A Regression Model For Alzheimer's Disease Progression Using The ADNI Database

Abstract

Alzheimer's disease is a neurodegenerative disorder and the most common form of dementia. It afects approximately 50 million people worldwide, and to date, no definitive cure has been found. As one of the leading causes of death among individuals over the age of 65, early diagnosis is crucial, as it can significantly improve life expectancy and quality of life. In recent years, numerous machine learning techniques have been applied to various biomarkers to support the early detection of the disease. The objective of this project is to conduct an in-depth analysis of the ADNI database in order to study the characteristics of individuals afected by Alzheimer's at diferent stages of the disease, using machine learning methods. The results of this study demonstrated that it is possible to distinguish four distinct stages of Alzheimer's progression-from cognitively healthy individuals to those severely afected-rather than the commonly discussed three. Notably, the analysis also revealed that women are disproportionately impacted by the disease, accounting for nearly 80% of the afected population.

1. Introduction

tions such as physical activity and cognitive exercises have also been shown to delay the onset of brain damage.

Alzheimer’s disease (AD) is a neurodegenerative disor- Given the slow and progressive nature of the disease, der and the most common cause of dementia, account- longitudinal research is essential to better understand ing for approximately 50–70% of all diagnosed dementia its development over time. Longitudinal studies are parcases. It can severely compromise an individual’s ability ticularly valuable in Alzheimer’s research because they to perform daily activities and alter their personality, as involve repeated evaluations of the same individuals, alit afects brain regions responsible for memory, language, lowing researchers to detect subtle changes and trends and cognitive function. In 2020, over 55 million people as the disease progresses. worldwide were afected by the disease, and projections Among the most prominent longitudinal resources estimate this number will rise to 139 million by 2050 [ 1 ]. available for Alzheimer’s research is the Alzheimer’s

The most frequently observed symptoms of AD in- Disease Neuroimaging Initiative (ADNI) database. One clude memory loss and behavioral changes, both of which of the main objectives of ADNI is the early detection are linked to the accumulation of specific biological sub- of Alzheimer’s disease and the identification of reliable stances [ 2 ]. These substances progressively disrupt neu- biomarkers to monitor its progression. ronal function, ultimately leading to widespread brain The ADNI study involves volunteers aged 55 to 90, atrophy. One of the most severely afected regions is recruited from various research sites across the United the hippocampus, the brain structure responsible for the States and Canada. Launched in 2004, the study has formation of new memories. This degeneration is caused evolved through several phases, each with specific objecprimarily by the abnormal behavior of tau proteins and tives and participant cohorts. The first phase, ADNI-1, -amyloid plaques, which form deposits in the brain and aimed to develop biomarkers for use as outcome meacontribute to neuronal impairment. sures in clinical trials. The cohort consisted of 200 cogni

Because of the irreversible damage caused to the neu- tively normal elderly individuals, 400 participants with ronal population, Alzheimer’s is a progressive and in- Mild Cognitive Impairment (MCI), and 200 patients dicurable condition that worsens over time. Despite its agnosed with Alzheimer’s Disease. The second phase, widespread impact, no definitive cure exists to date. How- ADNI-GO, focused on identifying biomarkers at earlier ever, early diagnosis is critical, as it allows for inter- stages of the disease. It expanded the ADNI-1 cohort by ventions that may slow the progression of the disease. including an additional 200 participants diagnosed with Current treatments include medications that can help early MCI. The third phase, ADNI-2, built on previous manage symptoms and maintain the patient’s functional eforts by refining biomarkers as predictors of cognitive independence for longer. Moreover, lifestyle interven- decline and as clinical trial endpoints. This phase incorporated participants from both ADNI-1 and ADNI-GO, SYSTEM 2025: 11th Sapienza Yearly Symposium of Technology, Engi- and enrolled an additional 150 cognitively normal indineering and Mathematics. Rome, June 4-6, 2025 viduals, 100 with early MCI, 150 with late MCI, and 150 $ s.russo©@20h25sCaonpytraiglhutfcoirath.iistp(apSe.r bRy uitssasuoth)ors. Use permitted under Creative Commons License with more advanced mild cognitive symptoms. Finally, Attribution 4.0 International (CC BY 4.0). .

