Neurocapsule : Alzheimer Risk Predictor and Digital Memory Vault

Neurocapsule : Alzheimer Risk Predictor and Digital Memory Vault

Rucha Gharve

Articial Intelligence and Data Science A.C.Patil College of Engineering Kharghar, India

Pratik Surwade

Articial Intelligence and Data Science A.C.Patil College of Engineering Kharghar, India

Pareenita Shirsath

Articial Intelligence and Data Science A.C.Patil College of Engineering Kharghar, India

Shejal Ambre

Articial Intelligence and Data Science A.C.Patil College of Engineering Kharghar, India

Suresh Pingale

Articial Intelligence and Data Science A.C.Patil College of Engineering Kharghar, India

AbstractAlzheimers disease is a progressive neurodegen- erative disorder that signicantly impacts memory, cognition,

and daily functioning. Early diagnosis plays a crucial role in slowing disease progression and improving patient care. This paper presents NeuroCapsule, an intelligent system designed for early detection of Alzheimers disease using brain MRI images combined with articial intelligence techniques. The proposed approach employs convolutional neural networks (CNNs) for automated feature extraction and classication of MRI scans into multiple cognitive stages, including Non-Demented, Very Mild Dementia, Mild Dementia, and Moderate Dementia. To address class imbalance and computational constraints, balanced sampling strategies and image downsampling techniques are applied. The system is evaluated on the OASIS MRI dataset, demonstrating high classication accuracy and efcient perfor- mance across both binary and multiclass tasks. Unlike traditional diagnostic models, NeuroCapsule extends beyond prediction by integrating a secure Digital Memory Vault that enables users to store reminders, notes, and personal records, supporting continu- ous cognitive management. Experimental results indicate that the proposed system achieves a favorable trade-off between accuracy and computational efciency, making it suitable for real-world deployment. The integration of predictive analytics with memory assistance enhances both clinical relevance and patient-centered care, positioning NeuroCapsule as a comprehensive solution for Alzheimers disease detection and management.

Index TermsAlzheimers Disease, Machine Learning, Deep Learning, MRI Analysis, Hybrid Classication Model, Digital

Memory Vault, Healthcare Articial Intelligence

1. Introduction

   Alzheimers Disease (AD) is one of the most prevalent neurodegenerative disorders, primarily affecting the elderly population and leading to progressive deterioration of mem- ory, cognitive abilities, and behavioural functions. With the global increase in life expectancy, the number of individuals

   Identify applicable funding agency here. If none, delete this.

   affected by Alzheimers Disease is rising steadily, resulting in signicant challenges for healthcare systems, caregivers, and families. Early prediction of Alzheimers risk plays a crucial role in slowing disease progression, enabling timely intervention, and improving the overall quality of life of patients.

   In recent years, articial intelligencebased approaches have been widely explored for Alzheimers detection using medical imaging modalities such as Magnetic Resonance Imaging (MRI), along with structured clinical and cognitive assessment data. Various machine learning and deep learning algorithms, including support vector machines, ensemble classiers, and convolutional neural networks, have demonstrated promising results in identifying Alzheimers-related patterns. Despite these advancements, existing studies often focus on individual algorithms or highly complex deep learning models, with limited emphasis on practical deployment, computational ef- ciency, and system-level integration.

   Moreover, most Alzheimers prediction systems are de- signed solely for diagnostic or risk assessment purposes and do not address the ongoing memory-related challenges faced by individuals at risk of the disease. Digital memory assistance tools, while useful for managing reminders and notes, typically operate independently of predictive intelligence and clinical insights. This separation highlights the need for integrated solutions that combine early risk prediction with continuous cognitive support.

