A stratified search strategy was employed to comprehensively cover core Chinese and English databases, including PubMed, Web of Science, Embase, CNKI, and Wanfang Data, with a search timeframe from January 2017 to April 2025 (encompassing the full research cycle after the Transformer architecture was proposed) (21). Search keywords combined disease terms (AD, mild cognitive impairment, etc.), technical terms (Transformer, multimodal fusion, deep learning, etc.), and diagnostic scenarios (early diagnosis, classification, prediction, etc.). Reference lists of included studies and cited literature in relevant reviews were also traced to avoid omissions (22). Inclusion criteria were: (1) Clinical studies on early AD diagnosis (including AD vs. normal control, MCI vs. normal control, and AD vs. MCI) (23). (2) Integration of at least two modalities (e.g., imaging, clinical indicators, genetic data) (24). (3) Explicit use of Transformer core architecture (self-attention mechanism or encoder-decoder structure) for multimodal fusion, with reported diagnostic efficacy metrics (ACC, SENS, SPEC, AUC, etc.) (25). (4) Sample size ≥30 cases per group (26). (5) Journal articles in Chinese or English. Exclusion criteria included single-modality analysis, non-Transformer models, duplicate publications, incomplete data, or non-journal literature (27).

Data extraction was independently performed by two researchers with backgrounds in medical imaging and deep learning, with discrepancies resolved through consultation with a third-party expert. Extracted information included basic study details (author, year, and region), design characteristics (sample source, modality combination, and sample size), model specifics (Transformer type, fusion strategy, training method, and validation approach), diagnostic efficacy (core metrics and 95% confidence intervals), and bias risk indicators (data preprocessing, blind method implementation, and missing data handling) (28). The modified QUADAS-2 tool was used to assess literature quality, focusing on patient selection bias, index definition bias, and model validation bias to ensure methodological rigor of included studies (29).

Heterogeneity was assessed using Cochran’s Q test and I² statistic. If I² ≤ 50% and p ≥ 0.1, a fixed-effect model (Mantel–Haenszel method) was used to pool effect sizes. If significant heterogeneity existed (I² > 50% or p < 0.1), subgroup analysis (modality type, fusion strategy, dataset characteristics, model architecture) or random-effects model (DerSimonian–Laird method) was employed to explore sources (30). Core analyses included pooling diagnostic efficacy metrics (AUC, Sens, Spec, and ACC) and Drawing Forest plot, with subgroup analyses comparing efficacy differences across modality combinations (bimodal vs. multimodal), fusion strategies (early vs. late vs. intermediate fusion), data characteristics (single-center vs. multicenter, sample size stratification), and model architectures (pure Transformer vs. hybrid models). Sensitivity analysis evaluated result stability by sequentially excluding individual studies. Publication bias was detected via Egger’s test and funnel plot symmetry analysis, with Trim-and-Fill correction applied if bias risk was identified (31).

Stata 16.0 was used for meta-analysis and visualization, RevMan 5.4 for bias risk assessment, and EndNote X9 for literature management, ensuring reproducible analysis processes compliant with statistical norms. This study aims to objectively quantify the diagnostic efficacy of Transformer-based multimodal fusion models through systematic search, strict quality control, and rigorous statistical inference, providing a scientific basis for clinical application and methodological optimization.

A total of 3,287 articles were obtained through a hierarchical retrieval strategy. After the initial screening of titles and abstracts, 2,142 duplicate and irrelevant studies were excluded. After a detailed reading of the full texts, 1,025 studies that did not meet the inclusion criteria (such as single–modality, non-Transformer architecture, data missing, etc.) were excluded. Finally, 20 eligible clinical studies were included, as shown in Figure 1. The included studies were all published from 2022 to 2025, covering six countries (six from the United States, eight from China, three from Germany, two from the United Kingdom, and one from Lithuania), and included 12,897 subjects (3,452 in the AD group, 4,121 in the MCI group, and 5,324 in the normal control group).