ADNI-3 emphasized the use of tau PET imaging and other Table 1

advanced functional imaging techniques in clinical trials. Acronyms It continued with the ADNI-2 cohort and added 133 more cognitively normal elderly participants. Acronym Explanation

The dataset is composed of the results obtained from a AD Alzheimer’s Disease series of tests that participants undergo at multiple time MCCNI CMoilgdnCitiovgenlyitNiveorImmaplairment points during the study. These tests fall into two main ADNI Alzheimer’s Disease Neuroimaging Initiative categories: clinical tests and cognitive tests. Clinical tests KNN K Nearest Neighbours are exploratory examinations that gather biological and NAN Not A Number physical data, aiming to confirm or exclude specific diag- RMSE Root Mean Square Error noses and to detect potential anomalies in the patient’s MADMASSE AMlzinhie-Mimeenrt’salDSitsaetaeseExAasmseisnsamtieonnt Scale condition. Cognitive tests, on the other hand, are de- RAVLT Rey Auditory Verbal Learning Test signed to evaluate mental functioning and to assess how TRABSCOR Time to complete Part B of the Trail Making Test the brain processes information. They usually consist of FAQ Functional Activities Questionnaire simple tasks or questions that the participant is required CDR Clinical Dementia Rating Scale to complete, ofering insight into memory, attention, language skills, and other cognitive domains.

In this project, we will use the ADNI database as it field of neurophysiological research, as no cure has yet is one of the biggest and most used in Alzheimer’s re- been discovered, and its progression varies significantly search. We will also analyze its longitudinal data in or- from person to person. AD is typically diagnosed with der to predict the evolution of the disease, to prevent high accuracy only in its later stages, which is why recent further deterioration and to take earlier clinical and cog- studies have focused on detecting the disease at earlier nitive actions. In this context, several machine learning phases. One of the most recent contributions to this approaches have been applied to extract patterns from line of research [ 9 ] highlights ongoing eforts toward ADNI data, with particular success in classification prob- adopting a continuous model of disease progression. lems. Among these, neural networks [ 3 ]have shown Until a few years ago, based on the degree of brain dampromising results. For example, probabilistic neural net- age, it was common to characterize Alzheimer’s progresworks (PNNs) [ 4 ] have been used to classify patients sion using three main stages. The first stage, referred to with Alzheimer’s versus those with mild cognitive im- as Cognitively Normal, includes individuals whose brains pairment or normal cognition. Other models, such as show no structural or functional damage and whose brain elliptical basis neural networks (EBNNs) [ 5 ], and other volume remains within the average range. The second hybrid neural network based systems [ 6 ] have been ap- stage, known as Mild Cognitive Impairment (MCI), correplied to model more complex decision boundaries and to sponds to the early stages of Alzheimer’s development. It integrate imaging data with neuropsychological scores. is considered a transitional phase between normal aging

Moreover, recent studies have started exploring the and dementia-related decline and is typically associated link between Alzheimer’s disease and deficits in theory of with memory and language dificulties. The final stage is mind [ 7 ], the capacity to understand others’ beliefs, emo- Alzheimer’s Disease, in which the brain is clearly afected tions, and intentions. This aspect, often investigated in so- by the disease, both structurally and functionally. cial cognition, is found to deteriorate early in Alzheimer’s As with many other illnesses, the known nature of and can provide sensitive behavioral markers of cogni- Alzheimer’s is primarily biological, involving both structive decline. In parallel, the role of sustained attention [ 8 ] tural and molecular changes in the brain. Therefore, the has gained increasing attention in the literature, since im- study of AD relies heavily on the analysis of biomarkpairments in maintaining focus over time are commonly ers. These biomarkers represent objective indicators of observed in early phases of the disease. These deficits, a patient’s medical state and are characterized by their measurable through specific cognitive tasks, may serve accuracy and reproducibility [ 10 ], making them essential as early indicators and help in designing targeted inter- for identifying the diferent stages of the disease. ventions aimed at preserving attentional control. According to the methodology described in [ 11 ], AD biomarkers can be broadly categorized into two groups.