   To address these limitations, this paper proposes Neuro- Capsule, an AI-based framework that integrates Alzheimers risk prediction with a Digital Memory Vault. Representative machine learning and deep learning algorithms reported in existing literature are experimentally evaluated to analyse their predictive performance and guide the selection of an optimal

   model. The selected model is then implemented as the core prediction engine of the NeuroCapsule system. In addition, the Digital Memory Vault provides a secure and user-friendly platform for managing daily reminders, notes, and multimedia memories, thereby extending support beyond prediction alone. By combining validated prediction techniques with real-world memory assistance, NeuroCapsule aims to provide a compre- hensive and practical solution for Alzheimers risk assessment and cognitive support.

2. Related Work

   Alzheimers Disease (AD) is a progressive neurodegen- erative disorder characterised by structural and functional changes in the brain. In recent years, machine learning and deep learning techniques have been extensively explored for early diagnosis of AD using neuroimaging data, particularly Magnetic Resonance Imaging (MRI), due to its non-invasive nature and high spatial resolution.

   Traditional machine learning approaches have employed handcrafted features extracted from MRI scans combined with classiers such as Support Vector Machines (SVM), Random Forest (RF), and gradient boosting methods. Shukla et al. demonstrated the effectiveness of ensemble-based classiers including Random Forest and XGBoost for Alzheimers dis- ease detection using MRI-derived features, reporting improved classication performance compared to single classiers. Such approaches highlight the robustness of ensemble methods but rely heavily on feature engineering and preprocessing quality. With the advancement of deep learning, Convolutional Neu- ral Networks (CNNs) have become the dominant approach for MRI-based Alzheimers disease classication. CNN-based models automatically learn hierarchical spatial features di- rectly from MRI images, reducing dependency on manual feature extraction. Foroughipoor et al. analysed multiple deep learning architectures applied to MRI scans and concluded that CNN-based models signicantly outperform traditional machine learning techniques in multiclass Alzheimers clas-

   sication tasks.

   Transfer learning has further enhanced deep learning perfor- mance in medical imaging applications, especially when train- ing data is limited. Pretrained architectures such as VGG16, ResNet, and MobileNet have been widely adopted as feature extractors for Alzheimers disease detection. Givian et al. demonstrated that transfer learning models ne-tuned on MRI datasets achieve higher accuracy and faster convergence com- pared to CNNs trained from scratch. Lightweight architectures such as MobileNet are particularly suitable for deployment in resource-constrained environments due to their reduced computational complexity.

   Several studies have also explored hybrid approaches that combine deep feature extraction with traditional machine learning classiers. In such frameworks, deep neural networks are used to extract discriminative features from MRI images, which are subsequently classie using ensemble classiers such as Random Forest or XGBoost. These hybrid models leverage the representational power of deep learning while

   maintaining the interpretability and stability of classical ma- chine learning methods.

   Although existing studies report promising results, most focus primarily on improving classication accuracy without performing systematic comparative evaluation across multiple deep learning and hybrid models using a unied experimental setup. Furthermore, practical aspects such as model efciency, usability, and extensibility towards real-world clinical support systems are often overlooked.

   To address these limitations, the proposed work conducts a comprehensive comparative analysis of baseline CNN, transfer learning models, and hybrid deep learningmachine learning approaches for multiclass Alzheimers disease detection using MRI images from the Image OASIS dataset. .

3. Methodology

   This section describes the complete methodological frame- work adopted for Alzheimers disease detection using brain MRI images. The methodology includes dataset selection, data preprocessing and balancing, image downsampling, model development, training strategy, and evaluation metrics.

   1. Dataset Description

      The experiments were conducted using the OASIS-1 Alzheimers Disease dataset, obtained from Kaggle. The dataset comprises structural brain MRI scans of 461 subjects spanning different cognitive stages. Originally, the OASIS dataset consists of three-dimensional (3D) MRI volumes. In the Kaggle version, each 3D scan is decomposed into 256 two-dimensional (2D) axial slices along the z-axis.

      Based on domain relevance and prior neuroimaging studies, slices in the z-axis index range of 100 to 160 were selected, as these regions capture critical structural changes associated with Alzheimers disease, particularly cortical thinning and brain tissue shrinkage. Each selected 2D slice was treated as an independent sample due to the consistent presence of Alzheimers-related biomarkers across these slices.

