Alzheimer's disease (AD) represents the predominant cause of dementia worldwide (Alzheimer's Association, 2024). In its typical course, AD progresses gradually, beginning with a prolonged prodromal phase characterized by subtle memory impairments. As the disease advances, cognitive deficits become more pronounced and extend to other domains, eventually resulting in significant functional decline (Peña-Casanova et al., 2012).

Nevertheless, the progression of AD exhibits considerable interindividual variability. This heterogeneity is influenced by a combination of environment, genetic predisposition, and levels of cognitive reserve, among other factors (Hersi et al., 2017; Duara and Barker, 2022). Moreover, the presence of copathology further increases the variability in neurocognitive trajectories (Lam et al., 2013).

In this context, despite advancements in diagnostic and early detection methods (Ansart et al., 2021; Garćıa-Gutiérrez et al., 2022, 2024a; Sharma et al., 2023; Lee et al., 2024), tools for estimating cognitive decline remain underdeveloped. This limitation is particularly critical, as accurate trajectory prediction is essential for clinical decision-making, therapeutic planning, and patient selection for clinical trials (Abdelnour et al., 2022; Hampel et al., 2022). In fact, anticipating disease progression can substantially reduce the number of participants needed in clinical trials, with corresponding reductions in cost and patient exposure (Duara and Barker, 2022).

Accordingly, much of the existing research on modeling cognitive decline has adopted a qualitative approach, primarily focusing on transitions between clinical stages, such as from mild cognitive impairment (MCI) to dementia, with reported classification accuracies ranging from 75% to 95% (Ansart et al., 2021; Sharma et al., 2023; Garćıa-Gutiérrez et al., 2024b; Yang et al., 2025). However, predicting diagnostic transitions tends to oversimplify the complexity of cognitive decline, as it overlooks the finer-grained dynamics of domain-specific cognitive trajectories.

In contrast, a smaller subset of studies has aimed to model cognitive decline quantitatively. These efforts typically involve forecasting future scores on a limited number of neuropsychological assessments within defined time windows (Dansson et al., 2021; Lei et al., 2022; Maheux et al., 2023; Devanarayan et al., 2024). Commonly targeted measures include the Mini-Mental State Examination (MMSE), the Clinical Dementia Rating (CDR), and the cognitive sub-scale of the Alzheimer's Disease Assessment Scale (ADAS-Cog), often in the context of the TADPOLE challenge (Marinescu et al., 2019). Notably, recent studies have expanded the range of neuropsychological tests considered, incorporating measures such as the Rey Auditory Verbal Learning Test (RAVLT) (Seo et al., 2024), the Preclinical Alzheimer's Cognitive Composite (Devanarayan et al., 2025), and composite cognitive domains (Moradi et al., 2025). These advancements represent meaningful progress toward modeling cognitive decline with enhanced clinical relevance.

Furthermore, although these studies utilize a wide range of biomarkers for modeling, the representation of neuroimaging data remains restrictive. Typically, features are extracted from a single imaging modality, or occasionally from multiple modalities, and treated as tabular data. However, this ignores spatial and topological information inherent to brain structure (Maheux et al., 2023; Devanarayan et al., 2024).

This approach contrasts with other applications in which the superiority of artificial intelligence (AI) models incorporating three-dimensional or graph-based information has been well established (Li et al., 2021; Sharma et al., 2023). This is particularly relevant given the inherent complexity of modeling cognitive decline, a task considerably more challenging than diagnosis or conversion prediction. Enhancing predictive performance in this setting requires more effective utilization of neuroimaging data, one of the most informative biomarkers (Nordberg et al., 2010).

In this context, we aim to model domain-specific cognitive trajectories and to predict diagnostic transitions between clinical stages of the AD continuum. To this end, we introduce a multimodal learning framework that exploits complementary data representations and maximizes the use of available neuroimaging information.

Whereas many prior approaches rely primarily on global cognitive assessments such as the MMSE, our method specifically targets the core cognitive domains typically affected in AD, including memory, language, visuospatial abilities, and executive functions. Furthermore, to obtain a more robust quantification of cognitive decline and to mitigate the noise associated with single time-point measurements, we focus on modeling longitudinal trajectories of cognitive performance.

Methodologically, the proposed framework incorporates two key strategies to improve predictive performance. First, we introduce a pre-training stage to exploit the full dataset, thereby alleviating common limitations associated with high-dimensional neuroimaging data and relatively small sample sizes. Second, we systematically explore alternative data representations to enhance model accuracy and efficiency. Specifically, we evaluate tabular features derived from multiple neuroimaging modalities, leverage three-dimensional information through convolutional neural networks (CNNs), and incorporate brain connectivity patterns via graph neural networks (GNNs).