All 20 studies adopted Transformer architecture combined with multimodal data (Table 1, feature summary table omitted), as follows:

In Figure 2, the basic characteristics of the 20 included studies reflect the methodological features and potential limitations of current early AD diagnosis research. In terms of modality combinations, trimodal, and higher-fusion studies accounted for 60% (12 studies), significantly higher than bimodal studies (40%). Additionally, 15 studies included structural imaging (MRI), and 12 integrated functional imaging (PET), indicating that multimodal imaging data remain dominant. However, the integration rate of non-imaging data such as clinical indicators was only 40% (8 studies), suggesting that cross-modal information fusion could be further strengthened in the future. In terms of model architecture, hybrid Transformer (Transformer +CNN/RNN) models accounted for 70% (14 studies), while pure Transformer models comprised only 30% (six studies), reflecting researchers’ preference for optimizing feature extraction by combining traditional networks with Transformer. Feature-level fusion (intermediate fusion) was the dominant strategy (55%), consistent with the subgroup analysis conclusion that this strategy yields the best performance. Regarding validation methods, 80% of studies used 10-fold cross-validation, but only 20% included external independent validation sets, which may affect the evaluation of model generalizability. Quality assessment showed that all studies achieved QUADAS-2 scores ≥11/14, but multicenter data were used in only 35% (seven studies), and there was significant variability in the transparency of blind method implementation (explicitly reported in 12 studies). These findings highlight the need to address the potential impact of data heterogeneity and methodological rigor on research outcomes.

After sequentially excluding individual studies, the AUC fluctuated between 0.920 and 0.928, with stable pooled effect sizes, indicating that the results were not significantly influenced by any single study. Egger’s test showed a p-value of 0.217, and the funnel plot exhibited good symmetry, suggesting no significant risk of publication bias.

In Figure 5, the funnel plot and sensitivity analysis indicate that after sequentially excluding individual studies, the AUC fluctuates only between 0.920 and 0.928, with highly stable pooled effect sizes. This suggests that the meta-analysis results are not dominated by any single study, demonstrating strong robustness. Egger’s test shows a p-value of 0.217, and the funnel plot exhibits good symmetry, indicating no significant publication bias and a balanced distribution of included studies. These two results jointly validate the reliability of the research conclusions, showing that the high efficacy of Transformer-based multimodal fusion models in early AD diagnosis does not originate from data bias or outliers in individual studies, providing more credible evidence support for the clinical translation of the models.

In the AD vs. MCI discrimination task, Khan et al.’s (20) Dual-3DM³AD model achieved an AUC of 0.945 (95% CI: 0.931–0.958) through triplet preprocessing and 3D hybrid Transformer, representing the current highest efficacy. For incomplete data scenarios, Gao et al.’s (16) multimodal Transformer generative network maintained an AUC of 0.912 (95% CI: 0.895–0.927) when MRI/PET data were missing, validating the model’s robustness. Odusami et al.’s (14, 15) pixel-level ViT fusion achieved an AUC of 0.897 (95% CI: 0.876–0.915) in single-modality MRI analysis, demonstrating Transformer’s high-resolution representation capability for imaging details.

In Figure 6, the performance differences of typical models in AD vs. MCI discrimination are demonstrated. Together, these findings indicate that the Transformer architecture significantly enhances the accuracy and adaptability of early AD diagnosis through modality integration, strategy optimization, and single-modality deepening. Based on the performance advantages of trimodal fusion as well as issues related to overfitting and validation datasets, the following analysis conducts a comparison using the subgroup data of 20 included studies from the dimensions of core diagnostic indicators, validation methods, and result stability. This comparison aims to provide more detailed support for the superiority of trimodal fusion. Table 1 presents the comparative evaluation results between bimodal and trimodal fusion.

In Table 1, the data clearly indicates that the diagnostic AUC of trimodal and above fusion is significantly higher than that of bimodal fusion (0.935 vs. 0.908, p = 0.012). The synergistic effect of multi-source data significantly improves the diagnostic performance for early AD. Although some studies did not adopt independent external validation, the subgroup heterogeneity of trimodal fusion is lower (65.8%), and the overall sensitivity analysis confirms the stability of the results (AUC fluctuation: 0.920–0.928). This suggests that the risk of overfitting is controllable, further verifying the advantages of trimodal fusion. Table 2 presents the results of further significance analysis.

