Machine learning survey maps 62 studies chasing early Alzheimer’s diagnosis

Alzheimer’s disease accounts for an estimated 60 to 80 percent of dementia cases, and a new systematic survey argues that machine learning could help clinicians catch it earlier. Published in Discover Computing, the review analyzed 62 research studies published between 2017 and 2026 that applied machine learning and deep learning techniques to the prediction and early detection of Alzheimer’s disease. The authors, led by Issa Qabajeh, followed the PRISMA guidelines to ensure methodological rigor and reproducibility, retrieving studies from IEEE Xplore, Scopus, PubMed, and Web of Science. Their central conclusion is sobering for anyone hoping for a single winning algorithm: no one method is always the best, and the differences in reported results stem largely from how data were prepared and tested rather than from any inherent superiority of one algorithm over another.

The scale of the problem motivating this work is considerable. The survey notes that roughly 50 million people were believed to be affected by Alzheimer’s disease in 2020, a figure that could reach 152 million by 2050 if diagnostic advances are not made. A patient with the disease may experience mood swings, behavioral disturbances, loss of balance, and gradual loss of acquired memories, with death possible 3 to 9 years after initial diagnosis. Annual treatment expenses in the United States are estimated at 100 billion dollars. Because the early stage is often mistaken for normal aging by professionals, family members, and friends, the authors emphasize that early diagnosis is essential for proper treatment of symptoms, lifestyle modifications, and planning for the future.

The review’s methodology was deliberately structured. Studies were screened independently by two reviewers, with disagreements resolved by a third, and each study was assessed against eight criteria: patient- versus image-level splitting, use of independent test data, external validation, class balancing, adequacy of sample size, application of nested cross-validation, leakage-free preprocessing, and reporting of confidence intervals. Each item was judged as present, absent, or unclear. To avoid inappropriate comparisons across fundamentally different diagnostic and predictive problems, the authors synthesized studies according to task type, including Alzheimer’s versus healthy control classification, mild cognitive impairment versus healthy control classification, multiclass disease staging, and conversion prediction from MCI to Alzheimer’s disease.

The algorithmic landscape that emerged is broad. Using the total of 62 included studies as the denominator, the survey counted support vector machines in 17 studies, logistic regression in 11, neural networks including CNNs and DNNs in 15, decision trees in 9, random forests in 14, K-nearest neighbors in 14, Naive Bayes in 8, linear discriminant analysis in 6, and other approaches such as AdaBoost, Extra Trees, GANs, or hybrid ensembles in 7. Notably, the most frequently used algorithm was not the one achieving the highest reported accuracy. Random forest appeared most often, yet the authors found that Naive Bayes achieved the highest accuracy among the reviewed studies, with optimized support vector machine and decision tree models also performing well.

Dataset choice dominated the results. ADNI was used in 47 of the 62 studies, OASIS in 7, BioFINDER in 1, OpenNeuro in 1, Kaggle MRI data in 3, other MRI or public datasets in 10, and other sources including clinical, proteomic, EEG, ROSMAP, and multimodal datasets in 14. MRI was the most frequently used modality, compared with EEG and fMRI, which appeared less often. The authors report that within ADNI studies using MRI to classify cognitively normal versus Alzheimer’s disease, deep learning models such as CNNs and LSTMs consistently achieved accuracies above 85 percent, whereas smaller OASIS cohorts showed more variable outcomes, with classical classifiers such as SVMs occasionally outperforming deep networks.

Individual studies illustrate both the promise and the pitfalls. Ali and colleagues developed a hybrid framework using MRI images with an optimized support vector machine, achieving accuracies of 99.35 and 99.55 percent on two datasets. Abd El Hamid and colleagues applied single-nucleotide polymorphisms with machine learning methods and reported 99.63 percent accuracy. Zolfaghari and colleagues combined two convolutional neural networks with an ensemble of SVM, random forest, and KNN, reaching approximately 99.06 percent. Zubair and colleagues used U-Net segmentation with EfficientNet feature extraction and SVM classification, achieving 97.78 percent with explainable AI techniques added. Yet the survey repeatedly flags that very high accuracies, in the 95 to 99 percent range, were typically reported in small studies using internal splits, raising concerns about overfitting and limited generalizability.