The first group includes imaging-derived biomarkers, 2. Related Works which are extracted from techniques such as MRI or PET scans. The second group comprises biochemical biomarkAs there will be some acronyms that will appear through- ers derived from cerebrospinal fluid (CSF), which reflect out the literature review, it is possible to check their the molecular composition and alterations occurring in explanation in the table 1. the brain. Both categories are crucial for understandAlzheimer’s disease represents a major challenge in the ing the progression of Alzheimer’s and for improving early-stage detection. chine learning algorithms are often preferred, as they

In order to define a more continuous approach to the allow for dimensionality reduction. Most studies applydisease, a deep understanding of the diferent biomarkers ing ML algorithms to AD rely on supervised learning is necessary, as some of them may vary significantly over approaches to classify patients using either neuroimagthe years due to age-related decline, while others may ing data or biomarkers, as shown in [18] and [19]. only manifest when the disease is in its later stages or ac- Based on this literature review, we observe Alzheimer’s tively progressing. Therefore, it is essential to understand disease has been predominantly studied through medthe nature of diferent biomarkers prior to conducting ical imaging and classification algorithms. This led us the research. to our research direction: focusing on the use of non

Since Alzheimer’s is a dynamic disease that primarily imaging biomarkers from the longitudinal ADNI dataset. afects the brain and its neuronal population, numerous Our analysis will employ unsupervised machine learning studies such as [ 12 ] and [ 13 ] have focused on identify- algorithms to group patients with similar medical condiing early stages of the disease using imaging-derived tions and subsequently examine the temporal evolution biomarkers. However, these abnormalities generally be- of key biomarkers. come apparent only in the later stages of Alzheimer’s, meaning that MRI scans may appear normal during the early phases. Moreover, conventional machine learning 3. Methodology techniques have shown limited efectiveness due to their reliance on expert users for complex feature extraction. In this project, we aim to classify Alzheimer’s disease These earlier studies aimed to build models that analyze into distinct stages, identify the most relevant characteranatomical and structural brain images obtained from istics associated with each stage, and perform a temporal MRIs, as well as assess brain function to identify defects analysis of how various parameters evolve. Our objecand abnormalities. Additionally, training these models tive is to find common traits among patients who are often required extensive image partitioning, increasing at the same stage of the disease. Since the progression the time and complexity of the process. As a result, deep of Alzheimer’s is not uniform across individuals, we delearning techniques have been increasingly adopted in cided to group patients based on their clinical test results, such cases [14]. without relying on predefined diagnostic labels. For this

On the other hand, novel biomarkers such as cere- reason, we work with unlabeled data, meaning that the brospinal fluid (CSF) concentrations of -amyloid, also ifnal diagnosis of each patient is not used in the analysis. analyzable via PET imaging, have gained attention [ 9 ]. Among the available unsupervised machine learning When combined with traditional neuropsychological as- algorithms for classification, we selected the K-Means alsessments, these biomarkers can better define disease gorithm, as it allows us to partition patients according to progression. This evolution can then be compared to that similarities in their clinical profiles. However, K-Means of healthy individuals over time, ofering a more accu- is known to be sensitive to the choice of the number rate characterization than a single imaging result. By of clusters. To address this issue and optimize the algoanalyzing the progression of CSF biomarkers, it becomes rithm’s performance, we applied the Elbow Method [20] possible to determine whether a patient’s decline is faster to determine the most appropriate number of clusters. than that of a healthy individual, which could support a The algorithm was executed for up to 20 clusters, and continuous model of disease progression and potentially the results showed a clear inflection point at = 4, enable earlier detection, before reaching severe stages. where further increases in the number of clusters led to

Although medical imaging is a valuable tool for only marginal improvements in performance. Therefore, Alzheimer’s analysis—ofering detailed views of brain we selected four clusters for the analysis. The correspondvolume, it remains a resource-intensive technique, as pa- ing Elbow graph can be seen in Figure 1. tients must undergo complex scanning procedures. In Once the number of clusters has been defined, we apcontrast, CSF biomarkers may ofer a more accessible ply the K-Means algorithm, which assigns each patient to and interpretable alternative for continuous monitoring. the cluster with the closest centroid. The algorithm oper

