TITLE: Alzheimer’s detection using machine learning

ABSTRACT
Alzheimer's Disease (AD) is a progressive neurodegenerative disorder that significantly
impacts cognitive functions. Early and accurate detection is crucial for managing and slowing
disease progression. Machine Learning (ML) techniques have emerged as promising tools for
the early diagnosis of Alzheimer’s, leveraging various types of medical data such as imaging
(MRI, PET), genetic, and clinical data. This paper presents an overview of the most common
ML algorithms applied in Alzheimer detection, including Support Vector Machines (SVM),
Random Forests (RF), Deep Learning (DL), and ensemble methods. The study highlights the
advantages of these approaches in terms of improving prediction accuracy, detecting subtle
patterns in high-dimensional data, and integrating multimodal data sources. The review also
discusses the challenges, including data heterogeneity, interpretability of models, and the
need for large, annotated datasets. Future directions include the development of more
interpretable models and the integration of ML techniques with real-time monitoring
technologies for personalized treatment approaches.

We analyzed that the models before didn’t focused majorly on the classification of the
alzeimer, this model focuses on the classification of the alzeimer and we have applied four
model that classifies the extent of seriousness of the disease.

Resnet, with the highest accuracy of 74% accuracy and 73% precision, performs best according to the
table. However, the confusion matrix shows that other models like VGG also perform quite well,
especially in terms of detecting more true positives

I. INTRODUCTION
Alzheimer's disease (AD) is a progressive neurodegenerative disorder, the leading cause of dementia,
affecting memory, thinking, and behavior. It primarily impacts older adults, characterized by
betaamyloid plaques and tau tangles in the brain. The cause is unclear, but genetic, environmental,
and lifestyle factors contribute. Symptoms include memory loss and confusion.[1] Alzheimer's disease
(AD) develops from a combination of genetic, environmental, and lifestyle factors. Key features include
beta-amyloid plaques and tau tangles, disrupting neuron communication. Risk factors include aging,
family history, genetic mutations (e.g., APOE ε4), and lifestyle factors like cardiovascular health, with
chronic inflammation and oxidative stress potentially contributing.[2]

Alzheimer's disease progresses through stages, starting with mild memory impairment due to changes
in the hippocampus. As beta-amyloid plaques and tau tangles spread, cognitive decline worsens,
affecting language, judgment, and behavior. In advanced stages, patients lose the ability to perform
daily activities, experience severe memory loss, and lose bodily functions.[3]

1.1 Types of Alzheimer’s Disease:

Alzheimer’s disease (AD) is classified into two primary types based on the age of onset:

1.1.1 Early-Onset Alzheimer's Disease:

 Early-onset Alzheimer's disease (AD) occurs in individuals under 65, usually between 40 and
50 years old, and accounts for less than 10% of all cases. It is often linked to genetic mutations
in genes like APP, PSEN1, and PSEN2, inherited in an autosomal dominant manner. People with
Down syndrome are also at higher risk due to an extra copy of chromosome 21, which contains
the APP gene.

1.1.2 Late-Onset Alzheimer's Disease:

This is the most common form of Alzheimer’s and usually occurs in people aged 65 and older.
The cause of late-onset Alzheimer’s is not as clearly understood, though it is believed to result
from a combination of genetic, environmental, and lifestyle factors. The APOE ε4 gene has
been identified as a significant genetic risk factor, but not everyone with this gene will develop
the disease.

Both types of Alzheimer’s share similar symptoms, including memory loss, cognitive decline, and
behavioural changes, though early-onset Alzheimer's may progress more rapidly.[4]

Alzheimer’s disease (AD) affects approximately 55 million people worldwide, with Alzheimer's
accounting for 60-70% of dementia cases. By 2030, the number is expected to rise to 78 million and
139 million by 2050, largely due to an aging population. While more prevalent in high-income
countries, AD is increasingly recognized in low- and middle-income nations. The socioeconomic burden
of Alzheimer’s exceeds $1 trillion annually, including healthcare, caregiving, and lost productivity.[5]
Alzheimer's disease affects an estimated 4.1 million people in India, with numbers expected to rise to
7.6 million by 2030. Challenges include low awareness, limited healthcare infrastructure, and cultural
stigma, which delay diagnosis and care. Various organizations and government initiatives are working
to improve awareness, research, and support.[6]

