Literature-Identified Serum miRNA Signatures for Cognitive Decline: Integrated Analysis and Machine-Learning Diagnostics in Alzheimer’s Disease

Background Because identification of preclinical Alzheimer’s disease (AD) still relies on costly or invasive A/T/N biomarkers, we systematically integrated published case-control studies to identify differentially expressed serum microRNAs (miRNAs) associated with cognitive function in patients with AD, and validated their diagnostic performance in the large public cohort GSE120584. Methods We searched Chinese- and English-language databases for studies reporting serum miRNA expression differences in AD and their associations with cognitive scale scores. Validated target genes were retrieved from miRTarBase; protein–protein interaction (PPI) networks were constructed using STRING; functional modules were identified with the Cytoscape plug-in MCODE; and enrichment analyses were performed for GO, KEGG, Reactome, and Hallmark gene sets. Differential expression analysis in GSE120584 was conducted using limma with covariate adjustment (age, sex, and APOE4), and partial correlations between miRNA expression and age, sex, and APOE4 were calculated. Literature-derived miRNAs were matched to GSE120584, and unmatched miRNAs were supplemented by same-family candidates when necessary. Based on correlation-network analysis and nested cross-validation, optimal miRNA combinations were selected to construct diagnostic models with age or age + sex as baseline predictors. Model performance was evaluated using out-of-fold ROC and precision–recall (PR) curves, calibration curves, and decision curve analysis (DCA). Results Twenty-three publications including 2,580 patients with AD and 2,261 controls were included. Twenty-one differentially expressed serum miRNAs were identified, including miRNAs positively (n = 15) or negatively (n = 7) correlated with Mini-Mental State Examination (MMSE) scores. Targets of positively correlated miRNAs were enriched in PI3K/AKT/mTOR, Wnt, and TNF-α/NF-κB signaling pathways, whereas targets of negatively correlated miRNAs were mainly involved in cell cycle regulation, the G2/M checkpoint, and oxidative stress responses. After matching and expansion in GSE120584, 25 significantly differentially expressed miRNAs were identified. The minimal miRNA signatures with optimal diagnostic value were miR-211-5p alone (K1) and the three-miRNA panel miR-211-5p, miR-128-1-5p, and miR-128-3p (K3). When age and sex were added, the “K3 + age + sex” model showed the best performance (AUC = 0.838, AP = 0.934, Brier score = 0.162), yielding the highest sensitivity (0.563) and the best PPV (0.954) at specificity ≥ 0.90. Conclusion Using dual validation from published literature and a large cohort, we identified cognition-associated serum miRNAs in AD and established a combined diagnostic model integrating miRNA levels with clinical characteristics (age and sex). miR-211-5p and miR-128 family members appear promising as peripheral blood biomarkers, but require validation in independent cohorts.