Machine learning-based prediction of immune infiltration patterns in extrahepatic cholangiocarcinoma.

BACKGROUND: Immune cell infiltration patterns play important roles in shaping the tumor microenvironment of cholangiocarcinoma. However, efficient approaches for classifying these patterns from transcriptomic data remain limited. This study aimed to develop a machine learning-based framework to predict immune infiltration patterns in extrahepatic cholangiocarcinoma (EHCC). METHODS: Transcriptomic data from the GSE132305 cohort were analyzed using CIBERSORT to estimate the proportions of 22 immune cell types. K-means clustering was applied to identify distinct immune infiltration patterns, with the optimal number of clusters determined by the silhouette method. Differentially expressed genes (DEGs) between clusters were identified using limma (|log2FC| > 2, adjusted p < 0.05). Functional enrichment analyses, including GO, KEGG, and GSEA, were performed to characterize biological differences between clusters. Hub genes were selected using five feature selection methods. Nine machine learning classifiers were then combined with these feature sets to construct 45 prediction models, which were evaluated using stratified five-fold cross-validation based on AUC, accuracy, sensitivity, and specificity. RESULTS: Two distinct immune infiltration patterns were identified among 182 EHCC samples, with significant differences in immune cell composition between Cluster 1 (n = 92) and Cluster 2 (n = 90). A total of 116 DEGs were identified between the two clusters. Functional enrichment analyses revealed marked differences in immune-related signaling and translational activity. Among the 45 models, the GBM-selected features combined with an SVM classifier achieved the best overall performance, with an AUC of 0.9958 and an accuracy of 96.7%. CONCLUSION: We developed a transcriptome-based machine learning framework for classifying immune infiltration patterns in EHCC. These findings improve the understanding of immune heterogeneity in EHCC and provide a computational basis for future validation in independent cohorts with clinical annotation.