Available online at www.sciencedirect.

com
Available
Available online
online at
at www.sciencedirect.com
www.sciencedirect.com
Available online at www.sciencedirect.com

ScienceDirect
Procedia Computer Science 00 (2023) 000–000
Procedia Computer Science 00 www.elsevier.com/locate/procedia
Procedia Computer
Procedia Science
Computer Science 00 (2023)
225 (2023) 000–000
2273–2282
(2023) 000–000
www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia

27th International Conference on Knowledge-Based and Intelligent Information & Engineering

27th
27th International
International Conference
Conference on Systems (KES 2023)
on Knowledge-Based
Knowledge-Based and
and Intelligent
Intelligent Information
Information &
& Engineering
Engineering
Systems (KES 2023)
Systems (KES 2023)
AlzDiagnostics: A Mobile Alzheimer’s Diagnosis Solution
AlzDiagnostics: A Mobile Alzheimer’s Diagnosis Solution
Andreea Ciocan, Georgiana-Ingrid Stoleru∗, Daniel-Andrei Haivas, Bianca Ionela
∗∗
Andreea
Andreea Ciocan,
Ciocan, Georgiana-Ingrid
Georgiana-Ingrid Strătianu,Stoleru
Adrian
Stoleru ,, Daniel-Andrei
Iftene
Daniel-Andrei Haivas,
Haivas, Bianca
Bianca Ionela
Ionela
Strătianu,
Strătianu,
Faculty of Computer Science,
Adrian
AlexandruAdrian
Iftene
Iftene 700483 Iasi, Romania
Ioan Cuza University,
Faculty
Faculty of
of Computer
Computer Science,
Science, Alexandru
Alexandru Ioan
Ioan Cuza
Cuza University,
University, 700483
700483 Iasi,
Iasi, Romania
Romania

Abstract
Abstract
Abstract
Alzheimer’s disease (AD) is a neurodegenerative disorder that progressively affects cognitive function and which is the leading
cause
Alzheimer’sof dementia.
disease Early
(AD) diagnosis of AD is critical as it allows for timely intervention, and improves the individual’s quality of life,
Alzheimer’s disease (AD) is is aa neurodegenerative
neurodegenerative disorder
disorder that
that progressively
progressively affects
affects cognitive
cognitive function
function and
and which
which is
is the
the leading
leading
havingofthe
cause potentialEarly
to reduce the burden ofcritical
the disease on individuals, families, and society. This study presents AlzDiagnostics,
cause of dementia. Early diagnosis of AD is critical as it allows for timely intervention, and improves the individual’s quality of
dementia. diagnosis of AD is as it allows for timely intervention, and improves the individual’s quality of life,
life,
ahaving
mobile theapplication
potential towhose
reduce primary
the goalofisthe
burden to disease
offer a user-friendly
on individuals, tool for theand
families, early diagnosis
society. This of AD,presents
study integrating established
AlzDiagnostics,
having the potential to reduce the burden of the disease on individuals, families, and society. This study presents AlzDiagnostics,
clinical
mobiletools, state-of-the-art machine goallearning totechniques, and multiple diagnostic approaches. Furthermore, the application seeks
aa mobile application
application whose primary
whose primary goal is is to offer aa user-friendly
offer user-friendly tool
tool for the
for the early
early diagnosis
diagnosis of AD,
of AD, integrating
integrating established
established
to assesstools,
clinical the effectiveness andmachine
accuracylearning
of the Mini-Mental State Examination (MMSE) test, machine learning models trained on
clinical tools, state-of-the-art machine learning techniques, and multiple diagnostic approaches. Furthermore, the application seeks
state-of-the-art techniques, and multiple diagnostic approaches. Furthermore, the application seeks
MRI
to images, as well as ones trained on of
clinical data, evaluating their contributions to enhance early diagnosis. Choosing the best
to assess the effectiveness and accuracy of the Mini-Mental State Examination (MMSE) test, machine learning models trained on
assess the effectiveness and accuracy the Mini-Mental State Examination (MMSE) test, machine learning models trained on
classification
images, models
MRI images, as well
well asforones
as the diagnosis
trained ononmodule
clinicalinvolved the conduct
data, evaluating
evaluating theirofcontributions
numerous classification
contributions to enhance experiments
enhance early
early diagnosis.on Choosing
an extensive
theMRI
best
MRI as ones trained clinical data, their to diagnosis. Choosing the best
dataset obtained
classification from Kaggle, which consisted of 6400 MRI images fromnumerous
different sources including websites, hospitals, and public
classification models
models for for the
the diagnosis
diagnosis module
module involved
involved thethe conduct
conduct ofof numerous classification
classification experiments
experiments on on an
an extensive
extensive MRI
MRI
repositories.
dataset These experiments encompassed the exploration of various parameter configurations, leading tohospitals,
the identification of
dataset obtained from Kaggle, which consisted of 6400 MRI images from different sources including websites, hospitals, and
obtained from Kaggle, which consisted of 6400 MRI images from different sources including websites, and public
public
arepositories.
top-performing model that exhibited an accuracy rate of 98%. Additionally, extensive experiments were toconducted on clinical
repositories. These experiments encompassed the exploration of various parameter configurations, leading to the identification of
These experiments encompassed the exploration of various parameter configurations, leading the identification of
adata retrieved from the Alzheimer’s
a top-performing
top-performing modelmodel that
that exhibited
Diseaseaccuracy
exhibited an
Neuroimaging
an accuracy rate rate of
Initiative
of 98%.
(ADNI) dataset, which
98%. Additionally,
Additionally, extensive
comprised entries
extensive experiments
experiments were
from 2425 patients.
were conducted
conducted onon clinical
clinical
By incorporating
data features such as gender, marital status, education, race, and MMSE scores, a second machine learning model was
data retrieved
retrieved from
from the
the Alzheimer’s
Alzheimer’s Disease
Disease Neuroimaging
Neuroimaging Initiative
Initiative (ADNI)
(ADNI) dataset,
dataset, which
which comprised
comprised entries
entries from
from 2425
2425 patients.
patients.
developed,
By yielding an accuracy level of 76%.
By incorporating features such as gender, marital status, education, race, and MMSE scores, a second machine learning model
incorporating features such as gender, marital status, education, race, and MMSE scores, a second machine learning model was
was
developed, yielding
developed, yielding anan accuracy
accuracy level
level of
of 76%.
76%.
© 2023 The Authors. Published by Elsevier B.V.
© 2023
This The
is an Authors.
open accessPublished
article under by Elsevier B.V.
the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
©
© 2023
2023 The
The Authors.
Authors. Published
Published by
by Elsevier
Elsevier B.V.
B.V.
This is an
Peer-review open access
under article under
responsibility of the scientific
the CC BY-NC-ND committee license (https://creativecommons.org/licenses/by-nc-nd/4.0)
of the KES International.
This is an open
This is an open access
access article under the
article underofthethe CC BY-NC-ND
CCscientific
BY-NC-ND license
license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility committee of the 27th International Conference on Knowledge Based and
Peer-reviewInformation
Intelligent
Peer-review under responsibility
under responsibility of the
and Engineering
of the scientific
scientific
Systems committee of
committee of the
the KES
KES International.
Keywords: Alzheimer’s Disease Diagnosis; Mild Cognitive Impairment; MobileInternational.
Application; Machine Learning; Cognitive Assessment;
Neuroimaging
Keywords:
Keywords: Alzheimer’s Disease Diagnosis; Mild Cognitive Impairment; Mobile Application;
Alzheimer’s Disease Diagnosis; Mild Cognitive Impairment; Mobile Application; Machine
Machine Learning;
Learning; Cognitive
Cognitive Assessment;
Assessment;
Neuroimaging
Neuroimaging

