A multimodal model for clinically defined MCI integrating plasma biomarkers and brain microstructural features: a...

Mild cognitive impairment (MCI) is clinically and etiologically heterogeneous. We aimed to develop and externally validate a multimodal model for distinguishing clinically defined MCI from normal cognition using plasma biomarkers and diffusion tensor imaging (DTI)-derived white-matter features.

We analyzed 310 participants from a local discovery cohort (n = 153; 80 MCI and 73 cognitively normal [NC]) and an independent Alzheimer's Disease Neuroimaging Initiative phase 4 (ADNI4) validation cohort (n = 157; 68 MCI and 89 NC). Plasma phosphorylated tau 217 (p-tau217), neurofilament light chain (NFL), and glial fibrillary acidic protein (GFAP) were log-transformed and standardized using parameters estimated exclusively from NC participants in the local discovery cohort and then frozen for application to ADNI. Twenty-seven tract-level fractional anisotropy (FA) measures were considered as candidate imaging predictors. DTI feature selection used L1-penalized logistic regression with the 1-SE rule, with feature selection repeated within nested 10-fold cross-validation. Logistic regression (LR), support vector machine (SVM), random forest (RF), and XGBoost models were developed using local data, and the finalized pipelines were subsequently applied to ADNI4 without feature reselection, hyperparameter tuning, or recalibration.

Plasma p-tau217, NFL, and GFAP were higher in MCI than NC in both cohorts (all cohort-specific P ≤ 0.005). LASSO selected fornix (FX) and cingulum hippocampus (CGH) FA, and both features were retained in all 10 outer cross-validation folds. Nested internal AUCs ranged from 0.791 to 0.839. Under strict external validation, RF achieved the highest AUC of 0.829 (95% CI 0.762-0.889), followed by SVM with an AUC of 0.802 (95% CI 0.730-0.870), XGBoost with an AUC of 0.790 (95% CI 0.714-0.860), and LR with an AUC of 0.770 (95% CI 0.688-0.843). Relative to plasma biomarkers plus demographics, the full multimodal models significantly improved external discrimination across all four classifiers (all P ≤ 0.004). Improvements relative to DTI plus demographics were smaller and classifier-dependent; only RF showed a statistically significant increment (ΔAUC = 0.063, 95% CI 0.004-0.125; P = 0.036). RF had the lowest external Brier score (0.171). In the local cohort, lower FX and CGH FA remained associated with MCI after adjustment for available vascular and metabolic risk factors.

A multimodal signature combining two selected DTI features with three prespecified plasma biomarkers and demographic covariates showed reproducible discrimination of clinically defined MCI across independent cohorts. The findings support further evaluation of this approach as an adjunctive MCI risk-stratification strategy but do not establish Alzheimer's disease (AD)-specific etiology or clinical readiness. Prospective validation incorporating amyloid/tau status, longitudinal conversion outcomes, and real-world calibration is required.