diagnosis and treatment of oncology: opportunities and challenges. Theranostics. 2019; 9(5): 1303.

45.

McKhann

Drachman

Folstein

Katzman

Price

Stadlan

. Clinical diagnosis of Alzheimer’s disease. Neurology. 1984; 34(7): 939. doi: 10.1212/WNL.34.7.939.

46.

Mohs

Rosen

Davis

. The Alzheimer’s disease assessment scale: An instrument for assessing treatment efficacy. Psychopharmacology Bulletin. 1983; 19(3): 448-450. Retrieved from http//ci.nii.ac.jp/naid/10029841229/en/.

47.

Mueller

Weiner

Thal

Petersen

Jack

Jagust

, … Beckett

. Ways toward an early diagnosis in Alzheimer’s disease: The Alzheimer’s Disease Neuroimaging Initiative (ADNI). Alzheimer’s and Dementia. 2005; 1(1): 55-66. doi: 10.1016/j.jalz.2005.06.003.

48.

Nasreddine

Phillips

Bédirian

Charbonneau

Whitehead

Collin

, … Chertkow

. The Montreal Cognitive Assessment, MoCA: A Brief Screening Tool For Mild Cognitive Impairment. Journal of the American Geriatrics Society. 2005; 53(1): 695-699. 10.1111/j.1532-5415.2005.53221.x.

49.

National Alzheimer’s Coordinating Center. (n.d). The NACC database. https://www.alz.washington.edu/index.html/.

50.

Niyas

Muhammed KP

Thiyagarajan

. Feature selection using efficient fusion of Fisher Score and greedy searching for Alzheimer’s classification. Journal of King Saud University – Computer and Information Sciences. 2021; ISSN 1319-1578, doi: 10.1016/j.jksuci.2020.12.009.

51.

Pal

Mitra

. Multilayer perceptron, fuzzy sets, classification, 1992.

52.

Paulavicius

Mizzaci

Tavares

DRB

Rocha

Civile

Schultz

Pinto

ACPN

Trevisani

VFM

. Bilingualism for delaying the onset of Alzheimer’s disease: a systematic review and meta-analysis. Eur Geriatr Med. 2020 Aug; 11(1): 651-658. doi: 10.1007/s41999-020-00326-x. Epub 2020 May 4. PMID: 32367493.

53.

Prakoso

Vitriana

Ong

. Correlation between Cognitive Functions and Activity of Daily Living among Post-Stroke Patients. Althea Med. J. 2016; 3: 329-333.

54.

Peng

Long

Ding

. Feature selection based on mutual information criteria of max-dependency, max-relevance, and min-redundancy. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2005; 27(8): 1226-1238.

55.

Pfeffer

Kurosaki

Harrah

Chance

Filos

. Measurement of functional activities in older adults in the community. Journals of Gerontology. 1982; 37(3): 323-329. doi: 10.1093/geronj/37.3.323.

56.

Quinlan

. Induction of decision trees. Machine Learning. 1986; 1(1): 81-106. doi: 10.1007/BF00116251.

57.

RapidMiner (n.d). RapidMiner Studio. https://rapidminer.com/products/studio/.

58.

Rasmussen

. Gaussian processes for machine learning. MIT Press, 2006.

59.

Reisberg

Ferris

De Leon

Crook

. The Global Deterioration Scale for Assessment of Primary Degenerative Dementia. Am J Psychiatry. 1982; 139(9): 1136-1139. Retrieved from http//www.ncbi.nlm.nih.gov/pubmed/7114305.

60.

Remeseiro

Bolon-Canedo

. A review of feature selection methods in medical applications. Computers in Biology and Medicine. 2019; 112: 103375.

61.

Foundation

. (n.d) What is R? https://www.r-project.org/about.html.

62.

Salvatore

Castiglioni

. A wrapped multi-label classifier for the automatic diagnosis and prognosis of Alzheimer’s disease. Journal of Neuroscience Methods. 2018; 302: 58-65. doi: 10.1016/j.jneumeth.2017.12.016.

63.

Samper-González

. et al. Yet Another ADNI Machine Learning Paper? Paving the Way Towards Fully-Reproducible Research on Classification of Alzheimer’s Disease. In: Wang

Shi

Suk

Suzuki

(eds). Machine Learning in Medical Imaging. MLMI 2017. Lecture Notes in Computer Science, vol 10541. Springer, Cham, 2017. doi: 10.1007/978-3-319-67389-9_7.

64.