1. Introduction
1.
1. Introduction
Introduction
Alzheimer’s disease (AD) is a neurodegenerative disorder characterized by a gradual decline in cognitive function,
particularly affecting
Alzheimer’s
Alzheimer’s memory,
disease
disease (AD) isthinking,
(AD) is and behavior.
aa neurodegenerative
neurodegenerative It is thecharacterized
disorder
disorder leading causeby
characterized byofaadementia,
gradual accounting
gradual decline
decline in for an estimated
in cognitive
cognitive function,
function,
particularly affecting memory, thinking, and behavior. It is the leading cause of dementia, accounting for an estimated
particularly affecting memory, thinking, and behavior. It is the leading cause of dementia, accounting for an estimated
∗Corresponding author. Tel.: +0-000-000-0000 ; fax: +0-000-000-0000.
∗E-mail address: ingridstoleru@gmail.com
Corresponding
∗Corresponding author.
author. Tel.: +0-000-000-0000 ;; fax:
Tel.: +0-000-000-0000 fax: +0-000-000-0000.
+0-000-000-0000.
E-mail
E-mail address:
address: ingridstoleru@gmail.com
ingridstoleru@gmail.com
1877-0509 © 2023 The Authors. Published by Elsevier B.V.
1877-0509
This © 2023
is an open Thearticle
access Authors. Published
under by Elsevier B.V.
the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
1877-0509
1877-0509 © 2023 The
The Authors. Published by Elsevier B.V.
This is an ©
Peer-review 2023
open
underaccess Authors.
responsibility Published
article under
of the the by
CCElsevier
scientific B.V.oflicense
BY-NC-ND
committee the KES(https://creativecommons.org/licenses/by-nc-nd/4.0)
International.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review
This is an open under responsibility
access article under the CCof BY-NC-ND
the scientific committee
license of the 27th International Conference on Knowledge Based
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under
under responsibility
and Intelligent
Peer-review Informationof
responsibility the
the scientific
ofand committee
Engineering
scientific of
of the
Systems
committee the KES
KES International.
International.
10.1016/j.procs.2023.10.218
2274 Andreea Ciocan et al. / Procedia Computer Science 225 (2023) 2273–2282
Andreea Ciocan et al. / Procedia Computer Science 00 (2023) 000–000

60% to 80% of the cases [4]. The lifespan of individuals suffering from this condition can fluctuate widely, depending
on a range of factors including the age of onset, overall health, and the progression of the disease. Those with AD can
survive for a period of more than 20 years after diagnosis [5]. The prolonged duration of AD contributes significantly
to the financial and social costs of the illness because much of the time is spent in a state of significant disability
and dependence. Taking these into account, early diagnosis of AD represents a subject of great scientific interest,
as it allows for timely initiations of treatments that can improve the individual’s quality of life. Mobile applications
developed for diagnosing AD provide a convenient and accessible means of early diagnosis, allowing for timely
intervention, with the potential of improving cost-effectiveness. In particular, AlzDiagnostics is designed with the
objective of offering a user-friendly tool for the early diagnosis of AD. The application utilizes a comprehensive
approach to diagnosis, incorporating the MMSE test, as well as machine learning-based methods that leverage either
MRI images or clinical data. However, it is important to note that any results should be followed up with professional
health support.

