Development of Plasma Protein Classification Models for Alzheimer’s Disease Using Multiple Machine Learning Approaches

Alzheimer's Disease (AD) management is challenging due to limitations in detection methods. Currently, cerebrospinal fluid (CSF) biomarkers involve assessing β-amyloid (Aβ) and phosphorylated tau proteins. The lumbar puncture procedure to obtain CSF is invasive and sometimes causes significant anxiety in patients. In contrast, plasma biomarkers would allow rapid, accurate, and cost-effective diagnosis, while minimizing invasiveness and discomfort. Using a dataset involving 120 plasma proteins from clinically diagnosed AD patients versus cognitively normal subjects, we developed classification models by applying various machine learning algorithms (EBlasso, EBEN, XGBoost, LightGBM, TabNet, and TabPFN) to plasma proteomic measurements. Gene ontology and pathway enrichment, and a literature review were used to evaluate the potential relevance of the biomarkers identified in AD-related mechanisms. Biomarkers identified were also evaluated for the enrichment of aging-related biomarkers. The models developed yielded high AUROC and accuracy, mostly >0.9. Proteins selected as predictors by all the models included Angiopoietin-2 (ANG-2), epidermal growth factor (EGF), Interleukin 1α (IL-1α), and platelet growth factor subunit B (PDGF-BB). Ample previous literature supported their relevance in AD. The pool of all the biomarkers identified was significantly enriched with known aging-related biomarkers (<i>p</i> = 0.040). Applying cutting-edge algorithms is expected to be advantageous for developing AD prediction models with plasma proteomic data, and future large studies to externally validate the constructed models in other populations to assess their generalizability is important. The proteins uncovered may represent novel preventative or therapeutic targets.