Accurate and early characterization of Alzheimer’s disease (AD) progression remains a major clinical challenge, particularly in biologically heterogeneous stages such as mild cognitive impairment (MCI). We propose an sMRI-based machine learning framework for multi-stage AD discrimination, encompassing six clinical groups (cognitively normal (CN, n = 103 ), subjective memory complaints (SMC, n = 54 ), early MCI (EMCI, n = 151 ), MCI ( n = 229 ), late MCI (LMCI, n = 117 ), and AD ( n = 114 )), with explicit β -amyloid (A β ) stratification of MCI subtypes ( + / − ). T1-weighted sMRI scans from 768 ADNI subjects were analyzed using histogram and GLCM texture features across multiple anatomical planes and multi-scale wavelet decompositions. The framework demonstrated robust performance across 36 pairwise comparisons under stratified 5-fold cross-validation with SMOTE and RUS, achieving AUC > 0.95 in well-separated tasks (e.g., AD vs EMCI − ) and in SMC vs AD and SMC vs MCI + . More challenging “gray-zone” distinctions (e.g., EMCI − vs EMCI + ) showed moderate performance ( AUC ≈ 0.62–0.67), with RUS providing the most overall balanced results. Feature analysis revealed dominant contributions from GLCM descriptors, complemented by histogram features, with highest discriminability in Level-2 wavelet sub-bands (HL2 and LH2). A β stratification enabled a more biologically grounded interpretation of disease progression, improving discrimination in later stages while remaining challenging in early-stage comparisons. Overall, sMRI radiomics provides an interpretable framework for capturing stage-specific structural patterns in AD, highlighting the importance of biological stratification and robust validation.
Menezes et al. (Tue,) studied this question.