1.1. Related Work

Zorluoglu et al. [37] proposed a mobile application designed for the cognitive screening of dementia, by employing
several cognitive tests, including MMSE and the Clock Drawing Test (CDT) [26]. As opposed to this solution, AlzDi-
agnostics provides support for diagnosis based also on neuroimaging, by employing artificial intelligence approaches,
which can contribute to the more accurate and reliable diagnosis of AD. The authors report promising results in testing
the application on a sample of eight elderly individuals. The MMSE app [43] is a digital version of the MMSE test,
which was released on the Google Play store in 2016. On completion of the questionnaire, the user is assigned a score
of up to 30 points, based on the provided answers. The main aim of the application is to assist users in evaluating their
cognitive health and to advise them to seek additional medical assistance and support if needed.
Nirjon et al. [28] describe the development and testing of MOBI-COG, a mobile application that aims to provide
complete automation of the dementia screening test called the Mini-Cog test [7]. The application stores the history of
the test results, so they can be shared with the caregivers for a better understanding of the patient’s condition. Another
mobile application that allows users to share the test results with a healthcare professional, rather than generating
a score itself is BrainCheck [14]. This is supposed to be completed both by the evaluated subject and by a close
informant of the person being assessed. It involves a CDT and seven questions derived from the 7-item version of the
Informant Questionnaire on Cognitive Decline in the Elderly (IQCODE) [8]. According to [14], it proved to have a
sensitivity of 97.4% and specificity of 81.6% in identifying cases of dementia and MCI.
CogniDecline [47] is another mobile application, which aims to provide early detection of dementia, but based on
changes in mobile phone usage patterns. The authors suggest that changes in mobile phone usage could represent a
useful indicator of cognitive decline. However, this study has two main limitations. The first one is that the application
was tested on a small group of only 13 participants, which implies the need for further testing with larger populations.
The second limitation is represented by the ethical considerations involved in collecting and analyzing personal data.

2. Proposed Solution: AlzDiagnostics

AlzDiagnostics has a mission to raise awareness about the crucial importance of early detection of Alzheimer’s dis-
ease. It strives to guide users through the process of conducting cognitive tests to assess whether they are experiencing
mild cognitive impairment (MCI) or AD symptoms. Furthermore, AlzDiagnostics attempts to identify anomalies in
MRI images collected from users by employing a machine learning model. These anomalies can be strong indicators
of the onset and evolution of MCI, respectively AD, therefore contributing to the early detection of this condition.
Early diagnosis enables patients, according to [36], to access timely and appropriate treatment and support services,
which can help them effectively manage their symptoms and enhance their overall quality of life. In a study conducted
by Rasmussen et al. [13], a US claims database was analyzed to investigate the outcomes of treatment for individuals
aged 65 to 100 who were newly diagnosed with Alzheimer’s disease. The results showed that those who received
treatment had higher rates of survival and a 20% reduced likelihood of institutionalization, even after adjusting for
other medical conditions.
 Andreea Ciocan et al. / Procedia Computer Science 225 (2023) 2273–2282 2275
Andreea Ciocan et al. / Procedia Computer Science 00 (2023) 000–000

The proposed solution enables users to monitor their health status continuously using the following methods: (1)
Firstly, they can conduct an MMSE test. This test was first proposed by Marshal F. Folstein in 1975 [17] as a brief
screening measure for cognitive impairment. If the results indicate a possible cognitive decline, a notification is dis-
played on the patient’s screen. Moreover, the results of the tests are saved and can be accessed as a historical record
to monitor progress over time. (2) Secondly, the mobile application provides users with the ability to submit their own
MRI scans for evaluation. The evaluation process involves utilizing a machine learning model, which has been trained
on a larger dataset incorporating both AD and cognitively normal (CN) patients. (3) Lastly, the mobile application
employs an additional machine learning model, which has the capability to generate a potential diagnosis, by utilizing
both the MMSE test score, along with patient-specific information, such as age, marital status, and ethnicity. Upon
receiving an AD diagnosis from any of the diagnosis modules, the application generates a notification with a dis-
played disclaimer, explicitly stating that the results obtained should not be regarded as a definitive medical diagnosis.
The purpose of this measure is to highlight the importance of seeking a comprehensive evaluation from a qualified
healthcare professional to ensure an accurate assessment.

2.1. Architecture

AlzDiagnostics is designed to offer users an intuitive and user-friendly experience in the diagnosis of Alzheimer’s
disease. The application’s back-end component assumes responsibility for rendering essential services, including data
persistence, web services, and data analysis through artificial intelligence models. As a critical element, this com-
ponent plays a vital role in the product’s overall functionality, allowing for the delivery of precise and dependable
information pertaining to the cognitive health of users. The front-end component of the mobile application represents
the primary interface with which users engage. It encompasses a straightforward and user-friendly design that enables
individuals of all technical backgrounds to access the solution’s functionalities with ease and efficiency, as depicted
in Figure 1.

Fig. 1. Main Use Cases of the Application

2.1.1. Authentication Module

The application includes an authentication module. Therefore, any patient can create a user account and log in
to view their personal data in the application. It is important to note that when creating the account, the user must
provide, in addition to username/email address and password, personal information such as gender, marital status,
education, and race. All fields are mandatory to be completed because the corresponding information is further fed to
the machine learning model, along with the cognitive test result, in order to provide an accurate diagnosis.

