Non-Invasive Early Detection of Alzheimer’s Disease Using AI-Powered Analysis of Clinically Silent Biomarkers

By Zoha Ali, Pakistan

Impact Statement

Alzheimer’s affects 55+ million globally, with irreversible damage beginning decades before symptoms. Our AI model forecasts its onset 24 months in advance, non-invasively, via MRI and blood biomarkers. With 92.4% sensitivity, our system offers a scalable, explainable diagnostic benchmark.

​

Abstract:

Early, accurate detection of Alzheimer's Disease (AD) is a pivotal factor in mitigating cognitive decline. Traditional diagnostic methods are ineffective at detecting AD during its pre-symptomatic phase, thereby delaying critical intervention. This paper proposes a sophisticated AI-powered framework utilizing multi-modal neuroimaging data (MRI scans) and biochemical biomarkers (CSF Aβ42/tau ratios, proteomic blood markers) to identify Alzheimer’s in its early, clinically silent stages. By integrating these data types into a hybrid convolutional neural network (CNN) and Gradient Boosting Machine (GBM) ensemble model, we achieved a performance of AUC = 0.942, accuracy = 91.4%, and recall = 90.8%. This performance significantly outperforms baseline models, underscoring the framework's efficacy in the early detection of AD. Our approach identifies preclinical AD biomarkers up to 24 months before the manifestation of cognitive decline, offering a transformative tool for early intervention and clinical trial optimization.

Introduction:

Alzheimer’s Disease (AD) is a neurodegenerative disorder causing progressive cognitive decline. Early detection is crucial for effective treatment, but current diagnostic methods, such as cognitive assessments, PET scans, and cerebrospinal fluid (CSF) biomarkers are invasive, expensive, and typically identify AD only in its later stages. Thus, non-invasive and accurate diagnostic approaches are urgently needed. Artificial Intelligence (AI), particularly Convolutional Neural Networks (CNNs), has shown promise in revolutionizing medical diagnostics, especially neuroimaging. However, integrating multi-modal data, such as neuroimaging and biochemical biomarkers, is an emerging frontier. AI-driven models have outperformed traditional methods in AD detection, reducing false positives and negatives and achieving higher accuracy. For example, while earlier models using traditional machine learning methods had detection accuracy rates around 70-80%, recent AI models have reached over 90% in identifying early-stage AD with greater precision. This paper proposes an AI-based framework utilizing multi-modal data to enable early-stage AD detection, improving diagnosis and intervention outcomes.

Literature Review:

Over the last decade, AI has increasingly been applied to Alzheimer’s detection, yet current models remain limited by their late-stage focus, dependency on invasive biomarkers, and lack of interpretability. Early machine learning efforts, using SVMs and random forests, showed moderate success but failed to generalize across diverse populations (Klöppel et al., 2008; Westman et al., 2012). The rise of deep learning, particularly CNN and CNN-LSTM architectures, brought substantial improvements in accuracy (Korolev et al., 2017; Wang et al., 2020), yet many models still function as opaque “black boxes” and often overlook the preclinical stages of neurodegeneration.

While attention-based networks and interpretability tools (e.g., Grad-CAM, SHAP) have recently emerged (Qiu et al., 2022), their integration into clinical settings remains minimal. For example, in one early-stage case involving a 63-year-old asymptomatic patient, our system identified subtle hippocampal atrophy undetected by conventional methods, later confirmed via PET.

Our model advances the field by offering a non-invasive, interpretable deep learning framework capable of detecting preclinical AD from MRI scans with both precision and transparency, addressing the major limitations of existing approaches and enabling intervention far earlier than current standards allow.

Methodology:​ 

Data
We utilized two of the most comprehensive publicly available datasets in AD research: the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and the OASIS-3 dataset. The ADNI dataset comprises over 10,000 neuroimaging and biomarker samples, including T1-weighted MRI scans, CSF biomarkers (Aβ42, tau), and longitudinal cognitive assessments. These samples span across cognitively normal (CN), mild cognitive impairment (MCI), and Alzheimer’s stages, providing a diverse dataset for training our model. The OASIS-3 dataset, which includes data on MCI and AD patients, was employed for model validation and external testing.

​

Data preprocessing included normalization of MRI scans and extraction of key features from proteomic blood biomarkers, utilizing statistical techniques such as Z-score transformation and feature scaling. Synthetic Minority Over-sampling Technique (SMOTE) was applied to address class imbalance, particularly for the underrepresented MCI class.

​

Model Architecture

The hybrid ensemble model designed for this study integrates two powerful machine learning techniques: DenseNet121 for convolutional feature extraction from neuroimaging data and Gradient Boosting Machines (GBM) for handling structured biomarker data. The DenseNet121 CNN architecture, chosen for its efficiency in learning hierarchical representations, was fine-tuned to detect subtle brain atrophy patterns indicative of early AD. Concurrently, GBM was employed to model the complex relationships between biochemical markers, such as Aβ42/tau ratios and proteomic profiles, which are pivotal in AD pathophysiology.

To enhance feature extraction from latent biomarkers, Recursive Feature Elimination (RFE) and Principal Component Analysis (PCA) were incorporated. Dropout layers within the CNN were utilized to prevent overfitting and promote the generalizability of the model to unseen data.

​

Model Training and Evaluation

The model was trained on a dataset split of 70% for training and 30% for validation. Additionally, we employed 10-fold cross-validation to further ensure robustness. The model’s performance was evaluated using a variety of metrics: accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC). Comparisons were made with baseline machine learning models, including logistic regression, support vector machines (SVM), and single-modality models that analyzed only neuroimaging data or biomarkers.​​

Results:

The hybrid CNN-GBM model demonstrated remarkable diagnostic performance. It achieved an AUC of 0.942, accuracy of 91.4%, precision of 89.6%, and recall of 90.8%, surpassing the performance of individual modality models (MRI-only: AUC = 0.869, Blood-only: AUC = 0.791). The model’s sensitivity in identifying MCI, a critical preclinical stage of AD, was particularly notable, with a sensitivity of 93.8% in detecting MCI conversion from CN up to 24 months before clinical decline. These results highlight the ability of the model to detect Alzheimer’s before overt cognitive decline, offering a crucial window for early intervention. The following statistics illustrate the comparative performance:

​​

​​​

​

​

​

​

​​

​

​​​​​​

​​

In terms of demographic subgroup analysis, the model maintained consistent performance across different age groups, genders, and ethnicities, suggesting high generalizability.

​

Discussion

This study illustrates that the integration of multi-modal neuroimaging and biochemical biomarkers enhances the detection of Alzheimer’s Disease during its earliest stages, well before cognitive symptoms manifest. The superior performance of the hybrid model, as compared to single-modality models, underscores the importance of combining imaging and biomarker data for a more holistic understanding of AD’s pathophysiology. The model’s high sensitivity and precision in detecting MCI, in particular, suggests its applicability as a screening tool for preventive healthcare initiatives, enabling early intervention and personalized treatment plans. Additionally, the innovative use of dropout within the convolutional layers aids in reducing overfitting, making the model more robust to unseen data.

Conclusion

This AI-powered framework, integrating neuroimaging and biomarker data, provides a novel approach for the early, non-invasive detection of Alzheimer’s Disease. Achieving superior diagnostic performance, the hybrid CNN-GBM model offers significant promise in improving early detection and enhancing clinical outcomes through timely interventions. Future work will focus on expanding the model’s scope by incorporating genetic markers, EEG data, and longitudinal studies to refine its predictive capabilities and further personalize AD risk stratification

​

Appendices

​

​