Application of machine learning methods in healthcare: Modeling on an open dataset and identification of risk factors for disease prediction

Aim: This study aimed to compare circRNA expression profiles across different stages of hepatocellular carcinoma (HCC), to identify potential biomarkers, and to evaluate their applicability in clinical decision support through explainable artificial intelligence (XAI) methods. Materials and Methods: Open-access circRNA microarray datasets were analyzed, including normal liver tissue, non-metastatic HCC, and metastatic HCC samples. Differential expression analyses were performed using statistical and bioinformatics approaches, and findings were visualized through UMAP, volcano plots, and MD plots. Eleven distinct machine learning algorithms were applied to classify groups, and SHAP-based explainability techniques were employed to ensure transparent interpretation of model decisions. Results: Distinct circRNA signatures were identified for both metastatic and non-metastatic HCC groups. In metastatic cases, circRNAs with high logFC and SHAP scores emerged as potential biomarkers associated with metastatic progression, whereas non-metastatic cases revealed specific molecular patterns reflecting early-stage tumor biology. Machine learning models demonstrated robust performance with high accuracy and ROC-AUC values, while SHAP analysis highlighted biologically meaningful transcripts contributing to classification. Conclusion: This thesis demonstrates that circRNAs may serve as stage-specific molecular signatures of HCC and highlights the utility of XAI-driven approaches in biomarker discovery. The findings contribute to identifying reliable biomarker candidates that can be integrated into clinical decision support systems, presenting a novel framework that combines transcriptomic analysis with interpretable artificial intelligence in HCC research. Keywords: Explainable Artificial Intelligence, Biomarker, circRNA, Hepatocellular Carcinoma, Machine Learning, SHAP.