      The dataset is categorized into four classes:

      • Non-Demented

      • Very Mild Dementia

      • Mild Dementia

      • Moderate Dementia

        A detailed analysis of the dataset revealed that the selected z-axis slices effectively represent brain tissue shrinkage, which is a key biomarker of Alzheimers disease. Since this shrinkage is consistently observable across these slices, each 2D MRI slice was treated as an independent sample.

        The complete dataset occupies approximately 1.3 GB of storage and contains nearly 86,400 MRI images. The class- wise distribution is as follows:

      • Non-Demented: 67,000 images

      • Very Mild Dementia: 13,700 images

      • Mild Dementia: 5,000 images

      • Moderate Dementia: 500 images

   2. Data Preprocessing and Class Balancing

      The dataset exhibits signicant class imbalance, which can negatively impact model generalization and bias predictions toward majority classes. To address this issue, class balancing was performed separately for binary and multiclass classica- tion tasks using random undersampling.

      1. Binary Classication:

         Non-Alzheimers: 488 randomly selected images from the Non-Demented class

         Alzheimers: 162 images each from the Very Mild, Mild, and Moderate Dementia classes

         This resulted in a balanced dataset with an approximate 50 percent class prior for each category.

      2. Multiclass Classication

      For multiclass classication, the Moderate Dementia class, containing the fewest samples, was used as the reference. An equal number of 488 images were randomly sampled from each class, ensuring uniform class representation. This resulted in a balanced dataset with 25 percent class prior per category.

   3. Image Preprocessing and Downsampling

      To reduce computational complexity and memory over- head, MRI images were converted to grayscale, normalized, and resized. Since the primary structural features relevant to Alzheimers disease are preserved at lower resolutions, down- sampling was performed without compromising diagnostic information.

      The images were processed at two different resolutions:

      1. 64 × 32 pixels (1/8 of original resolution)

      2. 128 × 64 pixels (1/4 of original resolution)

      This multi-resolution approach enabled the evaluation of model performance under varying computational constraints while maintaining diagnostic robustness.

   4. Model Architecture and Training Strategy

      The proposed Alzheimers disease detection system employs a Convolutional Neural Network (CNN) architecture due to its proven effectiveness in extracting spatial features from medical imaging data. CNNs are particularly suitable for MRI analysis as they can automatically learn hierarchical representations of structural brain patterns associated with neurodegeneration.

      The architecture consists of multiple convolutional layers followed by pooling and fully connected layers. Each convo- lutional layer applies a set of learnable lters to extract low- level features such as edges and textures in the initial layers, and higher-level anatomical patterns in deeper layers. Rectied Linear Unit (ReLU) activation functions are used to introduce non-linearity and improve convergence speed. Max-pooling layers are incorporated after selected convolutional layers to reduce spatial dimensions and computational complexity while retaining the most salient features.

      To mitigate overtting, dropout regularization is applied in the fully connected layers. The nal output layer uses a Softmax activation function for multiclass classication and a

      Sigmoid activation function for binary classication, enabling probabilistic interpretation of the predictions.

      The network is trained using the Adam optimizer, which provides adaptive learning rates for efcient convergence. Categorical cross-entropy is used as the loss function for multi- class classication, while binary cross-entropy is employed for binary classication tasks. The dataset is divided into training, validation, and testing sets using a standard split strategy to ensure unbiased performance evaluation.

      Fig. 1. Architecture

      Training is performed for a xed number of epochs with an empirically selected batch size to balance training stability and memory usage. Early stopping is incorporated to prevent over- tting by monitoring validation loss and terminating training when performance saturation is observed. This training strat- egy ensures robust learning while maintaining computational efciency and generalization capability.

