This study meticulously adheres to the Cochrane Handbook of Systematic Reviews and Interventions and aligns with the preferred reporting items for systematic reviews and meta-analyses (PRISMA) statement, including the PRISMA extension for network meta-analysis (PRISMA-NMA). Moreover, it was prospectively registered in PROSPERO, ensuring transparency and accountability in the review process.

Relevant studies were meticulously sought through thorough searches across multiple databases including PUBMED, Web of Science, Cochrane, Medline, and EMBASE, with the search concluding on May 27th, 2025. The search strategy was tailored to each database’s default options, employing a meticulously crafted combination of keywords and Boolean operators. Literature published from 2004 to 2021 were included in the search, and the following keywords and terms were used: “Machine learning” or “Deep learning” or “Alzheimer’s Disease” and “Network meta-analysis.”

The research question was designed according to the Population/Intervention /Comparison/Outcome(s) criteria (PICO):

Prospective cohort studies were considered eligible for inclusion if they examined the accuracy of AI algorithm methods as an outcome. Disagreements regarding eligibility were resolved through mutual consensus. We only included prospective cohort studies to prevent common biases like recall and selection biases inherent in case–control studies.

After database search, the results were imported to a reference manager software Endnote (Mendeley desktop v1.17.4, Elsevier, New York, New York) to remove duplicates automatically. Two independent researchers used the previously described eligibility criteria to screen the records by title and abstract. Eleven full-text articles were assessed for eligibility (Plant et al., 2010; Escudero et al., 2011; Diciotti et al., 2012; Skolariki et al., 2020; Moradi et al., 2015; Lama et al., 2017; Cohen et al., 2019; Pan et al., 2020; Liu et al., 2018; Zhang et al., 2019). Data were extracted by the same two independent reviewers. Standardized data extraction was performed in order to select the following data: study characteristics (author and year), study design, sample characteristics (machine learning/deep learning method, patients’ number, scanning type of MRI, speciality, accuracy and sensitivity). If the values were not completely reported in the original article, the authors were contacted via email to request the data. Articles with a lack of data not receiving an answer from the authors were excluded from the meta-analysis. When the articles reported the results as median and quartile values, the mean and standard deviation were estimated according to previously described methods. In case of disagreements about the eligibility of a study or data extraction, the authors consulted a third author to reach a common decision through consensus. The primary outcome measures considered for data extraction were accuracy of detecting AD and speciality as well as sensitivity of the detection. In order to be included in the network model analysis (NMA), the outcome measure had to have been assessed by at least three articles investigating the same intervention over a similar timescale (immediate, short term, medium term or long term).

For each included study we extracted information on the diagnostic task (AD vs. healthy control; AD vs. MCI; MCI conversion), MRI modality (T1-weighted alone or multimodal sequences), dataset origin (public dataset such as ADNI vs. local cohort), and validation strategy (k-fold cross-validation, leave-one-out, or hold-out test set). We summarized these characteristics in an expanded Table 1. To account for between-study variability we used a random-effects network meta-analysis model.

For the statistical analysis, the R Ver. 4.0.5 program (R Foundation for Statistical Computing, Institute for Statistics and Mathematics, Welthandelsplatz 1, 1020 Vienna, Austria) was used with the netmeta and dmetar packages. A frequentist NMA with DerSimonian–Laird estimator design was carried out assuming a random effects model for the network analysis. A network diagram was created in which each node represents an intervention, and the effect of pairwise comparisons of two interventions is shown as lines interconnecting the nodes, where the thickness of the lines represents the weight of pairwise comparisons. The number of studies contributing to each pairwise comparison is shown on each line. The assumption of transitivity was evaluated assuming that all the interventions analyzed present the same results regardless of the study to which they belong. The graph of the structure of the network was created, weighting the size of the nodes by said covariates, visually evaluating in which comparisons the covariates were not balanced. We transformed accuracies using the logit function (logit = log[p/(1 − p)]) and treated log odds ratios as the effect measure. Network meta-analyses were conducted within a frequentist framework using a random-effects model. We estimated between-study variance (τ²) and calculated a network I² statistic to quantify heterogeneity. Sensitivity and specificity were analyzed separately using the same approach. A presence of inconsistency was assessed using nodesplitting and by analyzing the level of significance (set at p < 0.05) of the Z statistic to detect disagreement between direct and indirect comparisons of each intervention. The contribution of individual studies to the network and its methodological quality was also analyzed through the contribution matrix. Heterogeneity was assessed by estimating the overall and decomposed within and between studies Cochrane’s Q test, as well as with the estimator I² as a measure of the proportion of observed heterogeneity that is due to true heterogeneity between studies, rather than random error, which was defined as: 0–30%: unimportant heterogeneity; 30–50%: moderate heterogeneity; 50–75%: large heterogeneity; 75–100%: important heterogeneity. These indicators show the likelihood that an intervention is more effective than the other interventions in the network. Publication bias was assessed by visual inspection of funnel plots and Egger’s regression test. The methodological quality of the studies was assessed according to the Cochrane scale by the two independent reviewers.

