Alzheimer’s Disease (AD) is a complex and progressive neurological disorder that presents differently among patients, making early diagnosis very difficult. Many diagnostic methods that rely on a single data source struggle to capture variations. Although multimodal strategies have improved accuracy, most still analyze patients individually and ignore the clinically useful relationships between them. In this paper, a patient-centered graph-based model designed for AD diagnosis and subtyping is introduced. Patient Similarity Network (PSN), is built by combining clinical records and MRI scans from the OASIS-1 dataset [1] along with synthetically generated data [2], where each patient is represented as a node, with edges reflecting similarity in clinical and imaging features. Mixed embedding approach and Principal Component Analysis (PCA) for feature extraction and Node2Vec for structural representation to represent both personal and relational information is used. It was observed, by using K-Means clustering on these hybrid embeddings, several patient subtypes were consistent with known clinical patterns. To perform the classification task, a supervised Graph Attention Network (GAT) is trained. The model achieved an accuracy of about 97.14%, which is better than several existing multimodal and unimodal techniques. Our framework offers a practical direction toward more patient-focused and relationship-aware approaches in Alzheimer’s disease studies.