Abstract
OBJECTIVES: To develop and validate a hybrid deep learning model to enhance diagnostic and predictive accuracy of Alzheimer's disease (AD) using readily available clinical data. METHODS: A triple-architecture joint model was constructed, integrating a Kernel Extreme Learning Machine (KELM), a Long Short-Term Memory (LSTM) network, and a Transformer. This framework was designed to capture nonlinear associations, temporal dynamics, and global feature dependencies. The model was trained and validated on 2,149 subjects from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database and externally tested on an independent cohort of 1,012 subjects. RESULTS: In internal validation, the model achieved state-of-the-art performance with 95.42% accuracy, 95.63% recall, and an area under the curve (AUC) of 0.981. It also demonstrated strong generalizability in the external cohort, achieving 93.81% accuracy and an AUC of 97.25%. In a longitudinal sub-analysis, the model accurately predicted the 3-year conversion from mild cognitive impairment to AD with 92.78% accuracy. Ablation analyses confirmed the essential contribution of each model component. CONCLUSIONS: The proposed KELM-LSTM-Transformer model provides a powerful and robust framework for AD prediction. Its high accuracy and strong generalizability suggest potential as an effective and accessible tool for early risk stratification, supporting timely clinical interventions.