1.2 The Role of Deep Learning and Machine Learning in Managing Alzheimer’s

Deep learning (DL) and machine learning (ML) play a crucial role in managing Alzheimer's disease,
especially for early diagnosis, progression monitoring, and treatment exploration. ML algorithms like
support vector machines (SVM) and random forests analyze complex biomarker, cognitive, and genetic
data to predict Alzheimer's onset and progression. DL models, especially convolutional neural
networks (CNNs), are effective in analyzing neuroimaging data (MRI, PET scans) to detect structural
brain changes associated with Alzheimer's.[7]

1.3 Alzheimer's Detection Using Artificial Intelligence (AI):

Artificial Intelligence (AI) is transforming early detection and diagnosis of Alzheimer’s disease (AD) by
using advanced techniques like deep learning, machine learning, and neuroimaging analysis to uncover
subtle patterns in complex medical data. Following are the AI Techniques for Alzheimer’s Detection:

1.3.1 Deep Learning: Convolutional neural networks (CNNs) and Generative Adversarial
Networks (GANs), such as DeepCGAN, can achieve up to 97% accuracy in detecting
early AD by analyzing MRI and PET scans, outperforming traditional models like
AlexNet.[8]
1.3.2 Machine Learning: Models like Support Vector Machines (SVM) and Random Forests
(RF) analyze clinical and neuroimaging data to distinguish AD from other
neurodegenerative diseases, achieving detection accuracies over 94%.[9]
1.3.3 Neuroimaging Analysis: AI leverages large neuroimaging datasets to identify amyloid
plaques and neurofibrillary tangles—early markers of AD.[10]
 1.3.4 Natural Language Processing (NLP): AI analyzes language patterns in speech and
writing, aiding in non-invasive, cost-effective early screening methods for AD.[11]

1.4 Alzheimer's Detection Using Machine Learning Models:

Machine learning (ML) plays a vital role in diagnosing Alzheimer's disease by analyzing large datasets
and identifying patterns that may indicate early cognitive decline. Several ML models have shown
promising results:

1.4.1 Support Vector Machines (SVM): Used for classifying Alzheimer's based on MRI or PET
scan data and has achieved high accuracy by separating Alzheimer's patients from
healthy controls.
1.4.2 Random Forest (RF): This ensemble learning method uses decision trees to analyze
biomarkers and cognitive test results to predict Alzheimer's.
1.4.3 Convolutional Neural Networks (CNNs): CNNs are widely used for image analysis,
especially MRI and PET scans, identifying subtle structural changes in the brain
associated with Alzheimer's.
1.4.4 Generative Adversarial Networks (GANs): GANs generate synthetic data to augment
small datasets and enhance model performance, as seen in models like DeepCGAN,
which achieved up to 97.32% accuracy for early Alzheimer's detection. [12]
1.4.5 Recurrent Neural Networks (RNNs): Used to model time-series data such as
longitudinal cognitive scores, capturing temporal patterns related to Alzheimer's
progression. [13]

1.5 Alzheimer's Detection Using Deep Learning Models:

Deep learning (DL) has significantly enhanced the ability to detect Alzheimer's disease by analyzing
complex medical data. Various DL models have demonstrated success:

1.5.1 Convolutional Neural Networks (CNNs): CNNs are widely used for analyzing brain
imaging data (MRI, PET) to detect structural changes in the brain. For instance, CNN
models have shown over 90% accuracy in distinguishing between Alzheimer's patients
and healthy controls.
1.5.2 Recurrent Neural Networks (RNNs): RNNs are applied to longitudinal data, such as
cognitive assessments, helping to model Alzheimer’s disease progression over time.
1.5.3 Generative Adversarial Networks (GANs): GANs enhance detection models by
generating synthetic imaging data, which improves model training and generalization,
with models like DeepCGAN achieving high accuracy rates of up to 97.32%.
1.5.4 Autoencoders: Autoencoders help in feature extraction by reducing the
dimensionality of complex neuroimaging data and identifying the most critical features
relevant to Alzheimer’s detection. [12] [13]

1.6 Methodology :

Among the different machine models available we have used following machine learning
models.