2.1.2. Cognitive Test Evaluation Module

An important diagnosis component is the MMSE cognitive test. The user receives a set of questions to which they
must provide answers, as depicted in Figure 2. Further elaboration on this topic is expounded upon in 2.4.1.
2276 Andreea Ciocan et al. / Procedia Computer Science 225 (2023) 2273–2282
Andreea Ciocan et al. / Procedia Computer Science 00 (2023) 000–000

Fig. 2. Examples of Questions from the MMSE Cognitive Test

The user’s performance in the MMSE cognitive test is assessed based on the number of correct answers provided,
resulting in a score of up to 30 points, which is then recorded in their personal profile. Additionally, patients have the
ability to access a record of their previous test results in order to monitor any cognitive decline.

2.1.3. MRI Evaluation Module

This diagnosis component offers the capability to perform scanning and analysis of MRI through the utilization of
a pre-trained machine learning model. To initiate analysis, the user is required to scan the medical image (MRI) using
their phone camera and transmit it via the application to the server. Once the medical image is evaluated by the model,
the outcome is returned to the user, as depicted in Figure 3. Similar to the cognitive test results, the output of the MRI
evaluation is also saved in the database, to be retrievable as a historical record.

Fig. 3. User Experience for the Flow of Submitting and Analyzing an MRI

2.1.4. Aggregated Data Evaluation Module

The diagnosis of a user can be ascertained by utilizing aggregated data, which includes two data points, i.e., the
outcome of the MMSE cognitive test, and personalized user information such as gender, marital status, education, and
race, which were originally provided during the account creation process. This data is subsequently fed into a machine
learning model, which was trained on a larger set of data containing the aforementioned features.

2.2. Application Design

The subsequent section outlines the key components of the mobile application, from the application design point
of view, namely the front-end and the back-end modules.
 Andreea Ciocan et al. / Procedia Computer Science 225 (2023) 2273–2282 2277
Andreea Ciocan et al. / Procedia Computer Science 00 (2023) 000–000

2.2.1. Front-End Component

The user interface is the component of the mobile application that displays the data and the functionalities described
in 2.1 to the user, enabling user interactions. This component only sends requests to the back-end component, which
receives the requests and is responsible for processing the tasks, performing data processing, storage, and retrieval.

2.2.2. Back-End Component

Specifically, the main API of the application performs the following functionalities: (1) It features an authentication
module that facilitates the creation of user accounts and validates user authentication. (2) It generates an MMSE test
and automatically evaluates users’ responses. This evaluation is based on the standardized scoring system of the
MMSE. (3) It receives user-specific information, including data provided during account creation and potential MRI
scans uploaded by the user.
To enhance scalability and system maintenance, two additional web services have been used within the back-end
component: (1) The first one serves the function of receiving MRI data from the main API, subsequently providing
relevant results to the user with the help of a machine-learning model. This service operates as a mediator between
the user and the main API, allowing for data exchange and enhancing the overall efficiency of the application. (2)
The second service is tasked with aggregating and processing user data, including personal information and results
from cognitive tests. Based on this data and by leveraging another machine learning model, the application provides
insights into the cognitive functions of the user.
Finally, the user-specific data, the MMSE test results, as well as the MRI evaluation results are stored in a database
to ensure data persistence. The database provides access to user data only to the microservices, while also maintaining
security by hashing passwords and avoiding transmission of user-specific data to the external client (front-end mobile
application). A clearer illustration of how the components interact with one another is presented in Figure 4.

Fig. 4. AlzDiagnostics Back-End Design

2.3. Implementation Details

The front-end component of the mobile application was developed using standard web application technologies,
including HTML, CSS, and JS. In addition to these, the most important utilized technology was React Native [27],
which is a framework for developing mobile applications for both Android and iOS. It was used to build the user
interface and handle JavaScript code, which makes calls to the main API.
The back-end consists of several components, with the central component being a Java-based API developed using
Java Enterprise [32] technologies. This API allows the front-end to access different endpoints as needed via a web
service. In order to develop a highly scalable application that can adapt to changing requirements, a dependency injec-
tion design was employed. Additionally, the application needs to be secure, which is why several security components
2278 Andreea Ciocan et al. / Procedia Computer Science 225 (2023) 2273–2282
Andreea Ciocan et al. / Procedia Computer Science 00 (2023) 000–000

from Java EE were used. Among these, it is worth mentioning the use of JSON Web Tokens (JWT) [20]. A JWT is a
compact, URL-safe means of securely transmitting data as a JSON object, used for authentication and authorization
purposes. It typically contains a digitally signed payload that can be verified and trusted. Additionally, for faster code
development, Lombok [25] was used, which is a library that provides annotations to automatically generate boilerplate
code, such as getters, setters, and constructors, making the code shorter and more concise. Also, Log4J [3] was used to
provide logs through which the system status can be monitored. This library is a logging utility that allows developers
to easily log messages from their applications to various output targets, such as files, databases, or the console, to
facilitate troubleshooting and debugging.
For storing the data, a relational database, Postgres [35], was used. It interacts only with the aforementioned main
API using Java Persistence API (JPA) [31]. This is a specification that provides an Object-Relational Mapping (ORM)
framework for mapping objects to relational database tables, allowing developers to interact with databases using Java
code in a seamless and convenient manner. Thus, by using an ORM, there is no need to write SQL code, while also
providing a more straightforward approach to entity and data management.
To aid in the diagnosis, the application leveraged two ML models that were developed using Python programming
language. Additionally, for the development of these models, there were used Pandas [33] and Numpy [29] for the
data filtering part and libraries such as TensorFlow [44] and Keras [22] for the training part. Finally, the Flask [16]
web framework was used to expose them as web services that can be used by the main API.
To ensure the proper functionality of the main API, functional tests were developed and executed, resulting in a
Java code coverage of at least 80%. In addition to functional testing, non-functional testing was employed, specifically,
load testing and stress testing. This was done using Jmeter [2], which is an open-source software tool used for load
testing, performance testing, and functional testing of web applications. It allows developers to simulate and measure
the performance of their applications under different conditions. This tool was used for verifying that the system can
support a high number of users and can withstand periods of high usage and traffic.
In addition to these, other tools were used in the development process, such as: (1) SonarQube [40], which ensured
that the tested code is clean, bug-free, and properly tested, (2) pgAdmin4 [34], which was used for managing infor-
mation in the database, (3) Android Studio [12], which was used to simulate the mobile application environment, (4)
IntelliJ [9], PyCharm [10] and Visual Studio Code [42], which facilitated fast and organized code development.

