Multi-marker discovery for mild cognitive impairment in metabolomics using machine learning with a global surrogate...

Dementia can be prevented through early intervention; hence, there is an urgent need for biomarkers to help diagnose mild cognitive impairment (MCI).

We aimed to develop a multi-marker panel composed of plasma metabolites to aid in the diagnosis of MCI.

We performed an analysis of a multi-marker panel of MCI metabolites using a random forest algorithm with variable selection methods and a global surrogate with principal component analysis and partial least squares (PLS).

By incorporating variable selection methods, we constructed a predictive model that demonstrated robust performance, with an AUC of approximately 0.85 in both cross-validation and test evaluations, using only five metabolites (methionine, quinic acid, hypoxanthine, O-acetylcarnitine, and 2-oxoglutaric acid). However, owing to the limited number of selected metabolites, it was challenging to infer the biological meaning of this multi-marker panel. To interpret this multi-marker panel biologically, we constructed a global surrogate model using PLS. By examining the PLS loadings corresponding to the scores with intergroup differences, we identified a relationship between 14 metabolites involved in neuronal energy metabolism and neurotransmission. This suggests that the multi-marker panel constructed in this study is related to abnormalities in energy metabolism and neurotransmission in patients with MCI.

The method used in this study may be broadly applicable for analyzing multi-marker panels of metabolites and their biological interpretation. This study included an independent validation, and further larger-scale studies using additional external cohorts are warranted to confirm the generalizability of this approach.