Shevade

Keerthi

Bhattacharyya

Murthy

KRK

. Improvements to the SMO Algorithm for SVM Regression. IEEE Transactions on Neural Networks, 1999.

65.

Hooshyar

Park

Lim

. Early diagnosis of dementia from clinical data by machine learning techniques. Applied Sciences. 2017; 7(7): 651.

66.

Thabtah

Spencer

. Peebles Common dementia screening procedures: DSM-5 fulfilment and mapping to cognitive domains. International Journal of Behavioural and Healthcare Research. 2022 April 4; 8(1-2): 104-120.

67.

Thabtah

Spencer

. Ye

. The correlation of everyday cognition test scores and the progression of Alzheimer’s disease: a data analytics study. Health Inf Sci Syst. 2020; 8: 24. doi: 10.1007/s13755-020-00114-8.

68.

Thabtah

Ong

Peebles

. Examining Cognitive Factors for Alzheimer’s Disease Progression Using Computational Intelligence. In Healthcare (Vol. 10, No. 10, p. 2045), (2022a, October). MDPI.

69.

Thabtah

Ong

Peebles

. Detection of dementia progression from functional activities data using machine learning techniques. Intelligent Decision Technologies. 2022b; 16(3): 615-630.

70.

The Python Software Foundation. (n,d) The Python Tutorial(nd). https//docs.python.org/3/tutorial/.

71.

Tinklenberg

Brooks

3rd Tanke

Khalid

Poulsen

Kraemer

Gallagher

Thornton

Yesavage

. Factor analysis and preliminary validation of the mini-mental state examination from a longitudinal perspective. Int Psychogeriatr. 1990 Fall; 2(2): 123-34. doi: 10.1017/s1041610290000382. PMID: 2101301.

72.

Trambaiolli

Lorena

Fraga

Kanda

PAM

Anghinah

Nitrini

. Improving Alzheimer’s disease diagnosis with machine learning techniques. Clinical EEG and Neuroscience. 2011; 42(3): 160-165. doi: 10.1177/155005941104200304.

73.

Trzepacz

Hochstetler

Wang

Walker

Saykin

. Alzheimer’s Disease Neuroimaging Initiative. Relationship between the Montreal Cognitive Assessment and Mini-mental State Examination for assessment of mild cognitive impairment in older adults. BMC Geriatr. 2015 Sep 7; 15: 107. doi: 10.1186/s12877-015-0103-3. PMID: 26346644; PMCID: PMC4562190.

74.

Vinutha

Pattar

Sharma

Shenoy

Venugopal

. A Machine Learning Framework for Assessment of Cognitive and Functional Impairments in Alzheimer’s Disease: Data Preprocessing and Analysis. The Journal of Prevention of Alzheimer’s Disease. 2020; 1-8.

75.

Welsh

Butters

Hughes

Mohs

Heyman

. Detection and Staging of Dementia in Alzheimer’s Disease. Archives of Neurology. 1992; 49(5): 448. doi: 10.1001/archneur.1992.00530290030008.

76.

Whitley

. A genetic algorithm tutorial. Statistics and Computing. 1994; 4(2): 65-85. doi: 10.1007/bf00175354.

77.

Wyman-Chick

Scott

. Development of Clinical Dementia Rating Scale Cut-off Scores for Patients With Parkinson’s Disease, 2015. doi: 10.1002/mdc3.12163.

78.

Zhang

Kwoh

Liu

Yin

Wirawan

Cheung

Wong

. MRMR optimized classification for automatic glaucoma diagnosis. In 2011; Annual International Conference of the IEEE Engineering in Medicine and Biology Society (pp. 6228-6231). IEEE, 2011.

79.

Zhu

Panwar

Dodge

Hampstead

Albin

Guan,

. et al. COMPASS: A computational model to predict changes in MMSE scores 24-months after initial assessment of Alzheimer’s disease. Scientific Reports. 2016 6 (October), 1-12. doi: 10.1038/srep3456.

Neuropsychological features evaluation of data related to Alzheimer’s disease progression using feature selection

Abstract

BACKGROUND:

AIMS:

METHODS:

RESULTS AND IMPLICATIONS:

Keywords

1. Introduction

2. Literature review

2.1 Background on dementia and cognitive assessments

Table 1 Common neuropsychological tests

3. Methodology

3.1 Data, features, and methods

4.1 Settings

4.2 Results analysis

4.2.1 MMSE features

Table 6 Feature selection results on the CDR-SOB data

Footnotes

Funding

References

Table 1
Common neuropsychological tests

Table 6
Feature selection results on the CDR-SOB data