Machine learning models of Alzheimer's disease spectrum using blood tests

INTRODUCTION: The diagnosis of Alzheimer's disease (AD) traditionally relies on cerebrospinal fluid and plasma levels of amyloid beta and phosphorylated tau. Although informative, these biomarkers represent a narrow, hypothesis-driven approach to intercept the disease. METHODS: Data-driven analysis was applied on demographic data, apolipoprotein E (APOE) ε4 allele, and 82 biomarkers obtained from blood tests of healthy controls (HC), mild cognitive impairment that remained stable within 36 months following blood collection (sMCI), and patients with AD. RESULTS: Statistical analyses revealed differences among groups in many cholesterol-related analytes. APOE ε4 and analytes such as amino acids, lipoproteins, and fatty acids emerged as the most influential features in machine learning (ML) classification algorithms. Glycolysis-related metabolites and amino and fatty acids were predictive for distinguishing sMCI and AD from HC. DISCUSSION: These findings support the hypothesis that systemic alterations also occur during the preclinical stages of dementia, which can be detected by ML models on blood biomarkers. Highlights: Machine learning on blood tests detects preclinical cognitive decline. Glycolysis metabolites are predictive for distinguishing stable MCI and AD from HC. Amino acids, lipoproteins, and fatty acids are the most predictive features. Inflammatory and metabolic biomarkers represent a biosignature of cognitive health.