Today, thanks to technological advancements, particu- ates as follows: first, it initializes or updates the centroids; larly in artificial intelligence, computer-aided diagnosis then, it calculates the distance between each patient and systems have been developed [15]. These systems assist all centroids; each patient is assigned to the nearest cenin the detection and diagnosis of medical data, serving troid. If, during this process, any patient changes cluster, as a "second opinion" for healthcare professionals. In the algorithm restarts until convergence is reached. Alzheimer’s disease, CAD systems are mainly applied to Working with high-dimensional data increases the medical image interpretation [16]. However, data pro- complexity of the analysis. Although we reduced the cessing becomes increasingly complex when dealing with number of features by removing those with insuficient high-dimensional feature spaces [17]. Consequently, ma- data, the dimensionality remained significant, with 95 mensionality reduction methods) is the loss of physical information. However, in this project, where the goal is to identify similar characteristics or features among patients at the same stage of Alzheimer’s, dimensionality reduction does not provide us with meaningful insights.

Therefore, we decided not to include feature reduction algorithms such as t-SNE or PCA in our analysis.

Once we identified the diferent stages of the disease, it became possible to train classification models that can assign new patient data to one of the defined stages. This Figure 2: T-SNE cluster results classification provides insight into the patient’s current condition and allows for a prediction of disease progression, as we can estimate how many stages remain before features still under consideration. Therefore, we consid- reaching the most severe form. The algorithms used ered the implementation of a dimensionality reduction for this task are the Regression Tree Classifier and the algorithm such as PCA or t-SNE. XGBoost Classifier.

Following the conclusion of [21], which highlights We selected these two algorithms because, although that the t-SNE algorithm performs better when dealing they process data diferently, both are capable of classiwith non-linear structures—unlike PCA, which is a linear fying new instances into the defined stages. Boosting algorithm—we decided to implement the t-SNE algorithm. algorithms are based on the idea of creating highly acThe main characteristic of this algorithm is its ability curate prediction rules by combining multiple weak and to preserve clusters from high-dimensional spaces even imprecise rules. eXtreme Gradient Boosting (XGBoost) after reduction to lower dimensions [22]. combines decision trees with gradient boosting to min

After applying the algorithm, the dataset’s dimension- imize execution time and maximize eficiency. On the ality was reduced to two components. The resulting other hand, regression trees integrate decision-making clusters can be observed in Figure 2. and prediction to classify new data efectively.

After analyzing these results, although the outcomes Figure 3 shows the accuracy diferences when both from feature reduction were satisfactory, as it was possi- models are applied to the same dataset, and Table 3 conble to group patients into four clusters—the main issue ifrms that the XGBoost algorithm performs better in claswith this type of algorithm (and probably with most di- sifying new cases into the identified disease stages.

These are the main features considered for the anal

ysis and will serve as input for the K-Means algorithm. However, since our goal is to partition the patients based solely on their clinical data, we excluded categorical and socio-demographic features from the clustering process. Nevertheless, these features were analyzed after the clusters were formed and the patients were assigned to each cluster, in order to gain further insights into the characteristics of each group.

4. Dataset and Treatments

The database used for the analysis was ADNI, specifically the ADNIMERGE dataset, which contains data from 2,419 diferent patients (1,150 females and 1,269 males) and includes 115 features for each patient. From this dataset, certain features or patients lacked suficient data for proper analysis, so we removed those features whose 5. Experiments and Results proportion of missing values (NaNs) was 80% or higher.

