TITLE:
A Cross-Cohort Prediction Study of Type 2 Diabetes Based on Gut Microbiome and Machine Learning
AUTHORS:
Zhiming Deng
KEYWORDS:
Type 2 Diabetes, Gut Microbiome, Machine Learning, XGBoost, Microbial Biomarkers
JOURNAL NAME:
Journal of Computer and Communications,
Vol.14 No.9,
September
17,
2026
ABSTRACT: Early screening for type 2 diabetes mellitus (T2DM) still lacks convenient and non-invasive biomarkers. This study aimed to utilize gut microbiome 16S rRNA data to construct an interpretable machine learning model, achieving accurate cross-cohort prediction of T2DM and identifying generalizable microbial biomarkers. Fecal 16S rRNA sequencing data from two independent cohorts, the Guangdong Gut Microbiome Project (GGMP, n = 2603) and the Shandong Gut Microbiome Project (SGMP, n = 968), were integrated. An ensemble learning model was constructed using the XGBoost algorithm, generating multiple sub-models through a random subsampling strategy. The contribution of microbial features to prediction was quantified using SHAP values, and the model’s generalization ability was validated in a cross-cohort scenario. The model achieved an AUC of 0.87 (95% CI: 0.85 - 0.89) in internal validation within GGMP and an AUC of 0.82 (95% CI: 0.79 - 0.85) in external validation within SGMP. SHAP analysis identified 12 cohort-consistent T2DM-associated microbial genera, with decreased abundance of Faecalibacterium, Roseburia, and Bifidobacterium being most strongly associated with increased T2DM risk. The model outperformed a baseline logistic regression model based solely on host clinical indicators (BMI, age, fasting blood glucose) (AUC = 0.76). This machine learning model based on the gut microbiome can effectively predict T2DM in diverse geographic populations, and the identified microbial biomarkers exhibit cross-cohort stability, providing a novel tool for non-invasive screening of T2DM.