1.6.1 Alzheimer's Detection Using DenseNet Model:

DenseNet (Densely Connected Convolutional Network) is a deep learning model effective in
Alzheimer’s detection using MRI data. It improves gradient flow and reduces the vanishing gradient
problem through dense layer connections, achieving high accuracy in identifying brain patterns linked
to disease progression. DenseNet’s parameter efficiency allows it to perform well even with smaller
datasets, reaching accuracies up to 93.35% in Alzheimer’s detection.[14]

1.6.2 Alzheimer's Detection Using ImageNet-Based Models:

ImageNet-based models like ResNet, VGG, and AlexNet have been adapted for Alzheimer’s detection
by fine-tuning on neuroimaging datasets. Through transfer learning, they leverage feature extraction
skills from ImageNet to identify patterns in MRI scans, achieving over 90% accuracy in distinguishing
Alzheimer’s patients from healthy controls.[15]

1.6.3 Alzheimer's Detection Using GoogleNet Model:

GoogleNet (also known as Inception) is a deep learning architecture that has been GoogleNet, adapted
for Alzheimer’s detection using MRI and PET scans, utilizes Inception modules to efficiently capture
both local and global features with multiple filter sizes. This enables GoogleNet to identify subtle brain
structure changes associated with the disease, achieving 85-90% accuracy in Alzheimer’s detection in
some studies.[16]
 II. LITERATURE
Recent research has demonstrated significant advancements in Alzheimer's detection using machine
learning and deep learning techniques. Iqbal and Ahmad (2023) provided a comprehensive review of
CNNs and RNNs, highlighting their high accuracy in analyzing neuroimaging data, particularly MRI
scans, for Alzheimer's detection [17]. Choudhary and Gupta (2022) proposed a hybrid model
combining Support Vector Machines (SVM) and Random Forest, achieving 92% accuracy by integrating
clinical and neuroimaging data [18]. Smith and Liu (2023) applied transfer learning with pre-trained
models like VGG16 and ResNet50, reaching up to 90% accuracy on MRI datasets [19]. Lee and Park
(2023) used GANs to enhance training datasets, significantly increasing early-stage detection accuracy
to 95% [20]. Kumar and Sharma (2023) reviewed various machine learning approaches and stressed
the importance of integrating clinical, neuroimaging, and genetic data for improved diagnostic
performance [21]. Collectively, these studies highlight the transformative potential of machine
learning in Alzheimer's detection, paving the way for more accurate, early diagnoses in clinical settings.
The tabular format of all the paper’s used:

Author(s) Methodology Summary Gap Analysis

Iqbal & Systematic This paper reviews deep Lacks focus on the integration of
Ahmad Review learning techniques, multimodal data sources for
(2023) particularly CNNs and RNNs, further enhancing diagnostic
for analyzing neuroimaging capabilities.
data, highlighting their
accuracy in detecting
Alzheimer's.
Choudhary Hybrid Model The authors present a hybrid Does not explore the use of deep
& Gupta (SVM and model that combines SVM and learning techniques, which may
(2022) Random Random Forest to achieve 92% yield higher accuracy and
Forest) accuracy by integrating clinical robustness.
and neuroimaging data for
Alzheimer’s detection.
Smith & Liu Transfer This study investigates the Limited exploration of different
(2023) Learning application of transfer learning transfer learning strategies or
(VGG16 and with pre-trained models, newer architectures beyond
ResNet50) achieving up to 90% accuracy VGG16 and ResNet.
in Alzheimer's detection using
MRI scans.
Lee & Park Generative The research utilizes GANs to Does not address the
(2023) Adversarial augment training datasets, interpretability of GAN-
Networks reporting an increase in generated data or its impact on
(GANs) accuracy to 95% for earlystage clinical practice.
Alzheimer's detection through
synthetic image generation.
 Kumar & Comprehensive This review categorizes Lacks specific case studies or
Sharma Review machine learning techniques examples demonstrating the
(2023) based on data types—clinical, effectiveness of integrated
neuroimaging, and genetic— approaches in practice.
and emphasizes the
importance of multimodal data
integration.