After removing these features, we applied a supervised The implementation of the K-Means algorithm on the machine learning model—specifically, the K-Nearest processed dataset allowed us to identify four Alzheimer’s Neighbors (KNN) algorithm—to fill in the remaining miss- stages, which are presented in Table 4. Having defined ing values in the dataset. The choice of this algorithm these four stages of the disease, and considering that aimed to provide more accurate estimations for the miss- Alzheimer’s progressively worsens over time, it is posing values by imputing them based on the values of the sible to arrange the pathological levels chronologically, most similar patients. Unlike the K in K-Means, in KNN with level 0 representing the earliest stage and level 3 the the parameter K represents the number of neighbors con- most advanced one. sidered for the imputation. To achieve greater accuracy, This represents a novelty compared to the three main we computed the Root Mean Square Error (RMSE) for characterizations (AD, MCI, and CN), as it introduces diferent values of K and selected = 13 as it yielded a new stage that can aid in identifying the disease bethe minimum error, with RMSE( = 13) = 0.2087. All fore it reaches a more severe and irreversible state. Since computed errors are presented in Table 3. Alzheimer’s causes brain shrinkage, Figure 4 shows the The pseudocode of the KNN algorithm is as follows: whole brain volume at each of the identified levels. In this image, it is clearly visible that the lower the brain 1. Find the Euclidean distance to all training data volume, the more advanced the disease. Considering points that the average cranial volume in men is greater than 2. Sort each distance 1500 cm3, a clear diagnosis of the disease can be estab3. Select the first k values lished when the whole brain volume is 1.05 · 103 cm3 4. Assign the value based on the selected points or less. Other types of dementia, such as Huntington’s or Parkinson’s disease, can also cause brain shrinkage.

In this project, we used the KNNImputer from the Therefore, since levels 1 and 2 in our classification may sklearn library in Python to complete the missing val- indicate early stages of Alzheimer’s, it is necessary to ues using the K-Nearest Neighbors algorithm. examine additional biomarkers to ensure an accurate

Once the dataset was completed, we carried out a fea- diagnosis. ture analysis to identify the most relevant variables for Another objective of the project was to obtain inforthe study. The features considered include categorical mation about patients in each group and to try to identify information such as age, gender, education, and ethnicity, possible correlations. In image 5, it is possible to see the as well as results from cognitive assessments like the gender distribution of patients and their corresponding Clinical Dementia Rating (CDR) and Functional Activi- Alzheimer’s level, showing a marked diference between ties Questionnaire (FAQ). In addition, clinical test scores males and females, with the female group being the most such as MMSE, ADAS, RAVLT, and TRABSCOR were afected. The research results confirmed that women analyzed to evaluate cognitive impairment levels. have a higher overall tendency to develop AD pathology compared to men, as also reflected in our findings, reinforcing the idea that biological gender is a significant risk factor for Alzheimer’s disease.

The distribution of categorical features across diferent AD levels is summarized in Table 5, expressed as percentages. The analysis revealed that neither race nor ethnicity appeared to be significant factors in the development of

Alzheimer’s Disease within our dataset. However, the

ifndings of Liu et al. [ 23] align with our results, highlighting that married individuals are less prone to developing AD. This is likely due to the protective efect of close afective relationships, in contrast to individuals who are divorced, widowed, or living alone. In other words, stable couple relationships may act as a bufer against both the onset and progression of the disease.

Recent evidence from Rodriguez et al. [24] further supports our conclusions, showing that women are more likely than men to develop AD—an observation consistent with our own findings. However, their study also reported a higher prevalence of AD among Hispanic and African American populations, which they attribute to a greater predisposition to diabetes. This risk is notably heightened during pregnancy due to hormonal imbalances. Nonetheless, we were not able to confirm this association in our analysis.

Although our dataset includes both male and female participants, as well as Hispanic and non-Hispanic individuals, we did not incorporate diabetes as a variable. Specifically, we did not examine whether any of the biomarkers in the ADNI database are indicative of diabetic conditions. As such, diabetes-related risk factors were neither considered nor analyzed.