   5. Comparative Models

      To validate the effectiveness of the proposed CNN-based approach, its performance is compared against established deep learning architectures, including VGG16, ResNet, and MobileNet. These models are selected due to their widespread adoption in medical image analysis and their varying com- putational complexities. Comparative evaluation highlights the trade-offs between model accuracy, inference time, and resource requirements, thereby justifying the architectural choices of the proposed system.

   6. Evaluation Metrics

   The performance of the models is assessed using standard classication metrics commonly adopted in medical imaging research. These include accuracy, precision, recall, F1-score, and confusion matrix analysis. For binary classication, addi- tional emphasis is placed on sensitivity (recall) to minimize false negatives, which is critical for early Azheimers detec- tion. These metrics provide a comprehensive evaluation of model reliability and clinical applicability.

4. Conclusion and Future work

This paper presented an AI-based system for the early detection of Alzheimers disease using brain MRI images. The

Fig. 2. Home Page

1. G. P. Shukla, S. Kumar, and S. K. Pandey, Diagnosis and detection of Alzheimers disease using learning algorithms, Journal of Emerging Trends in Computing, vol. 11, no. 2, pp. 3441, 2023.

2. S. Foroughipoor, K. Moradi, and H. Bolhasani, Alzheimers disease diagnosis by deep learning using MRI-based approaches, Neuroinfor- matics Research Journal, vol. 15, no. 3, pp. 112124, 2023.

3. B. S. Rao and M. Aparna, A review on Alzheimers disease through analysis of MRI images using deep learning techniques, International Journal of Advanced Computer Science, vol. 14, no. 6, pp. 8998, 2023.

4. E. M. Mohammed, A. M. Fakhrudeen, and O. Y. Alani, Detection of Alzheimers disease using deep learning models: A systematic literature review, Informatics in Medicine Unlocked, vol. 50, p. 101551, 2024, doi: 10.1016/j.imu.2024.101551.

5. M. Pahar, F. Tao, B. Mirheidari, N. Pevy, and R. Bright, CognoSpeak: Automatic, remote assessment of early cognitive decline in real-world conversational speech, IEEE Transactions on Biomedical Engineering, vol. 72, no. 4, pp. 10211032, 2025.

6. M. Shahin, B. Ahmed, and J. Epps, Zero-shot cognitive impairment detection from speech using AudioLLM, in Proceedings of the IEEE International Conference on Affective Computing, 2025, pp. 210217.

Fig. 3. Vault Page

proposed approach employed convolutional neural networks to automatically extract discriminative features and classify subjects into different stages of cognitive impairment. Ex- perimental results demonstrated that the model achieved reli- able performance for both binary and multiclass classication tasks while maintaining computational efciency. The use of balanced datasets and optimized image resolutions further enhanced diagnostic accuracy and robustness.

In addition to disease prediction, the NeuroCapsule frame- work extends beyond conventional diagnostic systems by integrating a Digital Memory Vault to support daily cogni- tive management. This dual-function design improves clinical relevance by combining medical decision support with patient- centered memory assistance.

Future work will focus on incorporating multimodal data sources such as speech signals, cognitive test scores, and longitudinal clinical records to improve prediction reliability. The integration of advanced explainable AI techniques will further enhance model interpretability for clinical adoption. Additionally, real-world validation through larger and more diverse datasets, as well as deployment in a cloud-based healthcare environment, will be explored to improve scalability and accessibility.

References

1. H. Givian and J.-P. Calbimonte, Early diagnosis of Alzheimers disease and mild cognitive impairment using MRI analysis and machine learning algorithms, Discover Applied Sciences, vol. 7, no. 1, Art. no. 27, 2025, doi: 10.1007/s42452-024-06440-w.

2. M. Sai Teja, K. Thanuja, N. Mani Deep, P. Ravindra Reddy, and O. Likhith Kumar Reddy, Prediction and analysis of Alzheimers disease using deep learning algorithms, International Journal of Computer Applications, vol. 181, no. 45, pp. 16, 2023.