ClinPatch-AD: Token-Level Fusion of 3D MRI Patch Transformers and Clinical Scores for Subject-Wise Alzheimer’s Disease Screening

Alzheimer’s disease (AD) is a leading cause of dementia, and early subject-level screening is important for timely clinical support. However, automated MRI-based AD diagnosis remains difficult because many deep learning studies rely on 2D slices or non-subject-wise splits, which can overestimate performance, and clinical variables are often fused only at the final stage. This study aims to develop a more reliable multimodal framework that uses full 3D MRI volumes and clinical information under strict subject-wise evaluation.We propose ClinPatch-AD, a 3D patch-based Transformer that converts each MRI scan into volumetric patch tokens and embeds age, sex, and MMSE as a dedicated clinical token. The clinical token is inserted into the same Transformer sequence as the MRI tokens, allowing token-level fusion through self-attention. An auxiliary MMSE regression head is also used during training to guide the shared representation with cognitive severity information. Experiments on the OASIS-1 cross-sectional cohort with a stratified subject-wise split achieved an AUC of 0.951, accuracy of 0.898, precision of 0.857, recall of 0.818, and F1-score of 0.840 on the held-out test set. Ablation results further showed that both the clinical token and the auxiliary MMSE objective improved performance over the MRI-only Transformer baseline. These findings show that token-level multimodal fusion improves realistic patient-level AD screening and supports the use of clinical context during MRI-based diagnosis.