In Table 2, the issue of confused validation types is significant: the AUC of cross-validation (0.931) is higher than that of independent external validation (0.905). Moreover, 80% of the studies rely on cross-validation, while only 20% adopt external validation. Combined analysis is likely to falsely inflate accuracy. It is necessary to split subgroups as recommended and take the results of external validation as the basis for core conclusions. In terms of confused diagnostic tasks, the AUC of AD vs. NC is the highest (0.942) due to obvious pathological features, whereas the AUC of MCI vs. NC, which is more critical for early diagnosis, is the lowest (0.897). Combined analysis will mask the model’s weakness in identifying mild cognitive impairment (MCI). It is required to present the results of each task separately and clearly define “early AD” to demonstrate the rationality of combination. The potential issue of dataset overlap is prominent: 75% of the studies rely on the ADNI dataset, and none of the studies verified the overlap of participants. There is a hidden risk of “false precision” in results caused by duplicate counting. It is necessary to supplement the dataset list of the 20 studies and optimize the analysis through sensitivity analyses such as excluding duplicate data. Regarding statistical models, all studies used the random-effects model to pool indicators individually, without adopting the bivariate/HSROC models recommended by PRISMA-DTA. This ignores the correlation between sensitivity and specificity as well as differences in diagnostic thresholds. It is essential to acknowledge this limitation, discuss its potential impact on result bias, and thereby improve the credibility of the study conclusions.

This study systematically evaluated the efficacy of Transformer-based multimodal fusion deep learning models in early AD diagnosis through meta-analysis. Results showed that these models demonstrated significant advantages in distinguishing AD from normal controls and mild cognitive impairment (MCI), with an overall AUC of 0.924 (95% CI: 0.912–0.936), significantly superior to traditional single-modality methods (previous studies reported ACC of approximately 0.78–0.82) (7, 12). This finding confirms the unique value of the Transformer architecture in capturing complex correlations in cross-modal data, as its self-attention mechanism effectively models long-range dependencies in multi-source data (such as MRI, PET, and clinical indicators), addressing the limitations of traditional methods in detecting subtle early pathological changes (9, 11).

Subgroup analyses reveal several key influencing factors: Trimodal and above fusion achieves a significantly higher AUC (0.935 vs. 0.908, p = 0.012), indicating a synergistic effect of multi-source data integration. This is consistent with Tang et al.’s. (13) conclusion that dynamic modality attention mechanisms can optimize cross-modal feature weight allocation. Intermediate fusion strategy (feature-level fusion) shows superiority (AUC = 0.931), further suggesting that integrating cross-modal information during the feature extraction stage is more conducive to capturing complex pathological features. This may be related to the strategy’s ability to preserve raw data details and avoid early information loss (15). Multicenter studies have higher AUC (0.930 vs. 0.918, p = 0.046), highlighting the importance of data heterogeneity management for model generalizability. However, no significant difference is observed in sample size stratification, indicating that the current data scale generally meets model training requirements (16).

Comparisons of typical studies highlight the clinical value of technological innovations: Khan et al.’s (20) Dual-3DM³AD model achieved an AUC of 0.945 through triplet preprocessing and 3D hybrid Transformer, validating the synergistic advantages of deep feature engineering and multi-task learning. Gao et al.’s (16) generative network maintained an AUC of 0.912 in scenarios with missing data, demonstrating the adaptability of cross-modal completion technology to real-world data. Odusami et al.’s (14) pixel-level ViT fusion reached an AUC of 0.897 in single-modality MRI analysis, proving Transformer’s capability for high-resolution representation of imaging details. These results collectively indicate that innovations in model architecture (such as hybrid Transformer) and optimization of data fusion strategies are core pathways to improving diagnostic efficacy.

Although this study confirmed stable results (AUC fluctuation: 0.920–0.928) and no significant publication bias (Egger’s test, p = 0.217) through sensitivity analysis, the following limitations should be noted: First, only seven of the included studies use multicenter data, and single-center bias may limit the model’s performance in cross-cohort generalization (4, 6). Second, the efficacy difference between hybrid Transformer and pure Transformer models do not reach statistical significance (*p* = 0.068), indicating that the feature complementarity mechanism between traditional neural networks and Transformer requires further validation (11). Additionally, insufficient model interpretability remains a major obstacle to clinical application, as the black-box nature of attention mechanisms struggles to meet the transparency requirements of diagnostic decision-making (28).

Future research needs to focus on three major directions: First, promoting standardized integration of multicenter data and reducing the impact of data heterogeneity through technologies such as federated learning. Second, developing interpretability modules, such as introducing attention heatmaps to visualize brain region-pathology associations (23). Third, optimizing lightweight model design by borrowing the attention bottleneck mechanism proposed by Kadri et al. (19) to balance computational requirements and diagnostic accuracy. With the deep integration of Transformer technology with medical imaging and clinical data, such models are expected to become core tools for early and precise AD diagnosis, providing critical support for achieving the clinical goal of “early detection and early intervention.”