A systematic search was carried out using libraries such as PubMed (MEDLINE), Scopus, and Web of Science. The search followed the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analysis) 2020 guidelines to maintain transparency, authenticity, and completeness of details of reporting [41]. The PRISMA checklist of this paper can be found in Multimedia Appendix 1. This search was carried out over the last 15 years and was centered on published studies specific to early-stage AD detection and classification (between January 2010 and March 2024). Limiting our review to the last 15 years of publication allowed us to focus on papers reflective of current trends in research.

The search strategy used the following keywords: “Alzheimer’s disease,” “machine learning,” “early detection,” “diagnostic accuracy,” “diagnosis,” “predictive models,” “biomarkers,” “deep learning,” “diagnostic accuracy,” “feature selection,” “AD biomarkers,” and “ML models.” The search strategy was (“machine learning” OR “artificial intelligence” OR “classification”) AND “Alzheimer’s disease” AND “MRI” AND “diagnosis” AND “classification.”

Full-text papers in the English language were considered. We have included in this study only published papers in peer-reviewed journals. The majority of the papers analyzed were centered on MRI data combined with ML models in AD diagnosis. Selected studies included patients diagnosed with early-stage AD and healthy controls. The papers published with a title or abstract containing at least 1 abovementioned keyword were considered for inclusion.

Papers written in a language other than English were excluded. We excluded studies that were not specifically conducted in the context of AD diagnosis using MRI and were not primarily focused on ML models. Papers published before 2010 were not considered. Studies in which ML in MRI was not explicitly linked to clinical diagnosis, medical training, or initiatives to improve AD diagnosis were excluded. This review excluded studies using PET and CT scans because the primary focus was on ML in MRI, which is specifically linked to clinical diagnosis, medical training, and initiatives to enhance AD diagnosis. The selection process excluded review papers, conference proceedings, and gray literature reports.

Multiple stages were involved in the paper selection process. The results of the systematic search were documented in a spreadsheet using the above strategy. The selected papers were equally distributed among the authors, and each paper was screened by examining titles and abstracts to identify potentially relevant publications. The selected papers were then reviewed comprehensively according to predefined inclusion and exclusion criteria in the subsequent phase. To facilitate synthesis, relevant information was extracted and organized in a tabular format, covering study design, datasets, performance metrics, model validation, and feature selection. As a result, a summary of each study’s main findings to discern trends, patterns, and common themes was done.

The Newcastle-Ottawa Scale [42] was used to assess the study quality based on different factors such as selection, comparability, and outcome, providing a structured approach to gauge the risk of bias. In terms of quality, scores ranged from very poor (0-3) to moderate (4-6) to excellent (7-9). The papers meeting the score (Newcastle-Ottawa Scale≥7) were only considered for final review. Two authors (GB and NC) independently assessed the quality, and any discrepancies were resolved through discussion or consultation with a third author (FA).

The statistical tests Egger regression [43] and Begg rank correlation [44] were used to address the potential bias of publications. To assess the strength of our findings against potential biases or variations in study characteristics, sensitivity analyses were performed. Lower methodological quality or different study designs were excluded. To identify the effect size measures and quantify the strength or magnitude of the relationship between variables or the magnitude of differences between AD groups, we applied the “PLOGIT” function to the logit transformation of the proportion [45]. The logit transformation is commonly used when dealing with proportions or probabilities, especially when they are bounded between 0 and 1. An inverse variance method has been applied that specifies the method for pooling effect sizes. There were 2 types of models considered in the meta-analysis: fixed effects and random effects. Using the fixed effects model when we observed a low level of heterogeneity, the test is not statistically significant.

The random effects model (REM) was considered for the heterogeneity test with statistical significance [46]. By calculating T², the amount of heterogeneity between the true effect sizes of different studies was quantified. An estimation method using a restricted maximum likelihood estimator that maximizes the likelihood function while accounting for other parameters of the model was used [47]. I² and Cochran Q statistic tests were conducted to assess the heterogeneity among the effect sizes of individual studies [48,49]. The measures of heterogeneity (T² and I²) indicate the variability in AD prevalence estimates across the studies [50].

