Standardized Comparison of Clinical, Cognitive, Genetic, Neuroimaging, and Fluid Biomarkers for Predicting 24-Month Progression from Mild Cognitive Impairment to Alzheimer's Disease

Identifying individuals with mild cognitive impairment (MCI) likely to progress to Alzheimer's disease (AD) is important for patient management and clinical trial enrollment. Although cognitive assessments, genetics, neuroimaging, and fluid biomarkers are each associated with disease progression, their predictive value has not been systematically compared using an identical cohort and evaluation framework. This study compared the predictive discrimination of clinical, cognitive, genetic, imaging, and cerebrospinal fluid (CSF) biomarkers, individually and combined, for 24-month progression from MCI to AD. A retrospective analysis used data from 2,430 participants with MCI enrolled in ADNI, including 547 who progressed to AD within 24 months and 1,883 who remained stable. Seven models were evaluated using identical preprocessing and modeling procedures: a clinical baseline (age and sex), the baseline plus a single modality out of cognitive assessment, APOE {varepsilon}4 genotype, structural MRI, CSF biomarkers, or PET biomarkers and a multimodal model combining all five. Performance was assessed using repeated 5 x 10 stratified cross-validation. Out-of-fold predictions from a separate 5-fold split were used to estimate confidence intervals and compare AUCs via DeLong's test with Holm/Bonferroni correction. Discrimination increased progressively across modalities. The clinical baseline achieved an AUC of 0.556; adding APOE e4 genotype increased performance to 0.692, CSF biomarkers to 0.729, PET biomarkers to 0.783, structural MRI to 0.836, and cognitive assessment to 0.918. Cognitive assessment significantly outperformed all other individual modalities, including MRI (difference in AUC = 0.079, P < 0.001). The multimodal model achieved the highest overall discrimination (AUC = 0.933), significantly outperforming cognitive assessment alone (difference in AUC = 0.016, P < 0.001), though it required complete data from only 20.5% of participants, versus 99.3% for cognitive assessment. Within a common evaluation framework, cognitive assessment demonstrated the greatest predictive discrimination among individual modalities for 24-month progression from MCI to AD, followed by structural MRI and PET. A multimodal model achieved the highest overall discrimination but required complete data from only one-fifth of the cohort. These findings suggest that routinely collected cognitive assessments capture substantial prognostic information, while full multimodal integration offers only modest incremental value relative to its reduced applicability.