BioPhysio-guided hybrid transformer-UNet for early Alzheimer's disease staging and longitudinal progression prediction using multimodal ADNI data

Detection of Alzheimer’s disease (AD) has proven difficult; further complicating the issue is determining the progression of the condition, as there are very slight structural changes and considerable variability among imaging datasets collected from each patient. Current approaches exacerbate these issues by relying solely upon data, while neglecting much of the biological understanding available on neurodegenerative diseases. This reliance results in poor generalization and difficulty interpreting results. We developed a novel hybrid architecture, the BioPhysio-Guided Hybrid Transformer-UNet (BPG-HTU), to integrate biologically and physically informed prior knowledge about AD into a longitudinal multimodal predictive model. BPG-HTU performs two functions using the same input (clinical data and T1-weighted MRI): early-stage classification (CN vs. EMCI vs. LMCI vs. AD) and prediction of future cognitive decline. Architecturally, BPG-HTU includes three components: (i) a Swin-Transformer encoder that captures global context, (ii) a temporal cross-attention module that models longitudinal dynamics, and (iii) a U-Net-based decoder that localizes features for both prediction tasks. AD’s atrophy mechanisms are incorporated throughout via physics-based constraints. Using 1,248 longitudinal T1-weighted MRI scans from ADNI, BPG-HTU achieved an average classification accuracy of 87.4% ± 2.1%, a macro F1-score of 0.85 ± 0.03, and an area under the curve (AUC) of 0.91 ± 0.02 for multi-class staging, exceeding a 3D U-Net (81.2%) and recent temporal models including AT-Attn variants. Recall for early-stage MCI transitions was 82.6%. Ablation studies showed that biological and physical constraints together produced a 0.04 (4.0-point) increase in macro AUC. Grad-CAM visualizations of region-specific activations supported high interpretability, and the model’s computational efficiency supports its potential as a practical, clinically dependable diagnostic tool. Clinical trial registration This study is not a clinical trial. It is a retrospective, secondary analysis of existing, publicly available longitudinal neuroimaging and clinical data obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. No prospective interventional study was conducted by the authors, and clinical trial registration is therefore not applicable.