6. Conclusion

Regarding age, the participants range from 55 to 90 ± 6.91 years. These results confirm findings from other years old. As shown in Figure 6, the mean age of indi- studies on the subject, including the observation that marviduals diagnosed with AD is 76.08 ± 6.91 years. We ital status is not a significant factor in the development mention this in connection to the diabetes-related dis- of the disease. cussion in [24]. While it is known that women face an One of the key aspects of this project has been the increased risk of developing diabetes during pregnancy, identification of four stages of Alzheimer’s disease, rangmost women in our dataset are likely post-menopausal ing from a healthy brain to one that is severely afected. and beyond childbearing age, making pregnancy-related These four stages allow the identification of two interdiabetes an unlikely contributing factor. Furthermore, mediate phases between health and severe illness, which since the ADNI dataset does not include variables related makes it possible to classify patients at an early stage of to pregnancy history, we were unable to investigate any the disease. This can facilitate treatment before it worspotential link between pregnancy-related diabetes and ens, thereby slowing its progression — one of the main AD development. objectives of this project. These four stages are associated with total brain volume, as it can be observed that the higher the disease stage, the lower the brain volume.

Afecting a large part of the population and expected to reach 139 million people by 2050, Alzheimer’s disease is one of the most common forms of neuropsychological dementia. It is a disease that is widely known and, at the same time, extremely misunderstood. Although it has been studied since the beginning of the 20th century, it still has neither a cure nor a full understanding of its progression, despite numerous studies and the substantial funding dedicated to the cause. Longitudinal databases such as ADNI facilitate research in this area by collecting data from patients, both healthy and afected by the disease, over the years; additionally, its public availability allows scientists to work freely on their projects.

Here, we have seen both in the literature review and in practice that machine learning algorithms, such as K-Means, applied to the ADNI database can help us understand relationships among features of diferent natures, from biological to physical, cognitive, and linguistic, when it comes to the detection of Alzheimer’s. It has been shown that women are more likely to sufer from the disease, representing 78.66% of diagnosed individuals.

Additionally, the average age of those afected is 76.08

Declaration on Generative AI

During the preparation of this work, the authors used

ChatGPT, Grammarly in order to: Grammar and spelling check, Paraphrase and reword. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content.

References

[1] M. Prince , R. Bryce , E. Albanese , A. Wimo , W. Ribeiro , C. P. Ferri , The global prevalence of dementia: a systematic review and metaanalysis , Alzheimer's & dementia 9 ( 2013 ) 63 - 75 .

[2] E. Joe , J. M. Ringman , Cognitive symptoms of alzheimer's disease: clinical management and prevention , Bmj 367 ( 2019 ).

[3] G. Capizzi , F. Bonanno , C. Napoli , A wavelet based prediction of wind and solar energy for longterm simulation of integrated generation systems , in: SPEEDAM 2010 - International Symposium Harbor perspectives in medicine 2 ( 2012 ) a006213 . on Power Electronics, Electrical Drives, Automa- [14] D. AlSaeed , S. F. Omar , Brain mri analysis for tion and Motion , 2010 , p. 586 - 592 . doi: 10 . 1109/ alzheimer's disease diagnosis using cnn-based feaSPEEDAM . 2010 . 5542259 . ture extraction and machine learning , Sensors 22

[4] F. Bonanno , G. Capizzi, C. Napoli , Some remarks ( 2022 ) 2911. on the application of rnn and prnn for the charge - [15] M. L. Giger , K. Suzuki , Computer-aided diagnosis, discharge simulation of advanced lithium-ions bat- in: Biomedical information technology, Elsevier, tery energy storage , in: SPEEDAM 2012 - 21st In- 2008 , pp. 359 - XXII . ternational Symposium on Power Electronics, Elec- [16] L. Billeci , A. Badolato , L. Bachi , A. Tonacci , Machine trical Drives, Automation and Motion , 2012 , p. 941 learning for the classification of alzheimer's disease - 945 . doi: 10 .1109/SPEEDAM. 2012 . 6264500 . and its prodromal stage using brain difusion tensor

[5] G. Lo Sciuto , G. Capizzi, R. Shikler , C. Napoli , Or- imaging data: A systematic review, Processes 8 ganic solar cells defects classification by using a (2020) 1071. new feature extraction algorithm and an ebnn with [17] S. Gupta , V. Saravanan , A. Choudhury , A. Alqahan innovative pruning algorithm , International tani, M. R. Abonazel , K. S. Babu , Supervised Journal of Intelligent Systems 36 ( 2021 ) 2443 - 2464 . computer-aided diagnosis (cad) methods for classidoi: 10 .1002/int.22386. fying alzheimer's disease-based neurodegenerative