Through this approach, we aim not only to enhance the predictive accuracy of cognitive decline but also to advance toward more detailed and clinically relevant modeling. This approach has the potential to optimize intervention strategies, support treatment personalization, and improve patient selection for clinical trials.

The following sections present the results of future diagnosis and cognitive decline prediction. Models were labeled as ni (non-imaging) when neuroimaging data were excluded, nd when neuropsychological and demographic information were included, and pt when pre-training was incorporated. By default, if none of these labels were specified, it was assumed that the models had been trained using only neuroimaging data.

This study addresses the pressing need for domain-specific prediction of cognitive decline in AD, moving beyond conventional global cognitive assessments. Despite the inherent challenges associated with neuroimaging-based AI applications, such as high data dimensionality and limited sample sizes, our framework demonstrates a strong capacity for modeling individualized cognitive trajectories across distinct cognitive domains. By incorporating a dedicated pre-training step, we significantly enhanced predictive performance, surpassing conventional baselines in capturing the heterogeneous patterns of decline in memory, language, executive function, and visuospatial abilities.

We first evaluated the classification capabilities of our models to assess the reliability of the proposed framework. To this end, we compared our approach with benchmark neuroimaging models reported in recent studies on clinical diagnosis prediction (Sharma et al., 2023; Kaur and Sachdeva, 2025). The resulting weighted macro–F1 scores, 0.745 for the 2-year and 0.779 for the 4-year forecasts, place our models within the range of high-performing neuroimaging classifiers. Most misclassifications occurred between CN and MCI groups (Figures 2c–d), which reflects the well-documented clinical ambiguity of these transitional stages (Petersen, 2004).

Furthermore, although the framework was not specifically optimized for conversion prediction, it achieved robust performance in identifying converters within four years (AUC = 0.909), consistent with current state-of-the-art results (Ansart et al., 2021; Sharma et al., 2023; Lee et al., 2024; Yang et al., 2025). Performance at two years (AUC = 0.772) was comparatively lower, likely reflecting diagnostic uncertainty and the limited number of conversion events in brief follow-up intervals (Farias et al., 2009; Garćıa-Gutiérrez et al., 2024b). These findings underscore the robustness of our multimodal modeling approach and align with previously reported AUCs ranging from 0.78 to 0.96 in longitudinal AD studies (Abrol et al., 2020; Wang et al., 2024; Dao et al., 2025).

In these diagnostic classification problems, the superior performance observed at the four-year horizon likely stems from intrinsic features of AD progression. Longer follow-up periods mitigate diagnostic noise and class imbalance, allowing disease-related patterns to manifest more clearly, whereas short-term variability is often influenced by individual factors such as cognitive reserve and comorbidities (Farias et al., 2009; Duara and Barker, 2022; Abdelnour et al., 2022). Consequently, extended observation windows provide more stable, pathology-driven predictions.

While these diagnostic results support the validity of our framework, the primary contribution of this study lies in advancing domain-specific modeling of cognitive decline rather than establishing new diagnostic benchmarks. Recent evidence suggests that conversion-prediction models may have reached a performance plateau, indicating that future progress toward clinical translation will depend more on expanding sample sizes, improving cross-cohort generalization, and enhancing model reliability (El-Sappagh et al., 2023). Within this context, domain-specific modeling of cognitive decline remains comparatively underexplored, despite its critical relevance for early intervention and clinical decision-making (Abdelnour et al., 2022; Hampel et al., 2022; Devanarayan et al., 2025).

Accordingly, our findings demonstrate that patterns of cognitive decline can be inferred from neuroimaging data, particularly through spatial features extracted from brain images. Although the models explained a moderate proportion of the variance in cognitive decline (29.4%–36.0%), the predicted trajectories were well correlated with the observed outcomes (correlation > 0.55). In the binary classification setting, the models also exhibited strong performance in identifying individuals whose decline deviated from normative aging trajectories. Notably, AUC values exceeded 0.83 in the memory, language, and executive function domains.

Visuospatial abilities constituted the main exception to these trends. In this domain, performance dropped markedly when cognitive decline was binarized, despite regression models achieving errors comparable to the other domains. This discrepancy is likely driven by the high sensitivity of the binary label to the normative cut-off, because annual change rates cluster near the threshold and small score fluctuations can flip individuals between stable and decliner categories, adding label noise. In addition, the visuospatial composite relies on tests with limited score range and pronounced ceiling and floor effects, particularly among CN and MCI populations (Kueper et al., 2018). This further increases the risk that subtle variations are misclassified as categorical change rather than genuine trajectory shifts. Future work should incorporate more fine-grained visuospatial measures and larger normative samples to reduce threshold sensitivity and stabilize qualitative predictions.