The prevalence of patients with AD across different subgroups within the overall population was also investigated. Subgroup analysis enables the identification of factors that can influence prevalence estimates and provide insight into the sources of heterogeneity [51]. A subgroup-specific meta-analysis model was used to calculate the pooled prevalence estimates for each subgroup, followed by a comparison of the prevalence estimates across subgroups to assess whether there were any significant differences. Data were subgrouped into 4 category-based AD classifications namely, 2-group classification, 3-group classification, 4-group classification, and 6-group classification. The 2-group classification involved individuals either without dementia (nondemented, ND) or with dementia (demented, AD). The 3-group classification includes CN, MCI, and AD. The 4-group classification comprises ND, mildly demented (MD), moderately demented (MoD), and AD. Meanwhile, the 6-group classification involves CN, SMC, EMCI, MCI, LMCI, and AD. Each subgroup data was recorded separately into a Microsoft Excel spreadsheet, which was further supplied as input to R software (version 4.3.3; R Foundation for Statistical Computing). For prevalence and summary meta-analysis, we used the “meta prop” functions available in the meta package.

During the identification phase, 5049 records were obtained from 3 major scientific databases using the given search strategy. Following the removal of duplicates (n=2355) and the assessment of ineligibility using tools (n=218), 2446 records were included in the screening stage. The inclusion and exclusion criteria determined that 2037 records were ineligible. We further screened 409 records, with 134 being excluded due to lack of full-text availability. In total, 251 records from the remaining 275 were excluded due to low-quality scores and publication bias. A total of 24 papers were included in the final analysis. Details on the procedures for selecting papers are summarized in Figure 1.

The data collected for this study were collected from various geographical locations and may have included memory clinics and neurology departments, suggesting a focus on cognitive impairment and related conditions. Table 1 displays the distribution of AD imaging sample data along with data sources.

The analyzed studies collected image data from various sources such as Alzheimer’s Disease Neuroimaging Initiative (ADNI) [76], Open Access Series of Imaging Studies (OASIS) [77], Australian Imaging Biomarkers and Lifestyle Flagship Study of Ageing [78], and public domains like Kaggle [79]. ADNI datasets were used more often for image collection [52,54,56-59,61,64-68,72-75,80]. The purpose of ADNI is to develop biomarkers for early detection and AD tracks through a multicenter study involving clinical imaging, genetics, and biochemistry. The studies that use ADNI datasets aim to detect AD at its prime stage. One study jointly applied 2 image datasets from ADNI and Australian Imaging Biomarkers and Lifestyle Flagship Study of Ageing [54].

OASIS brains aim to make it possible for anyone to access neuroimaging datasets of the brain through an initiative known as Open Access to Neuroimaging Datasets. Through this project, researchers can access and use a variety of brain imaging data for free. This resource assists neuroscience researchers in advancing their research by providing a comprehensive collection of brain imaging datasets. Cross-sectional OASIS 1 data were used by researchers for hypothesis-driven analysis, neuroanatomical atlases, and segmentation algorithms [69]. In another study, OASIS-3 was integrated with longitudinal neuroimaging, clinical, cognitive, and biomarker data [55]. The use of public datasets or participation in Kaggle competitions related to AD research helps as a platform for data science competitions and datasets [70]. Three studies collected data from 3 hospitals in China [53,63,71]. The findings indicate that a diverse dataset from multiple sources, such as clinical settings and publicly available datasets, could provide a comprehensive basis for AD research and analysis.

