This review was conducted in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 guidelines [19], with the search procedures reported following PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses literature search extension) [20] and developed using the principles outlined in the Cochrane Handbook [21]. These methods were applied to systematically identify and evaluate studies on computer-aided AD diagnosis, with a particular focus on those using multimodal clinical phenotyping datasets and multimodal linguistic-based cognitive impairment datasets.

We developed and internally reviewed independent search strategies (no external peer review). We manually searched multiple databases to identify AI-driven multimodal approaches for AD diagnosis, rather than using an integrated multidatabase platform. As this review targets methodological advances mainly reported in peer-reviewed computational literature, we did not search trial registries (ClinicalTrials.gov, World Health Organization’s International Clinical Trials Registry Platform). We also avoided validated or published filters, instead iteratively refining customized controlled-vocabulary and free-text terms for AD or dementia, multimodal data, and AI through pilot screening to maximize sensitivity.

Searches were performed in PubMed (447 records; January 1, 2019, to November 13, 2025), Scopus (1086 records; all years through November 13, 2025, filtered to PUBYEAR > 2018), IEEE Xplore (2229 records; January 1, 2020 to November 13, 2025), ACM Digital Library (2067 records; 1 January 2020, to 15 November 2025), Cochrane Library (1061 records; all available years through November 15, 2025), and arXiv (1081 records; all available years through November 15, 2025). We included the verbatim search strings for all databases, and because arXiv does not support bulk export, an arXiv search Python (Python Software Foundation) script is provided in Multimedia Appendix 1 [11,14,22-44].

The inclusion criteria were if studies were considered eligible if they met all the following conditions: (1) focused on AD, MCI, or related dementias as the primary clinical outcome; (2) applied AI or ML methods for computer-aided diagnosis, classification, or prediction; (3) used multimodal data, defined as any combination of at least two distinct modalities (eg, neuroimaging, clinical phenotyping, genetics, or linguistic features); (4) reported quantitative evaluation metrics; and (5) written in English.

The exclusion criteria were if studies met any of the following conditions: (1) single-modal approaches using only a single imaging modality, cognitive test, or biomarker, without any multimodal integration; (2) works without reported performance metrics or with insufficient methodological detail; (3) works not addressing diagnosis, classification, or prediction (eg, treatment response, drug trials, and lifestyle interventions); (4) duplicate publications or overlapping datasets without providing additional methodological contribution; and (5) non-English publications.

The study selection process followed the PRISMA 2020 guidelines, and the protocol was registered. The final search update was conducted in November 2025. All records retrieved from the databases were first imported into Zotero, where duplicates were automatically detected and removed.

The initial search identified 7435 records. After removing 3047 duplicates, 4388 records remained for title and abstract screening. A total of 4021 records were obviously irrelevant at the title and abstract level, 252 studies were excluded for the following main reasons:

Finally, 66 studies were included in the systematic synthesis, and all were successfully retrieved (reports not retrieved=0).

The workflow of AI-assisted AD diagnosis involves 3 stages, as illustrated in Figure 1. The initial stage involves comprehensive data acquisition, where information is collected from multiple modalities, including neuroimaging, biomarkers, genetics, and speech or behavioral signals. The second stage involves feature extraction and model development, followed by an interpretable and explainable analysis to ensure that AI models can effectively support clinical decision-making.

To ensure the clinical applicability and scientific rigor of computer-aided diagnosis models for AD, it is essential to systematically evaluate their performance using a variety of quantitative metrics. We have summarized all performance evaluation metrics in Multimedia Appendix 2.

We assessed methodological quality with QUADAS-2 [45], evaluating risk of bias in 4 domains. Patient selection raised the main concern: 61% (40/66) of outcomes were high risk due to poor reporting or nonrepresentative sampling. For the index test, 76% (50/66) were unclear risk because procedures and decision thresholds were insufficiently described, and 20% (13/66) were high risk. The reference standard showed 76% (50/66) unclear risk from limited methodological detail, with no high-risk ratings. Flow and timing had the greatest uncertainty: 85% (56/66) were unclear owing to missing information on testing intervals and participant flow. Figure 2 summarizes domain-level risk distributions; Multimedia Appendix 3 reports study-level assessments.

