A comparative study of machine learning models for microbiome-based diagnosis and multi-class staging of colorectal cancer

Colorectal cancer (CRC) is closely associated with gut microbiota dysbiosis; however, comprehensive benchmarking of machine learning models that integrate case–control differentiation with TNM staging remains limited. To address this gap, we systematically evaluated eight machine learning algorithms using 16S rRNA sequencing data derived from fecal samples. Our analysis included traditional methods (Logistic Regression), classic machine learning algorithms (Random Forest, XGBoost, SVM, KNN), and advanced deep learning architectures (CNN, MLP, GCN). These models were applied to both binary classification (CRC vs. healthy controls) and multi-class classification (TNM staging) tasks. Two complementary feature selection strategies (LEfSe and RFCV) were employed, followed by hyperparameter optimization, five-fold cross-validation on the training set (n = 510). Model performance was assessed on internal test sets (n = 210) and multiple independent cohorts (total n = 1039), with bootstrap analysis conducted to provide robust performance estimates. For binary classification, Random Forest achieved the highest and most consistent AUCs across all validation cohorts (0.8633–0.8672), significantly outperforming most other models (p < 0.05). For multi-class classification (Control vs. TNM I/II vs. TNM III/IV), Random Forest again demonstrated superior macro-average AUCs (0.7736–0.7816) on the independent validation set, with particularly high sensitivity (0.8571) for early-stage detection. Based on this comprehensive evaluation, Random Forest emerges as a highly robust and versatile algorithm for CRC diagnosis based on microbiome data, demonstrating balanced performance, effective feature selection, and promising potential for clinical staging applications.