Not every result was impressive. Khasanah and colleagues, analyzing 6400 pre-processed MRI images from a Kaggle dataset across four disease classes with K-nearest neighbors, obtained accuracies varying between 45.86 and 50.47 percent, a result the authors attribute to how imbalanced classes affect classification performance, particularly for the minority Moderate Demented class. Other studies faced structural constraints: Zhang and colleagues used a single ADNI dataset, limiting generalizability; Upadhyay and colleagues relied on the relatively small OASIS-1 dataset; and Maity and colleagues worked with a limited data volume from OpenNeuro and Harvard sources. Wong and colleagues addressed data scarcity by using generative adversarial networks to create synthetic MRI images from OASIS-3, achieving classification accuracy above 80 percent even with reduced real data, though without external datasets.

The survey also examined evaluation practices and found them uneven. Accuracy was reported in 29 studies, precision in 1, recall in 3, F1-score in 1, AUC in 4, ROC in 3, and both sensitivity and specificity in 1, while 30 studies reported no evaluation metrics at all. The authors caution that results derived from studies differing in dataset source, diagnostic task, imaging modality, or validation rigor cannot be considered evidence of any algorithm’s superiority. Studies employing independent test sets or external validation produced more reliable outcomes than those relying on internal data splitting. Reproducibility remains weak because few studies share their code or procedural steps, and interpretability of results is underdeveloped.

Explainable AI received particular attention. Methods such as SHAP and LIME suit tabular data, while Grad-CAM and saliency maps are commonly applied to imaging tasks. Some studies employed these approaches to generate explanations at the model level or for individual predictions. However, the stability of such explanations is critical, as perturbation-based methods can yield inconsistent results across runs, and saliency maps may be misleading if they highlight irrelevant regions or artifacts. Strikingly, none of the reviewed studies systematically evaluated explanation quality, and few engaged physicians to assess whether the explanations were medically meaningful, a gap the authors identify as significant for current research.

The authors conclude that meaningful progress requires rigorous testing on representative datasets, transparent reporting of research procedures, and cohorts reflecting real-world populations. Future research, they argue, should focus on more diverse datasets, standardized data preparation and evaluation methods, multiclass prediction of Alzheimer’s disease stages rather than binary classification, and data sharing among institutions to accelerate model development. They note that most reviewed studies focused on binary classification of Alzheimer’s versus healthy controls, which may oversimplify disease progression and fail to detect intermediate stages such as mild cognitive impairment. Hybrid approaches combining deep learning with other classification methods showed promise, with one reported accuracy of 96.87 percent, and the combination of multimodal data with neural networks may enhance clinical relevance, but scaling these systems for real medical settings remains a challenge.

Subject of Research: A systematic survey of machine learning and deep learning techniques for early Alzheimer's disease diagnosis

Article Title: A systematic survey of machine learning and deep learning techniques for Alzheimer’s disease diagnosis

Article References: Qabajeh, I., Alawneh, A., El-Emam, N., Qabajeh, R., Zain, H., Almakadmeh, M., & Aljawarneh, E. (2026). A systematic survey of machine learning and deep learning techniques for Alzheimer’s disease diagnosis. Discover Computing, 29(1), Article 846. https://doi.org/10.1007/s10791-026-10698-6

Image Credits: AI Generated

DOI: 10.1007/s10791-026-10698-6

Keywords: Alzheimer's disease, machine learning, deep learning, systematic review, early detection, MRI, ADNI, Naive Bayes, support vector machine, explainable AI, mild cognitive impairment, PRISMA