To date, few studies based on multimodal neuroimaging have modeled cognitive decline. Moreover, most existing works rely on tabular data and are typically limited to neuropsychological assessments of global cognition, such as the MMSE (Zhang et al., 2012; Huang et al., 2016; Franzmeier et al., 2020; Lei et al., 2020; Dansson et al., 2021; Lei et al., 2022; Maheux et al., 2023; Devanarayan et al., 2024). For example, in Franzmeier et al. (2020), the authors used MRI, FDG, and amyloid-PET data to predict cognitive decline rates—similar to those defined in this study—explaining approximately 25% of the variance in global cognition and memory. Similarly, in studies such as Dansson et al. (2021) and Devanarayan et al. (2024), cognitive decline was modeled in terms of MMSE score changes at two and four years, achieving R² values of 0.325 and 0.228, respectively. In Zhang et al. (2012), correlation coefficients of 0.697 and 0.739 were reported for MMSE and ADAS-Cog. Notably, one of the studies most closely related to our work modeled cognitive decline across the domains of memory, language, executive function, and visuospatial abilities, yielding correlation values between 0.42 and 0.50, lower than those obtained in the present study (Moradi et al., 2025).

In addition, several studies have focused on modeling future cognitive scores—i.e., point estimates of the expected value at a single time point—for tests such as the ADAS-Cog, CDR, or MMSE, usually reporting R² values above 0.65 (Huang et al., 2016; Lei et al., 2020, 2022; Maheux et al., 2023). However, this point-estimate approach presents limitations when assessing the true predictive capacity of the models, as baseline scores already account for a substantial proportion of the variance. As a result, the reported explained variance may be misleading. For instance, in groups where cognitive scores remain relatively stable over time, a naive strategy that simply reproduces baseline scores can yield high correlations with future outcomes. This highlights one of the key strengths of the approach presented here, which focuses instead on modeling general trends in cognitive trajectories rather than static future scores.

Our analysis also yielded several technical insights. First, models based on CNNs and GNNs consistently outperformed alternative approaches. Incorporating spatial features was particularly advantageous for complex tasks such as cognitive decline prediction and surpassed the performance of the tabular AI models that we implemented. These gains validate the extra implementation effort required to ingest volumetric or surface-based inputs and suggest that future neuro-AI pipelines should treat spatial encoding as a core design criterion rather than an optional add-on.

Second, although CNNs achieved the highest overall performance, their advantage over GNNs was often marginal, indicating that the increased memory demands of CNNs may not always be justified. In scenarios where hospitals deploy models on embedded hardware, or where cloud inference is billed by runtime and memory usage, the difference in resource usage between GNNs and CNNs may outweigh the slight accuracy advantage of CNNs, making GNNs the more pragmatic choice.

Finally, the proposed pre-training strategy consistently enhanced performance across tasks. Moreover, the learning curves suggest that larger pre-training datasets could lead to further improvements, which offer promising opportunities for model development in settings with limited data.

Within this context, the present study represents a substantial advancement over previous work by modeling cognitive decline across specific domains rather than relying solely on global cognition scores. This design enables a more granular characterization of neurodegenerative trajectories, an aspect of considerable clinical relevance.

Methodologically, our work extends previous multimodal fusion approaches (Zhang et al., 2023; Yu et al., 2024; Abdelaziz et al., 2025) by (i) shifting the focus from purely diagnostic endpoints to domain-specific cognitive trajectories, (ii) integrating several imaging modalities together with rich neuropsychological profiles within a unified comparative framework across FFN, CNN, and GNN backbones, and (iii) introducing a modality-specific multi-task pre-training stage that produces trajectory-aware embeddings for subsequent late fusion.

From an applied standpoint, accurately predicting cognitive decline in AD yields several clinically significant benefits. First, it enables genuinely personalized, dynamic therapy, since clinicians could adjust pharmacological or behavioral interventions in real time as the model detects accelerating progression (Jutten et al., 2021). Second, reliable forecasts support advance-care planning, giving patients and families evidence-based timelines for financial, legal, and caregiving decisions. Third, at the operational level, trajectory prediction helps healthcare systems triage high-risk individuals for closer surveillance and allocate scarce resources, such as specialist memory-clinic appointments, more efficiently (Alzheimer's Association, 2024).

