Kolmogorov-Arnold network model integrated with hypoxia risk for predicting PD-L1 inhibitor responses in hepatocellular carcinoma

Abstract

Hepatocellular carcinoma (HCC) is a leading cause of cancer-related deaths, with immunotherapy being a first-line treatment at the advanced stage and beyond. Hypoxia plays a critical role in tumor progression and resistance to therapy. This study develops and validates an artificial intelligence (AI) model based on publicly available genomic datasets to predict hypoxia-related immunotherapy responses. Based on the HCC-Hypoxia Overlap (HHO) and immunotherapy response to hypoxia (IRH) genes selected by differential expression and enrichment analyses, a hypoxia model was built and validated on the TCGA-LIHC and GSE233802 datasets, respectively. The training and test sets were assembled from the EGAD00001008128 dataset of 290 HCC patients, and the response and non-response classes were balanced using the Synthetic Minority Over-sampling Technique. With the genes selected via the minimum Redundancy Maximum Relevance and stepwise forward methods, a Kolmogorov–Arnold Network (KAN) model was trained. Support Vector Machine (SVM) combined the Hypoxia and KAN models to predict immunotherapy response. The hypoxia model was constructed using 10 genes (IRH and HHO). The KAN model with 11 genes achieved a test accuracy of 0.7. The SVM integrating the hypoxia and KAN models achieved a test accuracy of 0.725. The established AI model can predict immunotherapy response based on hypoxia risk and genomic factors potentially intervenable in HCC patients.