Detecting early symptoms of Alzheimer’s disease (AD) is critical for effective prevention and treatment. However, traditional detection methods are expensive and inefficient, presenting a major challenge in the pursuit of more cost-effective and efficient solutions. In this paper, we introduce a multimodal personal health graph aimed at improving the detection of early AD symptoms. Our approach leverages a cross-modal attention mechanism to capture diverse health characteristics by extracting both shared and unique information across modalities. Based on these features and their correlations, we construct personalized health graphs that reflect the diversity and individuality of patients. A flexible Graph Convolutional Network (GCN) is employed to derive diagnostic predictions. Experimental results on real clinical datasets demonstrate that our model achieves an accuracy of 87.22% on the Dem@Care dataset, and the personal health graph effectively distinguishes between groups by accurately capturing individual traits. We also verify that different modalities contribute differently to the diagnostic process. Therefore, our method exhibits strong potential for advancing AD diagnosis.