[6] G. Capizzi , F. Bonanno , C. Napoli , Hybrid neu- disorders, Computational and Mathematical Methral networks architectures for soc and voltage pre- ods in Medicine 2022 ( 2022 ). diction of new generation batteries storage , in: [18] T. Jo , K. Nho , A. J. Saykin , Deep learning in 3rd International Conference on Clean Electrical alzheimer's disease: diagnostic classification and Power: Renewable Energy Resources Impact, IC- prognostic prediction using neuroimaging data , CEP 2011 , 2011 , p. 341 - 344 . doi: 10 .1109/ICCEP. Frontiers in aging neuroscience 11 ( 2019 ) 220 . 2011 . 6036301 . [19] S. Grueso , R. Viejo-Sobera , Machine learning meth-

[7] N. Brandizzi , S. Russo , R. Brociek , A. Wajda , First ods for predicting progression from mild cognitive studies to apply the theory of mind theory to green impairment to alzheimer's disease dementia: a sysand smart mobility by using gaussian area cluster- tematic review, Alzheimer's Research & Therapy ing , in: CEUR Workshop Proceedings , volume 3118 , 13 ( 2021 ) 1 - 29 . 2021 , p. 71 - 76 . [20] M. Syakur , B. Khotimah , E. Rochman , B. D. Satoto ,

[8] F. Fiani , V. Ponzi , S. Russo , Keeping eyes on the Integration k-means clustering method and elbow road: Understanding driver attention and its role method for identification of the best customer proin safe driving , in: CEUR Workshop Proceedings , ifle cluster, in: IOP conference series: materials volume 3695 , 2023 , p. 85 - 95 . science and engineering, volume 336 , IOP Publish-

[9] R. J. Jutten , L. I. Thompson , S. A. M. Sikkes , ing, 2018 , p. 012017 . P. Maruf , J. L. Molinuevo , H. Zetterberg , J. Alber, [21] J. Pareek , J. Jacob , Data compression and visualizaD . Faust, S. Gauthier, M. Gold, J. Harrison , A. Lee , tion using pca and t-sne, in: Advances in InformaP . J. Snyder , A neuropsychological perspective on tion Communication Technology and Computing, defining cognitive impairment in the clinical study Springer, 2021 , pp. 327 - 337 . of alzheimer's disease: Towards a more continuous [22] L. Van der Maaten , G. Hinton, Visualizing data approach ., Journal of Alzheimer's disease : JAD using t-sne. , Journal of machine learning research ( 2022 ). 9 ( 2008 ).

[10] K. Strimbu , J. A. Tavel , What are biomarkers?, Cur- [23] H. Liu , Z. Zhang , S.-w. Choi, K. M. Langa , Marital rent Opinion in HIV and AIDS 5 ( 2010 ) 463. status and dementia: Evidence from the health and

[11] A. Lloret , D. Esteve , M. - A. Lloret , A. Cervera-Ferri , retirement study, The Journals of Gerontology: B . Lopez , M. Nepomuceno , P. Monllor , When does Series B 75 ( 2020 ) 1783 - 1795 . alzheimer's disease really start? the role of biomark- [24] K. Rodriguez , M. Trevino , G. Castro, H. R. Roers , International journal of molecular sciences 20 driguez , G. Ortiz, J. Garza , S. Aleman , A. Alvarez, ( 2019 ) 5536 . et al., The efects of alzheimer's: Gender, race , &

[12] A. S. Fleisher , W. S. Houston, L. T. Eyler, S. Frye, ethnicity, DHR Proceedings 2 ( 2022 ) 15 - 19 . C. Jenkins , L. J. Thal , M. W. Bondi , Identification of alzheimer disease risk by functional magnetic resonance imaging , Archives of Neurology 62 ( 2005 ) 1881 - 1888 .

[13] K. A. Johnson , N. C. Fox , R. A. Sperling , W. E. Klunk , Brain imaging in alzheimer disease , Cold Spring