The meta-analysis was conducted with a comprehensive and systematic approach to identify relevant studies, employing two complementary methods to ensure a robust dataset. The selection of studies drew from diverse sources, including database searches and citation screening. A meticulous search strategy covered diverse and reputable databases, including PubMed (n = 324), Web of Science (n = 135), Embase (n = 272) and the Cochrane Library (n = 23), ensuring a broad scope across different disciplines. The selection of these databases was strategic, leveraging their unique strengths to enhance the understanding of the research question. A total of 754 records resulted from the combined efforts of database searches and citation screening, with 579 duplicates removed during screening. Predefined criteria excluded 175 records before the formal screening process. Subsequently, 42 reports were identified for further retrieval, representing potential contributions to the synthesis of evidence. Rigorous screening excluded six reports based on various criteria. An additional 32 reports were excluded based on more specific criteria, resulting in 11 studies that met the inclusion criteria for the meta-analysis (Figure 1). These studies formed the foundation for subsequent analyses, chosen for their alignment with research objectives and adherence to predefined inclusion criteria. This process followed PRISMA-guided study identification and selection principles (see Figures 2–4).

As Table 1 suggested, datasets used in the studies vary, with some studies relying on local datasets while others utilize publicly available datasets like the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and AddNeuroMed. Each study employs a distinct set of machine learning algorithms for classification purposes, including Bayes, logistic regression (LR), support vector machine (SVM), multilayer perceptron (MP), low-density separation (LDS), extreme learning machine (ELM), feed-forward network (FFR), convolutional neural network (CNN), DenseNet, and ResNet. The participant populations across the studies consist of healthy controls (HC), individuals with MCI, and those diagnosed with AD. Sample sizes vary, reflecting the diversity of populations involved in the research and the broad scope of the studies. The included studies in Table 1 represent a comprehensive overview of research efforts focused on utilizing neuroimaging, particularly MRI, for the diagnosis and classification of Alzheimer’s disease (AD). Conducted by various authors across different years spanning from 2010 to 2021, these studies showcase the evolution of techniques and methodologies in the field. The included studies present two main applications, the classification of mild cognitive impairment (MCI) to AD conversion and distinguishing AD patients from healthy controls (HC). To achieve this, researchers predominantly utilize T1-weighted MRI scans, with some studies also incorporating T2-weighted MRI scans to enhance their analyses. Diagnostic tasks included AD vs. healthy controls (n = 5), AD vs. MCI (n = 3), and prediction of MCI conversion (n = 3). MRI modalities were predominantly structural T1-weighted scans, with two studies using multimodal data (T1 + T2). Datasets were mainly derived from ADNI; two studies used local hospital cohorts. Validation strategies varied (k-fold cross-validation in seven studies, hold-out test sets in three, leave-one-out in one). These characteristics and the extracted performance metrics are summarized in Table 1.