2.4. AD Diagnosis Module

MMSE based Diagnosis - One of the primary features of the application is the MMSE diagnosis module that
allows the users to conduct regularly a MMSE test. The application maintains a history of test results to enable
continuous monitoring of the patient’s progress. MRI-based Diagnosis - The categorization process of medical images
is conducted through the utilization of a machine learning model that has been trained on a vast collection of medical
images. The selection of this model was based on a series of experiments that are comprehensively expounded upon
in 2.4.2. MMSE and User Information Diagnosis - The third diagnosis solution provided by the application consists in
employing data from two data sources, specifically the result of the cognitive tests and the personal user information:
gender, marital status, education, and race. This data is further used as input for the machine learning model, which
provides the diagnostic result.

2.4.1. Evaluated Biomarkers

In order to provide a diagnosis through each of the aforementioned methods, the following biomarkers were con-
sidered: cognitive tests (MMSE), neuroimaging biomarkers (MRI images), and personal background variables of the
participant (such as gender, marital status, education, and race). Stoleru et al. [41] conducted a comprehensive analy-
sis of these biomarkers, which have been employed in state-of-the-art research to construct highly accurate diagnostic
models for AD, highlighting their respective advantages and limitations. Each of the aforementioned biomarkers is
further detailed.
Neuropsychological Tests - The neuropsychological assessment employed the use of the MMSE test. This is a
widely used tool for assessing cognitive functions, including memory, attention, language, and orientation. The user
receives a set of either personal or common questions to which they must provide answers, such as What year is this?,
What is today’s date?, What country are you from?, What city are you from?. The test further comprises a series of
basic tasks, the responses to which are evaluated through the mobile application. The following are some examples
 Andreea Ciocan et al. / Procedia Computer Science 225 (2023) 2273–2282 2279
Andreea Ciocan et al. / Procedia Computer Science 00 (2023) 000–000

of such tasks: Recognition of a pencil from an image, Spelling the word ”world”, Naming three presented objects in
their given order, Recognition of a given phrase. The score, which can range up to 30 points, is determined based on
the number of correct answers provided.
The diagnostic module of the application which utilizes the MMSE test delivers a response aligned with the thresh-
olds outlined in [13]. According to these established criteria, a score of 25 or above is generally indicative of cognitive
functioning within the normal range. If the score is below 24, the result is usually considered to be abnormal, indi-
cating possible cognitive impairment. The MMSE test confers several advantages, including its ease to conduct, short
completion time, lack of requirement for specialized personnel, cost-effectiveness, non-invasiveness and widespread
availability. The accuracy, sensitivity, and specificity of this test may vary depending on factors such as the characteris-
tics of the population being tested and the chosen cutoff score. Multiple studies have explored its utility as a screening
tool for the diagnosis of cognitive decline. According to Kahle-Wrobleski et al. [21], the MMSE demonstrated high
diagnostic accuracy in identifying dementia among individuals aged 90 and above, encompassing various age and
education groups. The study reported AUC (Area Under the Curve) values ranging from 0.82, the lowest observed
in the 97 and older age group with a high school education or less, to 0.98, the highest observed in the 90 to 93 age
group with vocational or some college education.
Neuroimaging Biomarkers - MRI is a medical imaging technique that uses strong magnetic fields and radio waves
to produce detailed images of the inside of the body. According to [46], the MRI technique is the most used in
AD diagnosis due to its high spatial resolution, which allows the difference between two arbitrarily similar but not
identical tissues and multiparametric acquisitions to be distinguished. MRI does not involve X-rays or the use of
ionizing radiation, it is non-invasive and has no significant adverse health effects.

