Currently, more than 55 million people worldwide are living with dementia, the seventh leading cause of death and one of the primary causes of disability and dependency among older adults (World Health Organization [WHO], 2024). Alzheimer’s disease (AD), a serious brain disorder caused by damage to neurons, is the main cause of dementia (Alzheimers Dementia, 2024; Taylor et al., 2017; Petersen et al., 2018; Rabinovici et al., 2019). As the global population ages, the number of individuals affected by AD is expected to rise. In 2020, it was estimated that 6.07 million older adults in the United States had clinical AD and this number was projected to increase by 18% to 7.16 million by 2025 and 128% to 13.85 million by 2060 (Rajan et al., 2021). Such an increase will impose a heavy financial burden on the society. The estimated global cost of caring for individuals with dementia is expected to rise to approximately $2 trillion by 2030 (World Health Organization [WHO], 2024).

Traditional clinical medications for AD, such as donepezil, rivastigmine, galantamine, and memantine, may cause side effects like arrhythmia, dizziness, headache, and nausea, and they do not alter the disease’s progression (Nham et al., 2024; Henstra et al., 2022; Maelicke et al., 2010; Tan et al., 2014). The efficacy of non-pharmacological interventions, such as music and reminiscence therapy, remains uncertain. New drugs like lecanemab aim to change the underlying biology of AD by targeting and slowing cognitive and functional decline. However, lecanemab is only suitable for patients with early AD, and the safety and efficacy of long-term use have not been observed yet (Terao and Kodama, 2024, 2024; van Dyck et al., 2023). Thus, it is particularly important to diagnose AD as early as possible. Mild cognitive impairment (MCI) represents a cognitive function state that lies between normal aging and dementia. Studies have shown that patients with MCI have a significantly increased risk of progressing to AD compared with the general population (Petersen et al., 2018; Rabinovici et al., 2019).

Many methods have been applied to the auxiliary diagnosis of AD, including the Clinical Dementia Rating Scale (CDR), magnetic resonance imaging (MRI), cerebrospinal fluid (CSF) biomarkers, and positron emission tomography (PET). However, there is still a high misdiagnosis rate in the early diagnosis of dementia in clinical practice (Beach et al., 2012; Gianattasio et al., 2019; Hunter et al., 2015; Nguyen et al., 2022). The underdiagnosis of MCI is even more pronounced, with only 8% of older adults with MCI in the United States receiving a diagnosis (Liu et al., 2024).

Artificial intelligence (AI) is the simulation of human intelligence in machines to make them think and learn like humans. It includes a variety of types, such as machine learning (ML), deep learning (DL), computer vision, and expert systems. AI has found widespread applications in medicine. For example, the application of AI in imaging diagnosis has progressed notably in the identification of tumors, such as prostate and breast cancer (Jiang et al., 2024; Molière et al., 2024). Furthermore, a great number of studies have been conducted in the realm of cognitive function diagnostics (Hussain and Borah, 2024; Basaia et al., 2019; Myszczynska et al., 2020). We conducted a literature search in the Web of Science (WoS) Core Collection and got 22,646 records, highlighting the vibrancy of research in this field.

Highly cited articles can serve as important indicators of the research status and trends within a specific field, providing valuable insights that can significantly enhance clinical practice. To effectively summarize key information and assist future researchers in gaining a clearer understanding of the characteristics of published works in this domain, we intend to analyze the top 100 cited articles on the combination of AI and AD/MCI using bibliometric methods.