There was no evidence of publication bias with Eggers (P=.49) or Begg (P=.38) tests. Figures 2-5 present the forest plot with the prevalences of participants with AD for 2, 3, 4, and 6 AD stage subgroups, respectively. Six studies with 1562 participants were identified among disease diagnoses with 2 stages including CN and AD [53,61,63,69,72,75]. The overall pooled prevalence of the REM reported 49.28% (95% CI 46.12%-52.45%; I²=15%; P=.32). Studies do not differ significantly in their estimates of prevalence, and the test of heterogeneity does not reveal substantial differences between them. Seven studies were identified with a total sample of 17,588 patients with AD with 3-stage AD classification including CN, MCI, and AD [56,57,62,65,67,71,73]. The overall prevalence of AD diagnosis is estimated at 29.75% (95% CI 25.11%-34.84%; I²=97%; P<.001). Each study provides an estimate of the AD prevalence among their respective populations with 95% CI. For example, Hazarika et al [56] found AD prevalence at 33.33% (95% CI 27%-40.15%). This indicates that if we were to combine the results of all the studies, this would be the estimated AD prevalence. I²=97% indicates that a large proportion of the total variation in prevalence estimates is due to true differences between study populations rather than random error. The significant P value (<.01) for the test of heterogeneity indicates that there is substantial variability in AD diagnostic prevalence estimates among the studies.

Five studies with 14,839 participants were included for the meta-analysis of 4-stage AD classifications as ND, MoD, MD, and overt AD [55,59,64,66,70]. This systematic review included 7 studies, but we excluded 2 studies [58,60] because they used the same dataset with 6400 ADNI participants. Overall prevalence estimation with REMs is 13.13% (95% CI 3.75%-36.96%; I²=99%; P<.001). There is significant heterogeneity in the studies based on the high I² and significant P value and a considerable variation in the prevalence of AD across these studies, according to these estimates. Different research studies have found prevalence estimates ranging from 1% [55] to 30.43% [66]. The CIs indicate the degree of uncertainty in these estimates. As a result of the high degree of heterogeneity observed in the study, the true prevalence of AD may vary significantly between populations and settings. Four studies with 3819 were considered for the calculation of the overall prevalence of AD diagnosis of 6 stages such as CN, SMC, EMCI, MCI, LMCI, and AD [52,54,68,73]. The estimated prevalence for each study is ranging from 16.18% [52] to 50.16% [54]. The overall estimate of prevalence from the REM stands at 23.77% (95% CI 12.22%-41.12%; I²=0.8020; P<.001). One study has a substantially greater estimated proportion of AD prevalence diagnosis than the other studies [54]. Compared to others, it reported the highest prevalence of 50.16% (95% CI 48.45%-51.88%) but does not differ weights (26.3%) significantly from other studies.

Meta-analysis through forest plots provides a comprehensive way of understanding meta-analysis results. It can be argued, however, that forest plots can only display CIs by assuming a fixed significant threshold (P<.05). It causes a replication crisis when hypothesis tests are conducted using P values. Based on P value functions, drapery plots were proposed to resolve this problem [82]. Using a drapery plot, an average effect and a confidence curve can be identified. The x-axis shows the effect size metric, and the y-axis shows the assumed P value. Multimedia Appendix 2 presents the drapery plots. There is a red curve showing the overall REM, which shows the P values for various effect sizes. Compared to the CI of pooled effects, the shaded area represents the prediction range. The prediction range is noticeably wider than the CI for the pooled effect. It indicates that the overall pooled effect does not fully capture the variability or uncertainty across different effect sizes.

In this work, we conducted a systematic review and meta-analysis based on the prevalence of patients with AD among different disease progression stages. For the systematic review, 24 studies were selected, among 22 selected for the meta-analysis. Due to their association with the same dataset of ADNI and similar sample size of patients with AD, these 2 studies avoid bias in the analysis [58,60]. The studies included in this review have explored the ML applications for AD diagnosis and intended to provide an understanding of AD progression, potentially with a focus on biomarker identification.

Different preprocessing techniques used to extract relevant features including cortical thickness [83], hippocampal volume [84], and brain activity patterns [85] from magnetic resonance images associated with AD were examined. According to the research objectives and AD stages being investigated, each study applied specific image preprocessing techniques. The progression of AD has been evaluated across multiple stages in our work. An accuracy range of 69%-75% is achieved with linear mixed-effects models that account for region of interest features with interparticipant variability of hierarchical structures [54]. Using image normalization, 1 study classified AD stages with different labeling with 84.03% accuracy by ensuring consistency in intensity and spatial properties [52,86]. Combining DL models with imaging techniques like MRI and PET has shown that structural and functional changes in the brain associated with AD can be detected [87,88]. Water molecules’ diffusion properties in brain tissue can be measured using diffusivity and kurtosis mapping. The results provided insight into microstructural changes for a maximum accuracy of 96.23% [53]. By conducting magnetic resonance image normalization, the authors proposed an MRI-based DL technique for 99.68% accurate AD detection [55]. Magnetic resonance images were investigated for pixel intensity distributions to detect AD abnormalities [56].