In the context of drug development, prognostic models could facilitate tighter enrichment of clinical-trial cohorts and more informed endpoint selection. For example, endpoints could be tailored to the specific cognitive domains most likely to decline within a given cohort, thereby lowering required sample sizes and increasing power to detect disease-modifying effects (Jutten et al., 2021; Duara and Barker, 2022). Ultimately, domain-level prediction of cognitive decline may be particularly valuable for informing personalized cognitive therapies that target the most at-risk cognitive functions. Collectively, these capabilities have the potential to advance both everyday clinical care and translational AD research.

Despite these promising results, the study has several limitations. Most notably, the predictive performance for cognitive decline remains moderate, which reflects the complexity of neurodegenerative processes. These processes are shaped by multifactorial and stochastic interactions among genetic, environmental, and lifestyle factors, resulting in substantial interindividual variability that is challenging to model with limited sample sizes (Alzheimer's Association, 2024). To improve model accuracy, increasing cohort sizes and incorporating complementary data sources, such as genetic profiles or plasma biomarkers, will be essential. In this context, federated learning offers a scalable and privacy-preserving approach to integrate multi-site data and enhance model generalizability (Rieke et al., 2020).

From a methodological perspective, our comparative non-deep learning baseline was an RF model, and we did not include other non-deep multimodal baselines or hybrid ensembles. Given that the available literature on this task is limited and predominantly deep-learning oriented (Kaur and Sachdeva, 2025), a more comprehensive benchmarking analysis should be undertaken in future work to better contextualize the state of the art in domain-level cognitive-decline prediction.

Another widely recognized limitation that applies to this study, and to much of the current literature, is the difficulty of generalizing models across cohorts. Our analyses were restricted to the ADNI dataset, limiting the findings to this specific population (Samper-González et al., 2017). This constraint may reduce the clinical applicability of the results, as research cohorts such as ADNI often differ from real-world clinical populations in demographic composition, disease spectrum, and imaging protocols. This issue is particularly relevant given the known variability in the demographics and clinical presentation of AD. Therefore, external validation on independent, multi-site cohorts (e.g., AIBL, OASIS) and, ideally, prospective clinical studies conducted in real-world settings will be essential to establish robustness, generalizability, and clinical utility. Future work should prioritize evaluating the proposed framework in such diverse cohorts as a necessary step toward translational deployment.

Future studies should also investigate whether incorporating longitudinal information, such as serial neuroimaging scans or repeated cognitive assessments, provides meaningful improvements in predictive performance. However, acquiring such data is logistically challenging, because repeated MRI or PET sessions are costly, time-consuming, and consequently rare in routine clinical practice. Rigorous analyses will therefore be required to determine whether the incremental predictive gains offered by longitudinal data justify the substantial logistical and financial burden associated with their acquisition.

Finally, future research should assess the biological validity of model predictions using explainable AI (XAI) techniques tailored to neuroimaging. Demonstrating that post hoc attribution methods (e.g., saliency maps, attention weights, or counterfactual lesioning) assign importance to neuroanatomically plausible regions would substantially strengthen the claim that the model leverages disease-relevant signals (Arrieta et al., 2020; Bloch et al., 2024). While such interpretability analyses were beyond the scope of the present study, which primarily focused on establishing predictive performance, they constitute a critical avenue for future work. Supporting the feasibility of this approach, previous work from our group has shown that XAI pipelines in related neuroimaging contexts can produce stable attributions in canonical disease-relevant circuits (Garćıa-Gutiérrez et al., 2024b). Conducting these analyses is essential, as they can help clinicians understand the rationale behind model predictions and thereby foster clinical trust, improve patient safety, and facilitate eventual regulatory evaluation.

In conclusion, we developed AI-based models using multimodal neuroimaging data and systematically compared alternative information-representation strategies to predict domain-specific cognitive decline in Alzheimer's disease. Theoretically, our findings show that spatially encoded multimodal features can capture individualized trajectories across memory, language, executive, and visuospatial domains, moving beyond conventional global cognition scores and point-estimate forecasting. Practically, our approach provides a scalable tool for stratifying patients according to their risk of accelerated decline in specific cognitive domains, thereby supporting personalized intervention planning, optimized resource allocation, and the enrichment and endpoint selection of clinical trials. By maximizing the use of all available neuroimaging information to generate domain-level trajectory predictions, this study offers both a conceptual and methodological advance in multimodal modeling and contributes to the development of clinically meaningful prognostic tools for neurodegenerative diseases.