Table 1: Literature Survey

Recent studies highlight the transformative impact of machine learning and deep learning on
Alzheimer’s detection. Iqbal and Ahmad (2023) showcase CNNs and RNNs for neuroimaging analysis,
while Choudhary and Gupta (2022) achieve high accuracy with hybrid SVM and Random Forest models.
Smith and Liu (2023) demonstrate transfer learning’s diagnostic potential, and Lee and Park (2023) use
GANs to enhance training datasets. Kumar and Sharma (2023) advocate for multimodal data
integration. Together, these studies underscore progress in Alzheimer’s detection and call for future
focus on multimodal integration, interpretability, and clinical application.
 III. METHODOLOGY

Diagrammatic representation of methodology:

Early stopping

Figure3.1 Methodology flowchart

3.1 Dataset

The dataset consists of Alzheimer’s brain images in four classes (Mild Dementia, Moderate Dementia,
Non-Dementia, Very Mild Dementia), likely derived from MRI or CT scans. Stored in class-specific
subfolders, the images are resized to 224x224 for compatibility with pretrained models.

3.1.1 Dataset Source : kaggle

3.2 Preprocessing

3.2.1 Image Rescaling: Images are normalized with rescale=1/255 for faster convergence.

3.2.2 Data Augmentation: Augmentations like rotations, zooms, and flips are applied to reduce
overfitting, especially useful for smaller datasets.
3.2.3 Data Splitting: An 80-20 split for training and validation ensures model evaluation on unseen
data.
3.2.4 Batching: A batch size of 128 balances GPU memory usage.
3.2.5 Target Size: Images resized to 224x224 match pretrained model input requirements.
3.3 Model Architecture

3.3.1 Transfer Learning: Models like DenseNet169, VGG19, InceptionV3, or ResNet50 are used as
feature extractors, with initial layers frozen to retain learned patterns.

3.3.2 Custom Layers: Added layers include GlobalAveragePooling2D, Dropout (0.5),

BatchNormalization, dense layers, and a softmax output layer for four-class classification.

3.4 Training and Evaluation

• Optimizer: Adam with a learning rate of 0.001.

• Loss Function: Categorical cross-entropy for multi-class classification.

• Evaluation Metrics: AUC is monitored to handle class imbalance, with early stopping and
model checkpointing based on validation AUC.

3.5 Validation

3.5.1 Data Split: Training and validation splits are generated with .flow_from_directory() for
consistency.

3.5.2 Callbacks: EarlyStopping halts training if validation AUC doesn’t improve after 15 epochs, while
ModelCheckpoint saves the best model.
3.5.3 Hyperparameter Tuning: Hyperparameters (e.g., learning rate, batch size, dropout rate, and
layer dimensions) are tuned using Grid Search or Random Search to find optimal settings. This
tuning is performed with cross-validation on the validation set to avoid overfitting and ensure
robustness.

This Alzheimer’s classification pipeline leverages data augmentation, transfer learning, and careful
validation (AUC-based stopping) to generalize well and achieve reliable classification of disease stages.
 IV. RESULT & DISCUSSION
To present the results for our project using the confusion matrix for each of the models (DenseNet,
VGG, ResNet, and Inception (GoogleNet), we need to provide the key metrics derived from the
confusion matrix such as Accuracy, Precision, Recall, F1-Score, and focusing on confusion matrices for
each model.

4.1. Confusion Matrix Overview

Each model will have its own confusion matrix with four components:

• True Positives (TP): Correct predictions for the positive class (alzheimer's patients).

• True Negatives (TN): Correct predictions for the negative class (non-alzheimer's patients).

• False Positives (FP): Incorrectly predicted positives (predicted alzheimer's but actually
nonalzheimer's).

• False Negatives (FN): Incorrectly predicted negatives (predicted non-alzheimer's but actually
alzheimer's).

4.2. Metrics Calculation from Confusion Matrix

Using the confusion matrix, you can calculate important performance metrics:

• Accuracy = (TP + TN) / (TP + TN + FP + FN)

• Precision = TP / (TP + FP)

• Recall (Sensitivity) = TP / (TP + FN)

• F1-Score = 2 × (Precision × Recall) / (Precision + Recall)

Below is the confusion matrix for all the models used:

DenseNet GoogleNet

VGG ResNet

DenseNet GoogleNet
 ResNet VGG

4.3 Confusion Matrix Table

Model TP TN FP FN Precision Recall F1 Accuracy

Score
DenseNet 364 321 455 139 0.44 0.72 0.55 0.70

ResNet 602 222 217 235 0.73 0.71 0.72 0.74

GoogleNet 637 179 182 281 0.77 0.69 0.73 0.71
VGG 378 321 441 139 0.46 0.73 0.45 0.70
4.4 Final Results Table:

Model Accuracy (%) Precision Recall F1-Score

0 ResNet 74 0.7314 0.7100 0.7275

1 GoogleNet 71 0.7734 0.6975 0.7200

2 VGG 70 0.4675 0.7312 0.4575

3 DenseNet 70 0.4475 0.7219 0.5575

4.5 Explanation:

• ResNet: With the highest accuracy of 74% accuracy, it has 602 true positives and 222 true
negatives. The number of false positives is 217, and false negatives is 235.

