Meta Analysis of the Diagnostic Efficacy of Transformer-Based Multimodal Fusion Deep Learning Models in Early Alzheimer's Disease

Hui GuoZiyu YangGaopan ZhangLingling LvXiongfei Zhao^*

• Xianyang Hospital of Yan 'an University, Xianyang, China

This study aims to systematically evaluate the diagnostic efficacy of Transformer-based multimodal fusion deep learning models in early Alzheimer's disease (AD) through a Meta-analysis, providing a scientific basis for clinical applications. Following PRISMA guidelines, databases such as PubMed and Web of Science were searched, and 20 eligible clinical studies (2022-2025) involving 12,897 participants were included. Study quality was assessed using the modified QUADAS-2 tool, statistical analyses were performed with Stata 16.0, effect sizes were pooled via random-effects models, and subgroup analyses, sensitivity analyses, and publication bias tests were conducted. Results showed that Transformer-based multimodal fusion models exhibited excellent overall diagnostic performance, with a pooled AUC of 0.924 (95% CI: 0.912–0.936), sensitivity of 0.887 (0.865–0.904), specificity of 0.892 (0.871–0.910), and accuracy of 0.879 (0.858–0.897), significantly outperforming traditional single-modality methods. Subgroup analyses revealed that: Three or more modalities achieved a higher AUC (0.935 vs. 0.908 for two modalities, P=0.012). Intermediate fusion strategies (feature-level, AUC=0.931) significantly outperformed early (0.905) and late (0.912) fusion (P<0.05 for both). Multicenter data improved AUC (0.930 vs. 0.918 for single-center, P=0.046), while sample size stratification (<200 vs. ≥200 cases) showed no significant difference (P=0.113). Hybrid Transformer models (Transformer +CNN) trended toward higher AUC (0.928 vs. pure Transformer 0.917, P=0.068) but did not reach statistical significance. Notable studies included Khan et al.'s (2024) Dual-3DM³AD model (AUC=0.945 for AD vs. MCI) and Gao et al.'s (2023) generative network (AUC=0.912 under data loss), validating model robustness and feature complementarity. Sensitivity analysis confirmed stable results (AUC range: 0.920–0.928), and Egger's test (P=0.217) and funnel plot symmetry indicated no significant publication bias. Limitations included a high proportion of single-center data and insufficient model interpretability. Future research should focus on multicenter data integration, interpretable module development, and lightweight design to facilitate clinical translation. Transformer-based multimodal fusion models demonstrate exceptional efficacy in early AD diagnosis, with multimodal integration, feature-level fusion, and multicenter data application as key advantages. They hold promise as core tools for AD "early diagnosis and treatment" but require further optimization for cross-cohort generalization and clinical interpretability.