Given frequent unclear and high risk in key domains, we interpreted diagnostic performance cautiously, especially without external validation or a clearly defined reference standard. Future benchmarking should emphasize transparent reporting, prespecified thresholds, and multicenter evaluation to reduce bias and improve reproducibility.

Following study selection (the complete selection process is summarized in the PRISMA 2020 flow diagram, Figure 3), we first summarize the overall profile of the included literature to contextualize the subsequent synthesis. Figure 4 provides a temporal overview (2019‐2025) of modeling approaches across included studies, illustrating how methodological focus has shifted over time and informing interpretation of the evidence base.

Ensemble learning improves generalization and robustness by combining multiple base models, including bagging and boosting methods such as AdaBoost (Adaptive Boosting), XGBoost (Extreme Gradient Boosting), and LightGBM (Light Gradient-Boosting Machine), and has been widely applied in AD detection and progression prediction [12,58-60]. However, ensemble models may introduce redundant features, offer limited gains on small datasets, and incur higher computational costs, which can restrict real-time or resource-constrained deployment.

Traditional single-modality ML approaches can achieve high performance in AD-related tasks; however, they are constrained by several inherent limitations:

First, regarding information completeness, structural MRI alone has limited sensitivity to functional and molecular changes and cannot fully capture AD-related cognitive and behavioral alterations. Combining MRI with PET, neuropsychological tests, speech, electroencephalography (EEG), and genetic or biomarker data provides a more complete, multidimensional view of disease progression and patient heterogeneity [61].

Second, regarding model robustness, in multimodal data, residual noise in one modality may persist despite denoising, but other modalities can provide complementary signals that improve robustness. Leveraging multisensory-style integration, multimodal models better reflect biological cognition and can yield more reliable decisions [62].

Third, in cross-modal learning, transformer architectures use cross-modal attention to learn associations between modalities. Some studies apply them in weakly supervised or cross-modal guided settings, using one modality to constrain or guide representation learning in another [63].

Fourth, in real-world decision-making, multimodal learning better matches real-world diagnosis, which integrates multiple information sources. Using diverse modalities aligns models with clinical workflows and improves translational potential in practice [64].

Therefore, this review analyzes the methodological strengths and limitations of multimodal models for computer-assisted AD diagnosis, focusing on how dataset grouping and classification choices affect evaluation. By classifying datasets, we enable model comparisons under a unified setup, allowing for more direct assessment of generalizability and cross-dataset stability.

Beyond multimodal clinical phenotyping datasets, multimodal linguistic-based cognitive impairment datasets represent an equally vital research resource. These datasets offer a noninvasive and cost-effective methodology for detecting cognitive decline, particularly valuable for identifying early-stage or subtle impairments where traditional neuroimaging or biomarker data may yield inconclusive results. Capturing spontaneous or semistructured speech and language patterns pushes the development of AI in speech data. Recent work is shown in Table 3.

Recent studies have shown that multimodal fusion of speech and text using transformer-based architectures substantially improves AD detection performance, with F₁-scores above 0.90 on ADReSS and ADReSSo (Alzheimer’s Dementia Recognition Through Spontaneous Speech 2021 Challenge) datasets [121,132,139]. Linguistic feature engineering and interpretable language models further enhanced classification accuracy, achieving up to 92.2% accuracy and F₁-scores of 0.955 using compact part-of-speech features [124-126,128,130]. Cross-lingual approaches based on language-agnostic and transfer learning methods enabled moderate generalization, with accuracies ranging from 69% to 73.9% in English-Greek transfer settings [23,127,133,136]. To support real-world deployment, lightweight and hierarchical models achieved around 80% accuracy with reduced computational cost [131,135]. In addition, data augmentation and ensemble strategies improved robustness in low-resource scenarios, yielding F₁-score gains of 5%‐7% and competitive challenge performance (accuracy 86.69%) [123,137,138].