These techniques encompass diverse methodologies ranging from normalization and smoothing to advanced mapping and feature extraction methods [89-91]. Several approaches have demonstrated high accuracy in identifying AD features, including image normalization, histogram-based approaches, and diffusion mapping [92-95]. Techniques like recursive feature elimination and outlier detection showcase promising results, emphasizing the importance of feature selection and data quality assessment in enhancing classification performance [57,63]. A similar study analyzed and segmented different tissue types within MRI scans using unified segmentation. A magnetic resonance image of the brain was segmented simultaneously into different tissue types with 85.12% accuracy [65,96]. KNN-trained data can be used to classify AD with 99.4% accuracy using the generalization method [66]. Moreover, the use of advanced DL architectures such as VGG16 [60] and ConvNet [62] for feature extraction underscores the significance of leveraging sophisticated computational tools in AD research. Augmentation methods, interpolation, and transfer learning also emerge as valuable strategies for improving classification accuracy and robustness [73-75].

By integrating statistical and ML algorithms with preprocessing techniques, AD diagnosis research further enhances its interdisciplinary nature. The CNN-long short-term memory model had an accuracy of 99.92%, followed by the multimodal AD diagnosis framework model with a precision of 96.88%. The accuracy of a customized CNN model was 99.68%, SVM with DKI was 96.23%, XGB was 96.20%, and multilayer perceptron was 99.44%. In addition, DenseNet121, CNN, DenseNet201, random forest, and gradient boosting achieved accuracy levels between 90% and 97%. While some models demonstrated higher accuracy, such as 3D CNN and SVM, others demonstrated lower accuracy, 85.12% and 80.36%, respectively.

Many ML modeling techniques have been explored, including SVM, LR, and DenseNet. Ensemble methods like gradient boosting and Ada boosting have highlighted the importance of aggregating multiple models to improve predictive accuracy and robustness, especially when dealing with complex neurological disorders like AD [69,97]. The identification of specific best-performing models further underscores the importance of optimization of methods and model selection to improve diagnostic accuracy. The use of SVM along with DKI or DenseNet201 in different studies illustrates the researchers’ tailored approach to leveraging each algorithm’s and feature representation’s strengths [98-100]. AD diagnosis is a nuanced process, where the choice of ML model can have a significant impact on model reliability and efficacy.

Data from magnetic resonance images have been analyzed using various ML models and validation techniques. To ensure robustness and generalization, the common technique used is K-fold cross-validation. Additionally, some authors have applied specific DL models along with traditional ML techniques, reflecting the diversity of approaches for modeling and validation [72,73]. Different mechanisms and approaches are used in each of these models to detect AD using magnetic resonance images. We have observed that SVM classifiers are largely used for 2-stage classification such as CN and AD [53,54,67,71]. Similarly, LR classifiers were used in other studies to assess MRI-based AD status interpretation and predictive factors for disease risk assessment. Based on learned discriminative patterns from magnetic resonance images, these models, as well as others mentioned, produce accurate AD detection predictions. Additionally, KNN can be used to identify magnetic resonance images with feature vectors similar to those associated with AD helping to detect patterns.

The meta-analysis shows that there is a great deal of variation between studies when it comes to estimating AD prevalence. The reason for this is probably because the study involved a wide range of diagnostic criteria and populations, not just prevalence rates. The prevalence estimates are diverse due to some studies focusing on specific AD stages while others cover a wider spectrum. The significant P values and I² statistics show that the diagnosis of AD is highly heterogeneous and requires a nuanced understanding of its epidemiology. The challenges associated with synthesizing prevalence data from disparate sources are revealed by this analysis. The prevalence of AD is subject to complex and variable research, which leads to wider CIs in some studies. Even after trying to use REMs to account for this heterogeneity, significant variation persists, suggesting that variables like demographics, study design, and diagnostic methodology may play a significant role. The provision of more reliable estimates requires the adoption of standardized protocols and collaboration in future research efforts, which stresses the importance of rigorous methodology and careful interpretation of results.

