Use of machine learning algorithms and NMR metabolomics to identify potential serum biomarkers in patients with amnestic mild cognitive impairment

Abstract

Objective

To identify potential serum biomarkers in patients with amnestic mild cognitive impairment (aMCI) using machine learning algorithms and nuclear magnetic resonance (NMR) metabolomics.

Methods

Seventy-four serum samples from 33 patients diagnosed with aMCI and 41 age-, sex-, and education-matched elderly controls (ECs) were subjected to analysis using a Bruker 850 MHz superconducting NMR instrument. We employed a multivariate analysis approach to identify the serum metabolic characteristics that distinguished patients with aMCI from ECs. Two machine learning (ML) algorithms, decision tree and random forest, were then employed to assess the discriminatory power of the resulting biomarker panel.

Results

Eighteen differential metabolites were identified by orthogonal partial least squares-discriminant analysis. Then, utilizing the LASSO regression algorithm, nine potential metabolic markers were selected for distinguishing patients with aMCI from the EC group. Based on this biomarker panel, the prediction models constructed using random forest discriminated aMCI from EC with a sensitivity of 0.912, specificity of 0.947, and an area under the receiver operating characteristic curve (AUC) value of 1.00. The decision tree model had a sensitivity of 0.958, specificity of 1.00, and AUC value of 0.979.

Conclusions

This study identified a panel of nine metabolites and could be used as potential biomarkers for aMCI. ML algorithms combined with NMR-based metabolomics can effectively provide valuable molecular-level metabolic information for the assessment of the early stages of Alzheimer's disease.