Skip to content
Open access

Interpretable Artificial Intelligence Assisted Plasma Metabolomics for Breast Cancer Classification

Sep 2026 · Computer and Decision Making: An International Journal · 0 citations

Abstract

Accurate discrimination of breast cancer from non-cancer controls remains important for the development of complementary minimally invasive diagnostic approaches. In this study, an interpretable artificial intelligence framework based on quantitative plasma metabolomics was developed and independently assessed for breast cancer classification. Publicly available targeted metabolomics data from 811 participants were analyzed across development (n=423), locked test (n=276), and independent-validation (n=112) cohorts. Four metabolites - inosine, uridine, L-phenylalanine, and threonine - were initially considered. Model construction was confined to the development cohort, with leakage-controlled preprocessing, repeated cross-validation, nested hyperparameter optimization, exhaustive assessment of metabolite subsets, and ablation analysis being applied. Among the algorithms examined, the strongest nested performance was obtained with the Explainable Boosting Machine. A parsimonious three-metabolite signature consisting of inosine, uridine, and L-phenylalanine was selected and frozen before validation. An AUROC of 0.947 and an AUPRC of 0.965 were obtained in the locked test cohort, whereas values of 0.922 and 0.964, respectively, were achieved in the independent-validation cohort. Inosine was identified by the explainability analysis as the dominant contributor. Overall, the findings support further evaluation of interpretable plasma metabolomics as a potential adjunctive tool for breast cancer classification rather than as a replacement for established screening or diagnostic methods. However, because participant-level tumor-stage information was unavailable, the present results should not be interpreted as direct evidence of stage-specific or population-level early-detection performance. Prospective validation in stage-characterized and clinically representative populations is required to determine whether the model provides incremental value alongside existing clinical and imaging-based pathways.  

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.