Answers are grounded in this paper and cite their sources.
Background Combining multiple diagnostic markers is a widely adopted strategy for enhancing diagnostic accuracy. Although numerous linear combination methods have been proposed, comparative studies have primarily focused on discrimination, with less attention to calibration. This study aimed to systematically compare linear combination methods within a unified evaluation framework. Methods We analyzed data from 860 participants in the Alzheimer’s Disease Neuroimaging Initiative (ADNI). Noninvasive markers, encompassing cognitive, functional, structural MRI measurements, plasma biomarkers, and multidomain markers, were evaluated across four diagnostic tasks: cognitively normal (CN) versus Alzheimer’s disease (AD), mild cognitive impairment (MCI) versus AD, CN versus MCI, and amyloid-β positivity classification. Linear combination methods, including logistic regression, suliu, min-max, distribution-free stepwise, and lasso-logistic methods were compared across different diagnostic tasks. Models were developed utilizing 5-fold cross-validation and assessed for discrimination, calibration, and robustness across tasks, marker domains, and subgroups. Results Lasso-logistic and stepwise multidomain models achieved AUCs exceeding 0.99 for CN vs. AD, 0.91 for MCI vs. AD, 0.84 for CN vs. MCI, and 0.77 for Aβ positivity, with no significant difference between two methods. Calibration performance varied across diagnostic tasks and methods; lasso-logistic showed the most consistent calibration, with slopes closest to 1 and intercepts closest to 0. In individual-domain models, logistic regression and stepwise approaches showed consistently good discrimination and calibration across domains and tasks. The suliu method showed relatively high slopes in some settings; while the min-max showed substantial deviations in calibration across domains and diagnostic tasks. Models based on ECogSP and RAVLT showed high AUCs for AD stage diagnosis, whereas plasma-based models showed relatively strong discrimination for amyloid-β positivity. Subgroup analyses showed higher accuracy in females, APOE ε4 carriers, and Aβ-positive individuals, but reduced separability in advanced-stage (MCI vs. AD). Conclusions A unified evaluation of discrimination, calibration, and robustness offers a comprehensive assessment of diagnostic marker combination methods. The stepwise and lasso-logistic exhibited comparable performance for multidomain marker integration, while logistic and stepwise methods showed relatively stable performance in individual-domain analyses. These findings provided guidance for selecting linear combination methods in diagnostic studies and for interpreting performance of noninvasive markers across AD stages.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: