Artificial Intelligence-Driven Early Detection of Neurological Disorders and Its Implications for Personalized Rehabilitation Strategies

Abstract

Early AD stratification is clinically relevant, and both structural MRI and cognitive–biomarker profiles contain complementary diagnostic information. We offer NeuroFusionAI, an ImageNet pre-trained ViT-Small/16 MRI branch (to extract relevant brain feature representation signals), combined with an FT-Transformer clinical branch via a bidirectional cross-attention module, and validate it on a four-class AD-stage task (NonDemented, VeryMildDemented, MildDemented, ModerateDemented). An important limitation is acknowledged: the imaging and clinical data are not sampled from the same patient cohort. The MRI set consists of 28,160 public access 2D brain images, while the clinical set is purely synthetic (6,000 records - 1,500/class) and is matched to images only at the level of diagnostic class, but never at the level of subject. It makes no claim of true multimodal correspondence per patient. One of the features (global CDR) was removed from consideration as a deterministic label proxy to avoid leakage. We circumvent optimistic bias from multiply-augmented near-duplicates with a perceptual-hash grouping procedure, which commits each near-duplicate group to exactly one data split. NeuroFusionAI achieves 98.18% accuracy and 0.9724 macro-F1 in under 5-fold stratified-group cross-validation with a held-out test set. Importantly, feature-concatenation fusion (0.9761) and a clinical-only XGBoost baseline (0.9770) equal or slightly outperform cross-attention fusion; differences between models are not statistically significant (Wilcoxon, n=5 folds). We view these results with skepticism; clinical-only performance near-ceiling is an expected byproduct of class-conditional synthetic generation, not evidence that clinical features alone will satisfy in real-world practice. Grouping and pairing tables available for full transparency.