A Hybrid Cross-modal Attention Framework for Early Alzheimer's Disease Prediction: A Multimodal Ablation and Benchmarking Study

Alzheimer's disease (AD) is a leading cause of disability among ageing populations worldwide, and early, accessible, and interpretable prediction tools remain a pressing need, particularly in low-resource settings. This study develops a hybrid cross-modal attention framework that fuses imaging-derived, clinical, and genetic features to predict cognitive status (Cognitively Normal [CN], Mild Cognitive Impairment [MCI], Alzheimer's Disease [AD]). It benchmarks it against classical machine-learning baselines under a controlled modality-ablation design. The framework was trained and evaluated on a harmonized ADNI-derived cohort of 1,163 participants (44.4% CN, 41.9% MCI, 13.8% AD). Each modality seven volumetric MRI-derived features, four clinical/cognitive features, and APOE4 genotype, is projected into a shared 8-dimensional embedding space and fused via 4-head cross-modal attention, and the resulting representation is trained jointly with a softmax classifier using Adam (custom NumPy implementation). A modality-ablation analysis shows that clinical features alone (MMSE, age, sex, education) already achieve AUC ≈ 0.81–0.84, with imaging-only and genetic-only models performing markedly worse (AUC ≈ 0.71–0.74 and ≈ 0.57, respectively); full multimodal fusion yields a modest but consistent improvement over any single modality. On a held-out ADNI test set (n = 233), the hybrid model achieved accuracy 0.678, balanced accuracy 0.664, macro F1 0.687, and AUC-OVR-macro 0.833 (95% bootstrap CI [0.794, 0.872], 2,000 resamples), with 5-fold cross-validated accuracy 0.695 ± 0.018 and AUC 0.860 ± 0.020. The hybrid model was competitive with, but did not uniformly outperform, the strongest classical baselines (Random Forest accuracy 0.704; Logistic Regression AUC 0.856) on the same test set, confidence intervals for all five models overlap substantially, and a paired bootstrap significance test (2,000 resamples, identical resample indices across models) found no statistically significant accuracy difference between the hybrid model and any classical baseline; on AUC, the hybrid model was significantly outperformed by Logistic Regression (mean difference −0.023, 95% CI [−0.041, −0.004], p = 0.014), with no significant AUC difference against the other three baselines. Integrated-gradients attribution identified MMSE score as the dominant feature (normalized importance 0.36), followed by hippocampal volume, sex, and APOE4 status; cross-modal attention weights showed each modality attending most strongly to itself (0.28–0.39) while incorporating meaningful cross-modal information. The study demonstrates that cross-modal attention fusion is a feasible and interpretable, but not statistically superior, approach to multimodal AD prediction relative to simpler classical baselines on this dataset, and identifies clinical features as the dominant driver of predictive performance. Fairness auditing and external validation are reported in a companion paper.