Blood based multi biomarker panel discriminates Alzheimer disease from healthy controls in a single center nested cross validation

AD involves complex pathological pathways that are poorly characterized by conventional univariate analyses. This study aimed to screen discriminatory blood biomarkers and evaluate their performance in differentiating patients with established AD from HCs using multimodal integrated analytical strategies. We recruited 200 participants (100 AD, 100 HCs) and measured 12 blood biomarkers, including neurotrophic factors, tau proteins, neuronal injury markers and peripheral inflammatory indices. PCA and OPLS-DA were applied for exploratory visualization. Nested stratified ten-fold cross-validation was used to avoid data leakage. L1-penalized LASSO logistic regression served as the primary feature selection method, with a stacking ensemble model constructed for benchmarking. Under lambda.1se criterion, eight predictive biomarkers were retained. BDNF, NGF, NFL and pTau217 were robust markers consistently selected by both pipelines; pTau231 and S100β represented second-tier stable markers. PLR and SII were excluded by LASSO-1se but occasionally selected in stacking, while Lym was retained only in LASSO-1se. Age, gender, neutrophil and platelet counts were eliminated in both models. Core AD neuropathology markers exhibited better cross-model reproducibility than non-specific systemic inflammatory indices. The LASSO-derived model achieved out-of-fold AUROC of 0.9999 and AUPRC of 0.9999; the stacking model reached AUROC = 0.999 and AUPRC = 0.999. All performance metrics were derived from single-centre internal cross-validation. External multi-cohort validation is mandatory before clinical application. This multimodal framework identifies multi-pathway blood biomarkers reflecting tau pathology, axonal injury, neurotrophic deficiency and peripheral immune perturbation, providing insights into AD pathogenesis.