Related Experiment Video
Updated: Sep 2, 2026
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
Urinary metabolomics-based machine learning for diagnosis of early gastric neoplasia: a retrospective diagnostic
Mengchen Luo1, Feng Chen2, Tao Ling1
1Department of Gastroenterology, Affiliated Jinhua Hospital, Zhejiang University School of Medicine, Jinhua, China.
Objectives:
Early gastric cancer (EGC) is often asymptomatic. Non-invasive and scalable screening tools remain limited. We aimed to develop and validate a urinary metabolomics-based machine-learning model for detecting early gastric neoplasia (EGN), including high-grade intraepithelial neoplasia (HGIN) and EGC, and to benchmark it against routine serum markers.
Methods:
This was a retrospective diagnostic case-control study using archived urine specimens collected during routine clinical care, with internal split-sample validation and an independent temporal validation cohort. Morning urine samples from two cohorts were profiled by liquid chromatography-mass spectrometry (LC-MS). Cohort 1 was split into a training set and a testing set. LASSO regression was used to identify candidate features, and a random-forest classifier was developed. An external targeted-quantification cohort was used for validation and single-biomarker assessment.
Results:
In this retrospective diagnostic case-control study, urinary metabolomics combined with machine learning was used to establish a three-metabolite diagnostic model (3-DM). In the internal testing set, the 3-DM showed discriminative potential with an AUROC of 0.81 (95% CI: 0.600-0.963). In the independent validation cohort, only 1-methylnicotinamide (1-MNA) remained significantly different between NC and EGN, whereas 3-pyridylacetic acid (3-PAA) and phenylacetylglutamine (PAGln) did not show consistent inter-cohort performance. 1-MNA showed a gradual decrease with disease progression, suggesting that it may represent the most reproducible candidate biomarker in this study.
Conclusions:
These preliminary findings suggest that urinary metabolomic profiling may have potential for non-invasive detection of EGN. Among the candidate metabolites, 1-MNA showed the most reproducible cross-cohort signal, but its diagnostic performance and clinical applicability require further validation in larger prospective cohorts.