The synthesis of accuracy across various deep and machine learning methods provides crucial insights into their efficacy in classification tasks, with confidence intervals offering a measure of the associated uncertainty. Bayes consistently emerges as one of the least effective methods across the studies. Its accuracy estimate consistently falls below zero, indicating inferior performance compared to most other methods. The confidence interval for Bayes ranges from −0.65 to −0.29, highlighting the certainty of its lower accuracy. On the contrary, CNN consistently outperforms Bayes and several other methods. Its accuracy estimate consistently stays positive, suggesting superior performance in classification tasks. The confidence interval for CNN ranges from −0.55 to −0.18, underscoring its robustness and reliability. DenseNet, while not always the top performer, often demonstrates competitive accuracy. It typically outperforms Bayes but occasionally lags behind CNN and other methods. The confidence interval for DenseNet ranges from 0.05 to 0.16, indicating its variability in performance across different studies. Extreme learning machine (ELM) also shows promise in classification tasks. While its accuracy estimate may vary, it generally performs better than Bayes. The confidence interval for ELM ranges from −0.19 to −0.04, reflecting its consistency in outperforming Bayes. Feed forward network (FFN) tends to perform similarly to Bayes in some cases but may outperform it in others. The confidence interval for FFN ranges from −0.17 to 0.08, suggesting a degree of uncertainty regarding its performance relative to Bayes. Low density separation (LDS) exhibits varying performance across studies, sometimes outperforming Bayes and other times performing similarly to it. The confidence interval for LDS ranges from −0.12 to 0.11, indicating its fluctuating effectiveness in different contexts. Logistic regression (LR) and multilayer perceptron (MP) show comparable performance to Bayes in many instances, with overlapping confidence intervals suggesting no significant difference in accuracy. The confidence interval for LR ranges from −0.09 to 0.15, while for MP, it ranges from −0.03 to 0.11. Simple vector machine (SVM) also demonstrates comparable performance to Bayes, LR, and MP in several studies. Its accuracy estimate overlaps with those of other methods, indicating similar efficacy in classification tasks. The confidence interval for SVM ranges from −0.09 to 0.07. ResNet consistently emerges as one of the top-performing methods across various studies. Its accuracy estimate consistently surpasses that of Bayes and many other methods. The confidence interval for ResNet ranges from −0.64 to −0.15, underscoring its superiority in classification tasks.

Across the network, deep learning architectures ranked highest overall. ResNet emerged as the top-ranked approach, followed by CNN and DenseNet, while several classical approaches demonstrated lower relative performance. Using the revised statistical scale (logit-transformed accuracy with back-transformed interpretation as odds ratios of correct classification), ResNet outperformed CNN, with a pooled OR of correct classification = 1.35 (95% CI 1.12–1.60) in the primary analysis. Between-study heterogeneity was moderate (τ² = 0.04, network I² = 48%), consistent with the variability in diagnostic tasks, MRI modalities, datasets, and validation strategies summarized in Table 1.

Prespecified subgroup analyses (by diagnostic task and MRI modality) suggested that model ranking was broadly stable, although the magnitude of pooled effects varied across subgroups, supporting cautious interpretation of pooled estimates in the presence of clinical and methodological heterogeneity.

Because diagnostic accuracy alone may conceal clinically important trade-offs, we evaluated sensitivity and specificity as secondary outcomes and summarized results in Table 2. In the network analysis of sensitivity, ResNet showed the highest sensitivity signal, with OR = 1.40 (95% CI 1.10–1.76) versus CNN. As a descriptive complement (based on extracted study-level values), the mean sensitivity patterns were consistent with the network ranking: ResNet (94.0%) ranked highest, followed by DenseNet (91.5%), LDS (89.0%), and SVM (86.8%) (Table 2). Specificity showed less separation between algorithms. In descriptive summaries, MP (90.0%) and ResNet (93.0%) were among the highest reported specificities where available, while some methods displayed clear sensitivity–specificity imbalance (e.g., LDS with lower specificity in the available study) (Table 2). Overall, these results indicate that high-ranked models by accuracy generally maintained strong sensitivity, while specificity differences were more modest and less consistently reported across studies.

Visual inspection of the funnel plot suggested slight asymmetry, and Egger’s test indicated potential small-study effects (p = 0.029). Given the limited evidence base and heterogeneous study designs, this finding was interpreted cautiously; nevertheless, it raises the possibility that pooled advantages for certain models could be inflated by selective reporting and emphasizes the importance of publishing negative or null diagnostic modeling results.