Detection of Hepatocellular Carcinoma Using Optimized miRNA Combinations and Interpretable Machine Learning Models

Early screening to improve the survival rate of hepatocellular carcinoma (HCC) patients remains a critical clinical challenge. However, current methods for early cancer detection still lack sufficient sensitivity and specificity. In this study, we developed an interpretable machine learning (ML) model for HCC prediction using miRNA-seq data from TCGA-LIHC. Five feature selection algorithms and 11 classifiers were employed to identify key miRNAs for HCC diagnosis, with the Sarsa-CatBoost combination yielding the best performance, achieving an accuracy of 96.0%, a matthews correlation coefficient of 89.2%, and an AUC of 97.6%. Three interpretability techniques—permutation feature importance, Partial Dependence Plots (PDP), and Shapley Additive Explanations (SHAP)—were integrated into the pipeline to enhance model transparency. Both permutation feature importance and SHAP analysis identified hsa-miR-93 and hsa-miR-139 as key miRNAs for HCC prediction. SHAP and PDP further revealed how these miRNAs influence HCC development and their potential synergistic interactions. Subsequently, survival analysis, target prediction, and enrichment analysis were performed to assess the biological significance of these miRNAs. The enrichment analysis suggested that these miRNAs may be implicated in the onset and progression of HCC. These miRNAs hold promise as potential biomarkers for HCC diagnosis and prognosis, warranting further investigation into their mechanistic roles and clinical applicability.