ALZHEIMER'S DISEASE DIAGNOSIS. THE ROLE OF PLASMA LIPIDOMICS AND ARTIFICIAL INTELLIGENCE

Abstract

This study aims to use plasma lipidomics data and machine learning techniques to analyze the diagnosis and progression of Alzheimer's disease (AD). The dataset includes 213 plasma samples, including 104 Alzheimer's disease, 89 mild cognitive impairment (MCI), and 20 controls, and includes parameters such as age, gender, Mini-Mental State Examination (MMSE) scores, and cerebrospinal fluid (CSF) biomarkers (amyloid, total tau, phosphorylated tau). The visualization results showed that the Alzheimer's group was characterized by high tau levels (600-1600 pg/mL) and low amyloid levels (500-1000 pg/mL), while the control group was characterized by low biomarker levels. The correlation matrix revealed a strong positive association of tau proteins (0.72) and a negative association between amyloid and tau (-0.45). Ten machine learning models were analyzed, with Extra Trees (97.7% accuracy, 95.4% F1-score) and Random Forest (93% accuracy, 91.9% F1-score) showing the highest performance. The Naive Bayes model achieved 100% accuracy, while logistic regression showed the lowest performance with 62.8% accuracy. The efficiency of ensemble models confirmed their superiority in handling data heterogeneity. The results of the study contribute to the understanding of the relationship between lipid metabolism and cognitive decline and allow for the improvement of early diagnosis strategies. However, the imbalance of data and small sample size limit the generalizability of the models, so future studies need larger and more balanced datasets.