There have been a few systematic reviews and meta-analyses about the importance of ML models in AD diagnosis. Table 3 summarizes the comparison between our work and the reviews that have already been published. In our analysis, we concentrated on using ML for AD diagnosis, while other studies were focused on using it for dementia forecasting [101]. In a similar study [102], the authors explored the effectiveness of both ML and DL models in AD diagnosis. In this study, the authors did not examine multistage AD cases but only the binary classification of AD. A single study [103] conducted a meta-analysis based on Wilcoxon signed rank tests and discussed multiple imaging modalities, including MRI, PET, and CSF. Despite this, there is a lack of discussion about feature selection techniques and their potential impact on ML accuracy. A prevalence-based meta-analysis on MRI-centered AD discussions is presented in our study along with an in-depth description of subcategories of AD. Our study stands out because it covers all aspects of ML in AD diagnosis, including imaging modalities and stages of AD. We reviewed and analyzed various imaging modalities, talked about feature selection methods, and delved deeper into AD subcategories in our research.

Data from open-access libraries such as ADNI, Kaggle, and others were used in studies, as evidenced by the analysis of datasets. Prospective validation studies should be carried out in the future to assess the accuracy of ML models for AD diagnosis across diverse populations and clinical settings. The incorporation of multimodal data, including imaging, genetics, and clinical information, into ML models can improve their accuracy and robustness in diagnosing AD and distinguishing it from other brain disorders [89]. To enhance their clinical utility and acceptability, ML models must be interpretable and explainable. It may be possible to use these models to predict the onset and AD progression based on longitudinal studies that track individuals over time [14,101]. Future research must incorporate ML models into diagnostic workflows and assess their influence on patient outcomes and health care delivery.

Despite its comprehensiveness, this study is characterized by some shortcomings. The availability and quality of data are essential for the effectiveness of ML approaches. The outcome of the meta-analysis may have been influenced by the limitations in access to complete datasets with different levels of quality. The potential for publication bias, in which studies with positive findings are more likely to be published, may lead to an overestimation of the effectiveness of ML approaches for diagnosing AD. The included studies may have experienced heterogeneity due to variations in study designs, patient populations, imaging modalities, and ML algorithms, making it difficult to draw definitive conclusions. Despite our best efforts to conduct a thorough review, some relevant studies may have been mistakenly excluded, potentially creating gaps in the analysis. The generalizability of ML models for AD diagnosis may be limited by their development and validation on specific datasets.

A summary comparison of current literature on ML approaches in AD diagnosis, along with a systematic review and meta-analysis, helps to understand the prevalence of disease at different stages. Our analysis of 24 relevant papers shows a significant difference in AD prevalence estimates, as individuals progress from CN to MCI and ultimately to overt AD. We observed a pooled prevalence of 49.28% during the CN to AD transition. This was followed by 29.75% for CN, MCI, and AD, 13.13% for CN, MoD, MD, and AD, and 23.75% for CN, SMC, EMCI, MCI, LMCI, and AD. Our analysis reveals the importance of adjusting diagnostic and management strategies to minimize the impact of demographic and setting characteristics on AD prevalence estimates. Due to the heterogeneity observed across studies, it is necessary to consider various factors to accurately estimate the prevalence of AD. Our study is different from other studies by comparing it to existing systematic reviews and meta-analyses, which provide an original contribution to the topic under evaluation. Unlike previous studies that have focused on imaging modalities and AD stages, our study has comprehensively analyzed ML in AD diagnosis. Multiple imaging modalities were reviewed and analyzed, feature selection techniques were discussed, and AD subcategories were explored, focusing particularly on MRIs. Although none of the biomarkers currently available can provide a precise diagnosis of AD, using ML approaches to identify prevalence patterns across disease stages will lead to progress in AD diagnosis.