Predicting Conversion from Mild Cognitive Impairment to Alzheimer's Disease: A Systematic Review of Deep Learning Models for Early-Stage Disease Classification.
INTRODUCTION Alzheimer's disease (AD) is a progressive neurodegenerative disorder for which early diagnosis-particularly the accurate prediction of conversion from mild cognitive impairment (MCI) to AD-is essential to enable timely and effective therapeutic interventions. Deep learning (DL) models have demonstrated substantial promise in this domain; however, critical challenges persist, including multiclass staging of disease progression, longitudinal data modeling, and effective multimodal data integration. This systematic review provides a critical appraisal of DL-based approaches for predicting MCI-to-AD conversion, with particular emphasis on these three key challenges. METHODS This systematic review was designed and conducted in accordance with the PRISMA guidelines. A comprehensive literature search was performed across PubMed, Scopus, IEEE Xplore, and Web of Science for articles published between January 1, 2019, and February 20, 2026. Following rigorous screening of titles, abstracts, and full texts, 60 studies were included that employed deep learning models for AD stage classification and/or MCI-to-AD conversion prediction. Data were extracted and synthesized regarding study design, datasets, input modalities, DL architectures, and classification tasks. RESULTS Cross-sectional approaches remained predominant (47 studies), while longitudinal designs were less common (13 studies) and showed heterogeneous but promising performance for MCI-to-AD conversion prediction. Heavy reliance on the ADNI dataset (42 studies) represents a major limitation to generalizability. Multimodal models were used in 24 studies and often reported strong performance, particularly for challenging MCI-related tasks; however, direct cross-study comparisons should be interpreted cautiously because of substantial heterogeneity in datasets, prediction tasks, validation strategies, and methodological quality. Convolutional neural networks (CNNs) dominated neuroimaging-based modeling (21 studies), while recurrent neural networks (RNNs) (4 studies) and transformers (4 studies) have emerged for capturing longitudinal dependencies and global relationships. The binary pMCI vs. sMCI classification proved the most challenging task (accuracy range: 71.71-96.3%), with performance declining as the number of classes increased in multiclass settings. Key limitations across studies include lack of diverse datasets, overfitting, and poor model interpretability. CONCLUSION Deep learning models hold considerable potential for predicting MCI-to-AD conversion, yet substantial barriers remain to their translation into routine clinical practice. Greater emphasis on longitudinal analysis, intelligent multimodal fusion, and interpretable architectures is essential for clinical impact. Future research should prioritize the development of diverse, multicenter datasets, advancement of explainable AI (XAI) techniques, and the design of personalized time-to-event models. This review offers a comprehensive roadmap to guide subsequent investigations toward more accurate, reliable, and clinically actionable diagnostic tools in Alzheimer's disease.