We conducted a bibliometric analysis of the top 100 most cited articles on AI in AD/MCI. The main findings include: (1) diagnostic and predictive studies are the most frequent research types, with MRI being the most commonly used clinical data source; (2) SVM and CNN are the most commonly applied AI methods; (3) the United States and China are the leading countries contributing to AI-driven AD/MCI research; (4) high-impact journals such as Neuroimage, Frontiers in Neuroscience, and Brain are key publishers in this field. These results highlight the central role of AI in advancing the diagnosis and prediction of AD and MCI, emphasizing the growing international contribution to this important area of research. We compiled the number of articles published by authors and analyzed the co-citation situation of each author. Shen Dinggang has the highest citation count, publishing 11 articles among the top 100. The author’s research has played a key role in advancing ML technologies for the diagnosis of AD and MCI, with a clear focus on DL, multimodal integration, and early disease detection. In the early stages, Shen’s research applied pattern classification methods to MRI scans to detect prodromal AD in MCI patients, achieving a diagnostic accuracy of up to 90%, and highlighting the diagnostic potential based on MRI (Davatzikos et al., 2008). Around 2011, the research expanded to methods involving more modalities, introducing multi-task learning to predict various clinical variables and AD classification by integrating MRI, FDG-PET, and CSF data (Zhang et al., 2012). Between 2014 and 2017, the author increasingly incorporated DL techniques, such as the Deep Boltzmann Machine (DBM) and Stacked Autoencoder (SAE), to create hierarchical feature representations from neuroimaging data (Suk et al., 2014; Cheng et al., 2015; Li et al., 2015; Suk et al., 2015; Suk et al., 2016). These models could automatically extract high-level features from MRI and PET scans, outperforming traditional hand-crafted feature-based approaches in classification tasks. During the period from 2017 to 2019, the author conducted research on multi-task multi-channel learning for brain disease classification and clinical score prediction using neuroimaging data along with demographic or genetic information, aiming to achieve higher accuracy (Liu et al., 2019; Zhou et al., 2019). Shen’s recent work focused on layered fully convolutional networks to locate brain atrophy in MRI scans, enhancing diagnostic accuracy and tracking disease progression (Lian et al., 2020). We compare two of Shen’s influential studies to reflect the methodological evolution in his research trajectory. Zhang et al. (2012) applied multi-modal multi-task (M3T) learning, integrating MRI, FDG-PET, and CSF features using a multi-task feature selection followed by a multi-modal SVM for regression and classification (Zhang et al., 2012). This approach effectively leveraged complementary information but relied on hand-crafted features and a relatively fixed processing pipeline. In contrast, Zhou et al. (2019) proposed a stage-wise deep neural network, first learning modality-specific latent representations, then progressively combining these features across modalities, and finally fusing them for diagnostic prediction in an end-to-end deep learning framework (Zhou et al., 2019). This evolution from feature-level fusion to hierarchical, data-driven fusion networks marks a significant shift toward more flexible, scalable models capable of better handling multimodal, heterogeneous datasets and incomplete data. The co-occurrence analysis of all keywords revealed emerging research trends and hotspots in this field. Both SVM and CNN are among the top 10 keywords in ML methods. In these studies, the SVM was used the most frequently, appearing in 25 articles. SVM is an ML algorithm primarily used to classify data into multiple predefined categories. By finding an optimal separating hyperplane in a high dimensional space, SVM can classify samples based on feature data; for instance, in neuroimaging research, it can group patients and healthy controls (Gao et al., 2022; Gaonkar et al., 2015; Sacchet et al., 2015). This process includes feature extraction and selection, training and testing classifiers, and evaluating their performance to optimize the algorithm’s predictive capability (Orrù et al., 2012). SVM is particularly well-suited for handling complex classification problems and can efficiently map data using kernel methods, enhancing classification accuracy in high-dimensional data.