Table 2 and Table 3 summarize the recent state-of-the-art models across the 2 major types of multimodal datasets, extracted according to the Cochrane Handbook. Full QUADAS-2 forms are available in Multimedia Appendix 5. Based on these results, the following quantitative synthesis compares performance trends across all multimodal datasets. Across the 4 major dataset categories, modality choices and model performance show clear dataset-dependent patterns as shown in Table 4. UK Biobank studies mainly combine MRI, clinical variables, and genetic features, with 2 diagnosis studies reporting an average accuracy of 71.4% (SD 5.2%) and 4 risk-prediction studies reaching an average AUC of 0.84 (SD 0.056). ADNI studies use the most comprehensive modality integrations, with 3 diagnosis studies averaging 92.5% (SD 3.8%) accuracy, 3 MCI-conversion studies achieving a mean AUC of 0.922 (SD 0.045), and risk-prediction studies reaching an average AUC of 0.81 (SD 0.06); these tasks collectively achieve the strongest results, with fusion models frequently reporting AUC values above 0.95. DementiaBank studies differ fundamentally by focusing on speech- and language-based modalities; 9 diagnosis studies report an average AUC of 0.813 (SD 0.042), and 5 cross-lingual AD-detection studies show a mean accuracy of 77% (SD 6.5%), where transformer architectures consistently outperform classical approaches, with models such as BERT + DeiT (Data-Efficient Image Transformers), BERT + ViT (vision transformer), and RoBERTa + (Robustly Optimized Bidirectional Encoder Representations From Transformers Approach) DNN (deep neural network) showing F₁-scores exceeding 0.90. Self-collected datasets are typically smaller and more heterogeneous; 3 diagnosis studies report an average accuracy of 96% (SD 2.4%), and lightweight models such as EEGNet or ViT-based hybrids demonstrate strong predictive capacity when applied to EEG or structural MRI.

To interpret these results and limit metric inflation, note that purely internal cross-validation tends to overestimate performance: AUC is typically ≈5‐15 points higher than with external validation. Small or tightly controlled datasets also report accuracies ≈10%‐20% above those in large, heterogeneous cohorts. Severe class imbalance can further raise accuracy while lowering F₁-score or sensitivity; without correction, imbalance may inflate results by ≈5%‐12%. Cross-sectional models often score higher in single-timepoint evaluations, whereas longitudinal designs usually yield lower but more stable estimates, which are more informative for follow-up and clinical use.

These findings should be interpreted in light of substantial heterogeneity and risk of bias. Variation in sample composition, task definitions, and evaluation procedures across datasets limits direct comparison of performance metrics. QUADAS-2 also indicated frequent unclear and high-risk in-patient selection, reference standards, and flow or timing, especially in studies using only internal validation or selected samples. Reported metrics, therefore, likely represent upper-bound estimates rather than expected real-world performance, and apparent gains often reflect dataset-specific effects rather than generalizable model superiority.

Overall, the evidence shows that modality effectiveness varies substantially across datasets, transformer models deliver the highest gains in speech-language tasks, and large clinical phenotyping datasets such as UK Biobank and ADNI still rely mainly on traditional machine-learning or custom fusion frameworks rather than modern cross-modal transformers. This gap highlights an opportunity to develop transformer-based multimodal integration approaches tailored to large, heterogeneous clinical datasets.

A structured multimodal fusion taxonomy clarifies the performance of different integration strategies across datasets (Tables 2 and 3). A total of 4 main paradigms are commonly used: early, intermediate, late, and attention- or graph-based fusion.

Early fusion concatenates low-level features and performs well for aligned modalities such as MRI + PET, often achieving AUC>0.95 in ADNI studies, but is sensitive to missing data and feature-scale heterogeneity. Intermediate fusion combines latent representations from modality-specific encoders and is effective for heterogeneous inputs such as MRI + speech or EEG + clinical data, as demonstrated by high performance in ADReSS-based models, although it may be unstable in small datasets. Late fusion aggregates model outputs and is robust to missing modalities, performing well in large datasets such as the UK Biobank, but underuses fine-grained cross-modal interactions.

Across paradigms, limited modality availability and high acquisition costs remain key challenges, underscoring the need for adaptive and clinically feasible fusion strategies.