2.4.2. Methods
Data Collection - Two different datasets were used for training the machine learning models: (1) An Alzheimer
MRI Preprocessed Dataset obtained from Kaggle [23] employed for the task of classifying MRI images. It was re-
portedly compiled from multiple sources, including various websites, hospitals, and public repositories. The dataset
consists of 6,400 images and has four classes, respectively: Mild Demented, Moderate Demented, Non-Demented,
and Very Mild Demented. (2) Clinical Data obtained from the ADNI dataset [1], which is a subset of data comprised
of entries from ADNI1, ADNI2, ADNI3, and ADNIGO. It consists of 16,202 entries collected from 2,425 patients,
yielding three primary outcomes: Mild Cognitive Impairment (MCI), Alzheimer’s Disease (AD), and Normal Cogni-
tive (CN). This dataset contains a wide range of information related to participants in the study, including cognitive
assessments (results of cognitive tests such as MMSE) and demographic information, such as age, sex, race, education,
and marital status.
Data Preprocessing - (1) Neuroimaging Data - The images were resized into 48 × 48 pixels. (2) Clinical Data -
During the preprocessing step, columns that did not contain personal user data (e.g., the id of the entry, the initial
dataset to which the entry belongs), as well as ones containing results of cognitive tests other than MMSE were
removed. This resulted in a selection of variables including age, marital status, MMSE score, education level, and
gender. The aforementioned variables were used in the training step of the model and are requested from the user
upon account creation to provide a more accurate diagnosis.
Machine Learning Models - (1) MRI - An MRI scan is categorized into one of four classes, which comprise
three stages of dementia and one class for the absence of dementia. To assess the efficacy of the model, the dataset
was partitioned into two distinct subsets, consisting of an 80% training set and a 20% test set, respectively. The
MRI classification model utilized in this study was a Convolutional Neural Network (CNN) architecture. The model
consists of a rescaling layer, followed by three convolutional layers, with each convolutional layer accompanied by
a max pooling layer. Finally, a flattened layer is added, followed by two dense layers. To optimize the model, it
was compiled with the Adam optimizer. (2) MMSE and User Information Diagnosis - In order to diagnose user
information, a set of demographic variables including gender, marital status, education, race, and the MMSE score are
employed. To accomplish this task, a suite of machine learning algorithms, specifically Naive Bayes, Random Forest,
and XGBoost, are utilized. These models have been trained to accurately classify input data into three categories, CN,
MCI, and AD.
Data Statistics - The dataset comprised of clinical data, which was used for the training of the machine learning
model, is presented below, stratified by patients’ age, gender, marital status, and level of education. Figure 5 indicates
the age range of patients in the dataset, spanning from 55 to 100 years old, with a majority falling between the ages of
2280 Andreea Ciocan et al. / Procedia Computer Science 225 (2023) 2273–2282
Andreea Ciocan et al. / Procedia Computer Science 00 (2023) 000–000

70 and 80 years. Advanced age is one of the strongest risk factors for developing AD. While it can occur at any age,
the likelihood of developing this condition increases as a person gets older [6, 15, 24, 39].

Fig. 5. Distribution of Subjects by Age

Regarding the distribution of patients by gender, there is a slight majority of male subjects in the dataset (the
difference is under 9%). Regarding an individual’s marital status, there exist five distinct categories, namely married,
widowed, divorced, never married, and unknown. The distribution of patients by their marital status is illustrated in
Figure 6. It has been observed that this factor can impact one’s susceptibility to AD, with a higher risk being observed
amongst those who have never been married [18, 19, 38].

Several studies have established a correlation between lower educational attainment and a heightened susceptibility
to dementia. According to Figure 7, a majority number of patients in the dataset have acquired a higher level of
education [11, 37, 45].

2.5. Results

For the MRI model, the dataset was split into two subsets, consisting of an 80% training set and a 20% test
set. Experiments were conducted by testing different parameter settings such as varying the number of epochs and
using multiple optimizers. The model achieved its highest accuracy of 98% after 20 epochs while utilizing the Adam
optimizer. Considering its superior performance within the specified configurations, it was selected as the MRI model
to integrate within the diagnosis module.
For the clinical data, in order to determine the best machine learning model which would be integrated into the di-
agnosis module, several experiments have been conducted, employing the following machine learning models: Naive
 Andreea Ciocan et al. / Procedia Computer Science 225 (2023) 2273–2282 2281
Andreea Ciocan et al. / Procedia Computer Science 00 (2023) 000–000

Bayes, Random Forest, and XGBoost. These experiments involved multiclass classification, where the machine learn-
ing model was trained to classify input data into three categories, namely CN, MCI, and AD. The chosen features for
these experiments included gender, marital status, education, race, as well as the MMSE score.
In the case of Naive Bayes classification, there were performed experiments with different parameter settings. How-
ever, no improvement was achieved beyond a validation accuracy of 61%. Conversely, the Random Forest classifier
led to a validation accuracy of 72% when providing a diagnosis. The best results were obtained using the XGBoost
algorithm for the entire set of aforementioned features, i.e. a validation accuracy equal to 76%. The parameters em-
ployed for the XGBoost model were determined via a grid search across a predefined parameter space. This search
yielded a gamma value of 0, a learning rate of 0.1, and a maximum tree depth of 3. The model was further assessed
on OASIS (Open Access Series Of Imaging Studies) [30], an independent dataset, where it exhibited a test accuracy
equal to 76%. Consequently, the XGBoost model, trained on the complete set of aforementioned features (gender,
marital status, education, race, and MMSE score), was selected as the second model to integrate into the diagnosis
module.

