Public Health.

Area of Science:

• Artificial Intelligence
• Deep Learning
• Neuroscience

Background:

• Existing AI for Alzheimer's Disease and Related Dementias (ADRD) prediction often focuses on specific subtypes and homogenous, older populations.
• This limits the applicability of AI tools to diverse demographics and other forms of dementia.
• This study proposes an AI-based deep learning framework for early ADRD detection in heterogeneous, midlife populations.

Purpose of the Study:

• To develop and evaluate an AI-based deep learning framework for early detection of ADRD risk.
• To assess the performance of Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) models in predicting ADRD risk.
• To apply the framework to a diverse cohort from the European Prevention of Alzheimer's Dementia (EPAD) and PREVENT Dementia Programme.

Main Methods:

• A harmonized cohort of 2796 individuals without dementia diagnosis was curated from EPAD and PREVENT datasets.
• Individuals were categorized into High, Medium, and Low risk groups based on ApoE4 allele presence and family history of AD.
• CNN and LSTM models were developed and optimized using 5-fold cross-validation.

Main Results:

• The harmonized cohort comprised 2796 individuals (mean age 62, range 40-89 years; 57.5% female; 95% Caucasian).
• CNN models achieved higher accuracy (7% points higher) and F1-score than LSTM models.
• CNN models demonstrated superior mean Area Under the Receiver Operating Characteristic Curve (AUROC) scores (97% vs. 94%) and lower validation loss (0.36 vs. 0.46).

Conclusions:

• The superior performance and low validation loss of the CNN model indicate its generalizability for ADRD risk prediction.
• The model is currently optimized for Alzheimer's Disease (AD) but Transfer Learning will be employed to predict other ADRD subtypes.
• Future work will explore multimodal features and explainability for enhanced CNN architecture.