This review synthesized multimodal AI studies for AD across diverse dataset families, including clinical phenotyping and cognitive-linguistic datasets. Multimodal fusion generally outperformed unimodal baselines, but the gain is dataset-dependent and should be interpreted cautiously. Strong performance in curated cohorts and constrained speech benchmarks may not generalize to population-based or multicenter settings. QUADAS-2 also indicated frequent risk of bias and unclear reporting across domains, likely inflating metrics and limiting comparability. Accordingly, headline accuracy and AUC should be treated as upper-bound estimates unless supported by external validation and transparent reporting.

In recent years, multimodal models have demonstrated remarkable potential in computer-aided diagnosis and risk prediction for AD. While these methods have achieved significant successes, several challenges remain that warrant careful examination. In this section, this systematic review summarizes the common limitations identified in existing studies and proposes directions for future research to advance the field.

Multimodal AI could support AD diagnosis through several clinical pathways. In memory clinics, models combining MRI, cognitive scores, and blood biomarkers could triage referrals, prioritizing patients for specialist review or PET. In general practice, speech-based and routine clinical-feature models could be embedded in consultations to flag early cognitive change. In radiology, MRI-clinical fusion could act as a second reader, reducing interobserver variability and supporting less experienced clinicians. Where imaging or specialist access is limited, speech, digital questionnaires, and basic clinical data could enable telemedicine-based screening and follow-up. At the population level, these models could support risk stratification and targeted monitoring. To enable real-world deployment, research should prioritize external multicenter validation, integration with electronic health records, and evaluation of regulatory feasibility, cost-effectiveness, and clinical impact.

Deploying multimodal AI for AD diagnosis requires ethical and regulatory safeguards. As datasets often combine imaging, clinical records, genomics, and speech, they fall under strict privacy regimes (eg, General Data Protection Regulation in the European Union; HIPAA [Health Insurance Portability and Accountability Act] in the United States), requiring explicit consent, data minimization, and secure handling, with added complexity for sensitive modalities such as speech and genomic data. Clinical deployment is also shaped by medical-AI governance frameworks (eg, the European Union AI Act, Food and Drug Administration Software as a Medical Device guidance, and UK Medicines and Healthcare products Regulatory Agency Good Machine Learning Practice), which emphasize transparency, risk management, and postdeployment monitoring. Fairness is essential because demographic imbalance can yield uneven performance across age, ethnicity, and language groups. Interpretability (eg, imaging attention maps and linguistic saliency) supports clinical accountability and aligns with explainability expectations. Future work should incorporate privacy-preserving methods, bias audits, and regulatory-aligned validation pipelines to enable responsible clinical integration.

Access to multimodal AD data remains severely restricted by privacy regulations and ethical constraints, which limit data sharing and external validation. This restricts the sharing and usage of comprehensive datasets needed for robust external validation and generalizability.

Federated learning (FL) provides a technically viable privacy-preserving solution; however, differences in data formats and institutional infrastructures still impede its large-scale deployment. For instance, Meerza et al [134] pioneered FL for AD speech diagnosis using mel-frequency cepstral coefficients and pause features, maintaining model performance while ensuring privacy through q-FedAvg/q-FedSGD optimization. Nambiar [129] validated an ALBERT (A Lite Bidirectional Encoder Representations From Transformers) + BiLSTM (bidirectional long short-term memory) hybrid model on the ADReSS dataset, achieving strong performance without compromising data confidentiality. In parallel, multi-institutional collaborations leveraging publicly available datasets such as ADNI, UK Biobank, and OASIS have enabled richer external validation while adhering to rigorous privacy standards [15,79,88,95,100,139,140].

Despite encouraging results, FL still lacks harmonized protocols and interoperable platforms. This limits cross-center reproducibility and weakens clinical credibility. International collaboration also remains constrained by regulatory differences. Future work should prioritize unified federated frameworks with standardized protocols and privacy-preserving methods to enable secure global data collaboration [143,144].

As most datasets lack fully matched modalities per participant, multimodal fusion often relies on representation- or population-level integration rather than early fusion. Early fusion requires paired samples and is therefore infeasible across datasets. By contrast, late fusion and embedding-level integration can train unimodal models separately and combine them via meta-learners, cross-modal transformers, or probabilistic ensembles. Domain adaptation, transfer learning, and harmonization can also combine heterogeneous cohorts at the population level to improve generalizability. A standardized benchmark could further support this by defining shared preprocessing, label taxonomies, and evaluation metrics, enabling meaningful comparison or representation-stage fusion even without subject-level pairing.