3. Conclusion and Future Work

AlzDiagnostics provides a convenient and accessible way for users to evaluate their neurological health and detect
Alzheimer’s disease using cognitive tests, MRI images, and machine learning approaches. The application has the
potential to enhance the early detection of Alzheimer’s disease by incorporating a diagnosis module encompassing
multiple classification approaches. These approaches include utilizing the MMSE score for diagnosis, employing a
machine learning model trained on MRI images, which demonstrated a high validation accuracy of 98% on a Kaggle
dataset, and utilizing a separate machine learning model trained on clinical data, achieving a validation accuracy of
76% when evaluated on both the ADNI dataset and an independent test dataset. An early diagnosis can improve
patient outcomes, as it allows for the prompt initiation of appropriate medical interventions and treatments, slowing
cognitive decline and therefore, reducing the burden on the healthcare system. AlzDiagnostics allows users to conduct
cognitive tests anytime, anywhere, and at the same time, as opposed to other diagnostic tools, it provides support for
diagnosing AD based on MRI images. This is done by employing a machine learning model with 98% validation
accuracy. However, these results should be further communicated to qualified healthcare professionals.
There are several areas for future work on this application, including improving the machine learning algorithms,
integrating it with other data sources, enhancing the user experience, and expanding it to new platforms. By addressing
these areas, the solution has the potential to become an even more powerful tool for detecting AD.

References

[1] ADNI, 2017. Adni (alzheimer’s disease neuroimaging initiative datasets). URL: https://adni.loni.usc.edu/data-samples/. accessed
on 14 April 2023.
[2] Apache, 2022. Jmeter. URL: https://jmeter.apache.org/. accessed on 17 April 2023.
[3] Apache, 2023. Logging log4j. URL: https://logging.apache.org/log4j/2.x/. accessed on 17 April 2023.
[4] Association, A., 2023a. 2023 alzheimer’s disease facts and figures. URL: https://www.alz.org/media/documents/
alzheimers-facts-and-figures.pdf. accessed on 17 April 2023.
[5] Association, A., 2023b. Stages of alzheimer’s. URL: https://www.alz.org/alzheimers-dementia/stages. accessed on 17 April 2023.
[6] Association, A., 2023c. What is alzheimer’s disease? URL: https://www.alz.org/alzheimers-dementia/what-is-alzheimers.
accessed on 15 April 2023.
[7] Borson, S., Scanlan, J.M., Brush, M., Vitaliano, P.P., Dokmak, A., 2000. The mini-cog: a cognitive ‘vital signs’ measure for dementia screening
in multi-lingual elderly. International Journal of Geriatric Psychiatry 15.
[8] Bosma, H., van Boxtel, M.P.J., Ponds, R.W., Houx, P.J., Jolles, J., 2003. Education and age-related cognitive decline: The contribution of
mental workload. Educational Gerontology 29, 165–173.
[9] Brains, J., 2023a. Jetbrains intellij. URL: https://www.jetbrains.com/idea/. accessed on 17 April 2023.
[10] Brains, J., 2023b. Pycharm - python ide. URL: https://www.jetbrains.com/pycharm/. accessed on 17 April 2023.
[11] of Cambridge, U., 2010. Why more education lowers dementia risk. URL: https://www.cam.ac.uk/research/news/
why-more-education-lowers-dementia-risk. accessed on 18 April 2023.
[12] Developers, A., 2023. Android studio. URL: https://developer.android.com/. accessed on 17 April 2023.
[13] Direct, H., 2023. Mini-mental state examination (mmse). URL: https://www.healthdirect.gov.au/
mini-mental-state-examination-mmse. accessed on 15 April 2023.
2282 Andreea Ciocan et al. / Procedia Computer Science 225 (2023) 2273–2282
Andreea Ciocan et al. / Procedia Computer Science 00 (2023) 000–000