A study titled “Using SVM to identify imaging biomarkers of neurological and psychiatric disease: a critical review” (Orrù et al., 2012), which describes the use of SVM to analyze MRI and PET for early diagnosis of MCI and AD, is the most cited article. Of the top 100 cited articles, 19 used SVM to analyze MRI, and 5 analyzed PET. This seems to indicate that the use of SVM methods to analyze patients’ neuroimaging data for early diagnosis of MCI and AD is a major research hotspot. It is worth noting that another paper published in 2020 used SVM methods in the analysis of EEG (Ieracitano et al., 2020). EEG has been found to have the potential to assist in the diagnosis or prediction of a variety of diseases such as AD, perioperative neurocognitive dysfunction (PND), MCI, subjective cognitive decline (SCD), and depression (Boby and Veerasingam, 2025; Ouchani et al., 2021; Deiner et al., 2015; Engedal et al., 2020). Moreover, AI methods facilitate a more convenient and accurate analysis of EEG results (Huang et al., 2012; Hosseini et al., 2021). Utilizing methods such as SVM to analyze patients’ EEG for diagnosing AD or MCI may be a future research direction. Additionally, PET and CSF have long been used to detect molecular and biochemical characteristics related to AD, but their usage frequency is still significantly lower than that of MRI in the current highly cited literature. Sufficient attention should be paid to these relatively understudied modalities and their integration with MRI information is expected to further enhance the sensitivity and specificity of the model for disease progression. For instance, MRI can provide structural or functional changes, while PET and CSF reflect pathological processes such as amyloid and Tau; EEG provides highly time-resolved functional dynamics. The strategy of multimodal integration may become an important approach for promoting the clinical application of AI methods. The frequency of CNN usage is also quite high, with 24 articles employing this method. CNN is a type of DL model particularly suited for processing image data. Compared to traditional methods, CNNs are better at capturing subtle changes in images, thereby enhancing the capability for early diagnosis. In the studies we included, the application of CNN methods for examining AD achieved a sensitivity of 97.96% and a specificity of 97.35% (Wang et al., 2018). A study conducted in 2022 reported a 100% accuracy in classifying AD and Normal Control (NC) subjects using CNN methods (Habuza et al., 2022).

To further contextualize our findings, we compared our bibliometric results with those of a recent comprehensive analysis by Song et al. (2025), which examined 2,316 AI-related AD papers from 2004 to 2023. In line with our findings, both studies identified the United States and China as the leading contributors, but our focus on the top 100 cited articles revealed a more pronounced dominance of the United States (54% of publications vs. 29% in Song et al.), suggesting that the most influential research is concentrated in a few top-tier American institutions such as the University of North Carolina at Chapel Hill and the University of Pennsylvania. In contrast, Song et al. reported more balanced global contributions, with China showing rapid growth but lower citation impact. Journal patterns also differed: our top-cited articles were mainly published in high-impact Q1 journals (e.g., NeuroImage, Brain), whereas Song et al. found higher productivity in broader-scope journals such as Journal of Alzheimer’s Disease and Frontiers in Aging Neuroscience. This implies that high-impact studies tend to cluster in neuroimaging and methodological domains, while the broader field encompasses more clinical and interdisciplinary work. Methodologically, both analyses highlighted DL and ML approaches, yet our dataset showed a greater presence of established algorithms such as SVM and CNN among highly cited works, reflecting their enduring influence and reliability. These comparisons suggest that the top-cited literature represents a distinct high-impact subset characterized by concentration in leading institutions, Q1 journals, and robust AI methodologies, complementing the broader bibliometric patterns of the entire field. In addition, we analyzed the IF, JCR Category Quartile, JCR category, and category of these articles, hoping to help future researchers better submit their papers. Of the 100 articles, 68 were classified within the JCR Q1 zone, which also reflects the importance of this research area. The novelty of our study lies in its exclusive focus on highly influential works, covering both AD and MCI, and its integration of AI methods with clinical data modalities. This multidimensional approach provides insights into the methodological and thematic features that drive high scientific impact in AI-assisted dementia research.

This study is not without its limitations. Despite our endeavors to employ a diverse array of search terms, it is possible that key studies have been overlooked. The ever-evolving nature of citation counts for articles means that recently published papers may not have had adequate time to amass significant citations, thereby excluding them from our analyses. Future updates could incorporate normalized metrics to better capture emerging high-impact studies. Moreover, our article retrieval was not exhaustive across all databases, inevitably leading to some omissions in the collected literature. Additionally, in this study, AI methods are classified into traditional ML and DL categories. However, the delineation is not always clear-cut, as some techniques can straddle multiple categories. For example, SVM, a conventional ML approach, is occasionally incorporated into DL frameworks, blurring the lines of classification. Finally, our analysis summarized AI method frequencies in aggregate but did not classify these methods by publication year, and future work incorporating year-based stratification may help clarify how approaches such as CNN and SVM have evolved over time.

We analyzed the top 100 cited articles on AI in AD and MCI by examining authors, journals, countries, institutions, keywords, and AI methods. According to these analyses, we find that the application of AI methods to diagnose AD or MCI is the primary research direction among these 100 articles, and MRI data is the most common type of clinical specimen collected for such studies. Moreover, SVM and CNN are the most commonly used AI methods.