Severe class imbalance remains a major obstacle, biasing training toward the majority class and inflating accuracy while masking low sensitivity to early disease. In addition, datasets such as the UK Biobank are dominated by White European ancestry, limiting generalizability across racially and ethnically diverse populations. Addressing this requires both technical mitigation and proactive recruitment of underrepresented groups so models better reflect population heterogeneity.

Researchers have applied data-level interventions such as generative adversarial network–based augmentation, diffusion models, and resampling [123,137,138,145,146]; algorithm-level solutions, including cost-sensitive, loss-focused, ensemble, and class-weighted training schemes [147-152]; and evaluation-focused remedies [153] have been developed to mitigate biases.

Current methods frequently introduce new challenges, such as overfitting or inadequate performance in minority classes. Moreover, efforts to increase diversity remain inadequate. Future directions should focus on novel adaptive resampling methods, generative methods for synthetic minority data creation, and dedicated efforts to include and characterize underrepresented populations to ensure equitable and robust clinical applicability across diverse populations.

Differences in acquisition protocols and diagnostic criteria across datasets limit comparability of imaging, cognitive, and biomarker outcomes. Longitudinal evidence is also constrained: even in relatively standardized resources such as ADNI, limited long-term follow-up hampers modeling the temporal dynamics of disease progression.

Future work should standardize key acquisition elements and diagnostic criteria across longitudinal studies and strengthen coordination across institutions. Building on this, a multimodal benchmark spanning imaging, clinical, biomarker, behavioral, and linguistic modalities would enable cross-dataset validation, improve comparability, and support reproducible evaluation of new models. These steps would strengthen temporal modeling and provide more reliable evidence for clinical translation.

Data imbalance is prevalent across many AD datasets, but the nature of this issue varies substantially between cohorts. This review, therefore, outlines the dataset-specific limitations of commonly used AD cohorts and corpora.

ADNI participants are generally healthier, with fewer comorbidities and a restricted age range (55‐90 y), limiting representativeness. Protocol differences across centers and evolving diagnostic standards introduce heterogeneity, while frequent reliance on subsets hampers comparability [154].

Of UK Biobank, dementia outcomes are derived mainly from health records, leading to potential misclassification and delayed ascertainment. Participants show strong volunteer bias, and PET or cerebrospinal fluid biomarkers are limited to a small subset, constraining multimodal analyses [155].

OASIS provides open neuroimaging data but with relatively small AD/MCI sample sizes and inconsistent modality coverage. Limited longitudinal depth and cross-scanner variability further reduce reproducibility [156].

Of NACC, data are aggregated from multiple centers with heterogeneous recruitment and diagnostic protocols, making harmonization challenging. The cohort is clinic-based rather than population-representative, and missing biomarker modalities are common [157].

Although high-quality, Australian Imaging, Biomarkers and Lifestyle Study is smaller than ADNI and NACC and is often used only for validation. Regional recruitment and protocol differences reduce ethnic diversity and cross-cohort comparability [158].

Of Pitt Corpus, this is the most widely used speech dataset, but remains small and imbalanced. Tasks are constrained, limiting ecological validity, and cross-linguistic generalizability is poor [159].

Of the ADReSS series, the ADReSS benchmarks provide standardized speech corpora but are modest in size and restricted to English. Narrow task design and small training partitions raise concerns of overfitting and limited external validity [18].

Of self-collected cohorts, locally collected datasets often involve small, single-site samples with heterogeneous acquisition protocols. Missing modalities, limited follow-up, and selection bias further restrict their generalizability [153].

Dataset challenges are compounded by unrepresentative cohorts, incomplete modalities, and poor cross-center consistency, limiting model robustness and cross-dataset generalization in AD diagnosis. Future work should improve data coordination and standardization, enable more practical sharing mechanisms, and adopt cross-cohort validation where feasible. Strengthening data quality and access is essential for translating multimodal AI methods into clinical use.