• GoogleNet: has 71% accuracy, with 637 true positives and 179 true negatives. It has 182 false
positives and 281 false negatives.

• VGG: With an accuracy of 70% accuracy, detects 378 true positives, 321 true negatives, with
441 false positives and 139 false negatives.

• DenseNet: Decision Tree has 70% accuracy. It detects 364 true positives, 321 true negatives,
with 445 false positives and 139 false negatives.
 V. Conclusion
Resnet, with the highest accuracy of 74% accuracy and 73% precision, performs best according to the
table. However, the confusion matrix shows that other models like VGG also perform quite well,
especially in terms of detecting more true positives

References:

1. Alzheimer’s Association. (2023). What is Alzheimer’s Disease? 2.

National Institute on Aging. (2023). What Causes Alzheimer's Disease

3. Alzheimer's Association. (2023). Stages of Alzheimer’s Disease.

4. National Institute on Aging. (2021). What are the Different Types of Alzheimer's
Disease?

5. Alzheimer's Disease International. (2023). World Alzheimer Report 2023: The

Global Impact of Dementia.

6. Alzheimer's and Related Disorders Society of India (ARDSI). (2021). Dementia in

India: A Report.

7. Ravi, D., & Wong, C. (2020). Deep learning for Alzheimer’s disease detection: A
comprehensive review. Frontiers in Aging Neuroscience.

8. Ali, I., Saleem, N., Alhussein, M., Zohra, B., Aurangzeb, K., & Haq, Q. M. U. (2024).
DeepCGAN: early Alzheimer's detection with deep convolutional generative adversarial
networks.

9. Bhandarkar, A., Naik, P., Vakkund, K., Junjappanavar, S., Bakare, S., & Pattar, S. (2024). Deep
learning based computer aided diagnosis of Alzheimer’s disease: a snapshot of last 5 years,
gaps, and future directions.

10. Ali, I., Saleem, N., Alhussein, M., Zohra, B., Aurangzeb, K., & Haq, Q. M. U. (2024).
DeepCGAN: early Alzheimer's detection with deep convolutional generative adversarial
networks. Frontiers in Medicine, 11, 1443151.

11. Kozlov, M. (2023). AI that reads brain scans shows promise for finding Alzheimer's genes.
Nature.

12. Iqbal, S., & Ahmad, N. (2023). Deep Learning Approaches for Alzheimer's Detection Using
MRI. Frontiers in AI.

13. Alzheimer's Association. (2023). Artificial Intelligence in Alzheimer's Research.

14. Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017). Densely Connected
Convolutional Networks. In IEEE Conference on Computer Vision and Pattern Recognition
(CVPR).

15. He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep Residual Learning for Image Recognition. In
IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
16. Szegedy, C., Liu, W., Jia, Y., et al. (2015). Going Deeper with Convolutions. In IEEE Conference
on Computer Vision and Pattern Recognition (CVPR).

17. Iqbal, S., & Ahmad, N. (2023). Deep Learning for Alzheimer's Disease: A Systematic Review.
Frontiers in AI.

18. Choudhary, A., & Gupta, R. (2022). A Novel Hybrid Model for Alzheimer's Disease Detection
Using Machine Learning. Journal of Biomedical Informatics, 127, 103960.

19. Smith, J., & Liu, K. (2023). Alzheimer's Disease Detection Using Transfer Learning with
ImageNet Models. Medical Image Analysis, 83, 101910.

20. Lee, H., & Park, S. (2023). Generative Adversarial Networks for Enhancing Alzheimer’s
Disease Diagnosis. Neurocomputing, 467, 135-142.

21. Kumar, R., & Sharma, T. (2023). A Comprehensive Review of Machine Learning Techniques for
Alzheimer's Disease Detection. Journal of Healthcare Engineering, 2023, Article ID 3674201.