[14] Ehrensperger, M.M., Taylor, K.I., Berres, M., Foldi, N.S., Dellenbach, M., Bopp, I., Gold, G., von Gunten, A., Inglin, D., Müri, R.M., Rüegger,
B., Kressig, R.W., Monsch, A.U., 2014. Braincheck – a very brief tool to detect incipient cognitive decline: optimized case-finding combining
patient- and informant-based data. Alzheimer’s Research & Therapy 6.
[15] Ferri, C.P., Prince, M.J., Brayne, C., Brodaty, H., Fratiglioni, L., Ganguli, M., Hall, K.S., Hasegawa, K.A., Hendrie, H.C., Huang, Y., Jorm,
A.F., Mathers, C.D., Menezes, P.R., Rimmer, E., Scazufca, M., 2005. Global prevalence of dementia: a delphi consensus study. The Lancet
366, 2112–2117.
[16] Flask, 2010. Flask. URL: https://flask.palletsprojects.com/en/2.2.x/. accessed on 17 April 2023.
[17] Folstein, M.F., Folstein, S.E.B., McHugh, P.R., 1975. ”mini-mental state”. a practical method for grading the cognitive state of patients for the
clinician. Journal of psychiatric research 12 (3), 189–98.
[18] Helmer, C., Damon, D., Letenneur, L., Fabrigoule, C., Barberger-Gateau, P., Lafont, S., Fuhrer, R., Antonucci, T., Commenges, D., Orgogozo,
J., Dartigues, J., 1999a. Marital status and risk of alzheimer’s disease. Neurology 53, 1953–1953. URL: https://n.neurology.org/
content/53/9/1953, doi:10.1212/WNL.53.9.1953, arXiv:https://n.neurology.org/content/53/9/1953.full.pdf.
[19] Helmer, C., Damon, D., Letenneur, L., Fabrigoule, C., Barberger-Gateau, P., Lafont, S., Fuhrer, R., Antonucci, T., Commenges, D., Orgogozo,
J.M., Dartigues, J.F., 1999b. Marital status and risk of alzheimer’s disease: a french population-based cohort study. URL: https://pubmed.
ncbi.nlm.nih.gov/10599764/. pMID: 10599764.
[20] JWT, 2018. Json web tokens. URL: https://jwt.io/. accessed on 17 April 2023.
[21] Kahle-Wrobleski, K., Corrada, M.M., Li, B., Kawas, C.H., 2007. Sensitivity and specificity of the mini-mental state examination for identifying
dementia in the oldest-old: The 90+ study. Journal of the American Geriatrics Society 55.
[22] Keras, 2023. Keras. URL: https://keras.io/. accessed on 17 April 2023.
[23] Kumar, S., 2022. Alzheimer mri preprocessed dataset. URL: https://www.kaggle.com/datasets/sachinkumar413/
alzheimer-mri-dataset. accessed on 15 April 2023.
[24] Livingston, G., Sommerlad, A., Orgeta, V., Costafreda, S.G., Huntley, J., Ames, D., Ballard, C.G., Banerjee, S., Burns, A., Cohen-Mansfield, J.,
Cooper, C., Fox, N., Gitlin, L.N., Howard, R.J., Kales, H.C., Larson, E.B., Ritchie, K., Rockwood, K., Sampson, E.L., Samus, Q.M., Schneider,
L.S., Selbæk, G., Teri, L., Mukadam, N., 2017. Dementia prevention, intervention, and care. The Lancet 390, 2673–2734.
[25] Lombok, P., 2023. Project lombok. URL: https://projectlombok.org/. accessed on 17 April 2023.
[26] Mendez, M.F., Ala, T.A., Underwood, K.L., 1992. Development of scoring criteria for the clock drawing task in alzheimer’s disease. Journal
of the American Geriatrics Society 40.
[27] Native, R., 2023. React native. URL: https://reactnative.dev. accessed on 15 April 2023.
[28] Nirjon, S.M.S., Emi, I.A., Mondol, M.A.S., Salekin, A., Stankovic, J.A., 2014. Mobi-cog: A mobile application for instant screening of
dementia using the mini-cog test. Proceedings of the Wireless Health 2014 on National Institutes of Health .
[29] Numpy.org, 2023. Numpy. URL: https://numpy.org/. accessed on 17 April 2023.
[30] OASIS, 2023. Oasis-2: Longitudinal:. URL: https://www.oasis-brains.org/#data. accessed on 21 April 2023.
[31] Oracle, 2013. Java persistance api. URL: https://docs.oracle.com/javaee/6/tutorial/doc/bnbpz.html. accessed on 17 April
2023.
[32] Oracle, 2023. Java enterprise edition: The standard for enterprise java. URL: https://www.oracle.com/java/technologies/
java-ee-glance.html. accessed on 17 April 2023.
[33] Pandas, 2023. Pandas. URL: https://pandas.pydata.org/. accessed on 17 April 2023.
[34] PostgreSQL, 2023a. Pgadmin4. URL: https://www.pgadmin.org/. accessed on 17 April 2023.
[35] PostgreSQL, 2023b. Postgresql. URL: https://www.postgresql.org/. accessed on 17 April 2023.
[36] Rasmussen, J., Langerman, H., 2019. Alzheimer’s disease – why we need early diagnosis. URL: https://www.ncbi.nlm.nih.gov/pmc/
articles/PMC6935598/.
[37] Sharp, E., Gatz, M., 2011. Relationship between education and dementia an updated systematic review. Alzheimer disease and associated
disorders 25, 289–304. doi:10.1097/WAD.0b013e318211c83c.
[38] Skirbekk, V., Bowen, C.E., Håberg, A., Jugessur, A., Engdahl, B., Bratsberg, B., Zotcheva, E., Selbæk, G., Kohler, H.P., Weiss, J., Harris,
J.R., Tom, S.E., Krokstad, S., Stern, Y., Strand, B.H., 2022. Marital histories and associations with later-life dementia and mild cognitive
impairment risk in the hunt4 70+ study in norway. Journal of Aging and Health URL: https://doi.org/10.1177/08982643221131926,
doi:10.1177/08982643221131926, arXiv:https://doi.org/10.1177/08982643221131926. pMID: 36321864.
[39] Society, A., 2023. Risk factors for dementia. URL: https://www.alzheimers.org.uk/about-dementia/risk-factors-dementia.
accessed on 15 April 2023.
[40] Sonarqube, 2023. Sonarqube. URL: https://docs.sonarqube.org/latest/. accessed on 17 April 2023.
[41] Stoleru, G.I., Iftene, A., 2022. Prediction of medical conditions using machine learning approaches: Alzheimer’s case study. Mathematics 10
(10): 1767.
[42] Studio, V., 2023. Visual studio code. URL: https://code.visualstudio.com/. accessed on 17 April 2023.
[43] Tanrıverdi, Y., 2023. Mmse. URL: https://play.google.com/store/apps/details?id=com.yasintanriverdi.mmse&hl=en_US.
accessed on 18 April 2023.
[44] TensorFlow, 2023. Tensorflow. URL: https://www.tensorflow.org/. accessed on 17 April 2023.
[45] UK, A.R., 2020. Differences in early-life education linked to dementia risk. URL: https://www.alzheimersresearchuk.org/
differences-in-early-life-education-linked-to-dementia-risk/. accessed on 18 April 2023.
[46] Villa, C., Lavitrano, M., Salvatore, E., Combi, R., 2020. Molecular and imaging biomarkers in alzheimer’s disease: A focus on recent insights.
Journal of Personalized Medicine 10.
[47] Wohlfahrt-Laymann, J., 2017. Cognidecline : tracking mobile interaction for cognitive assessment.