A major limitation of multimodal ML models in clinical AD diagnosis is limited interpretability and transparency. Many high-performing models provide insufficient insight into their decision processes, which can hinder clinical adoption and reduce confidence among end users.

Efforts that have been made toward model interpretability include designing inherently transparent models. For example, some studies demonstrate emerging explainability strategies, including hybrid neuro-symbolic models [160] that generate interpretable reports and post hoc methods such as SHAP, LIME (Local Interpretable Model-Agnostic Explanations), gradient-based saliency, and graph-masking techniques [161,162], which collectively enhance transparency in multimodal AD diagnosis.

Current interpretability methods often fail to produce explanations that clinicians can use reliably. Future work should prioritize clinically grounded explainability, including interactive visualizations and concise workflow-aligned natural-language summaries. Hybrid designs that combine deep learning with structured reasoning can further improve transparency by making decision logic explicit. For deployment, models should also report prediction uncertainty and demonstrate compatibility with clinical systems and regulatory requirements.

Beyond technical advances, incorporating patient and public involvement can improve multimodal AI development for AD. Patients and caregivers can help shape evaluation and result communication, not just act as end users, aligning explanations with patient priorities and addressing transparency and fairness. Engaging patient and public involvement earlier in model design may therefore support more interpretable and clinically usable diagnostic tools.

Integrating data across studies is challenging because single datasets rarely cover all modalities, forcing combinations such as ADNI with UK Biobank. However, differences in cohorts, imaging protocols, and cognitive assessment frameworks create substantial heterogeneity that limits direct pooling and comparability.

This heterogeneity hinders building unified models that generalize across nonoverlapping cohorts, so single-dataset models often fail out of domain. Platform-agnostic methods that tolerate missing or inconsistent modalities are therefore needed. Proposed solutions include shared latent-space learning [163], multibranch networks [164], and mixture-of-experts architectures [165] to support partial fusion and cross-dataset adaptation, but most still assume strong cross-domain alignment or require substantial retraining under domain shift.

Despite recent progress, multimodal methods often assume strict cross-domain alignment and require extensive retraining under domain shift or missing modalities. Future work should develop robust, platform-agnostic frameworks that adapt to changing modality availability and distribution shifts with minimal performance loss and advance representation learning to derive stable joint embeddings from heterogeneous data.

Although multimodal AD models have advanced, most studies still omit uncertainty quantification (eg, confidence or prediction intervals). Models typically provide deterministic outputs without communicating reliability, despite clinicians relying on uncertainty to guide management and treatment decisions. Future work should embed uncertainty metrics into diagnostic models to better align with clinical needs and improve interpretability, reliability, and real-world adoption.

Another limitation is data leakage, which can inflate performance. Common forms include subject-level leakage (samples from the same participant in both training and test sets), patch-level overlap in MRI slice and patch models, and transcript or utterance-level leakage in speech datasets when multiple segments come from 1 individual. Many studies did not report whether participant-independent splits were enforced. Clearer reporting of partitioning and rigorous participant-level cross-validation are therefore essential to ensure real-world generalizability.

This review synthesizes evidence on multimodal AI approaches for AD across clinical, neuroimaging, genetic, and linguistic data, systematically comparing modeling strategies, validation practices, and performance trends across heterogeneous datasets. In contrast to prior modality-specific reviews, the findings show that multimodal models generally outperform unimodal approaches, although performance varies substantially with dataset characteristics, modality availability, and cross-source alignment. High accuracies are often reported in curated or internally validated cohorts, whereas population-based and externally validated studies yield more modest but clinically realistic results, reflecting substantial heterogeneity and risk of bias.

Despite these limitations, the evidence demonstrates that multimodal AI captures complementary biological and behavioral signals relevant to AD, offering clear advantages for diagnosis and risk prediction. Transformer-based architectures and speech- or behavior-derived modalities show promise for scalable and noninvasive early detection. However, meaningful clinical translation will require harmonized benchmarking, transparent reporting, and rigorous external validation. Overall, this review advances the field by contextualizing performance gains within their methodological constraints and by outlining practical directions for developing robust, interpretable, and generalizable multimodal AI systems. These insights support the responsible integration of AI into real-world dementia screening, risk prediction, and early intervention strategies.