Machine-Learning Classification Models to Predict Liver Cancer with Explainable AI to Discover Associated Genes

Abstract

Hepatocellular carcinoma (HCC) is the primary liver cancer that occurs the most frequently. The risk of developing HCC is highest in those with chronic liver diseases, such as cirrhosis brought on by hepatitis B or C infection and the most common type of liver cancer. Knowledge-based interpretations are essential for understanding the HCC microarray dataset due to its nature, which includes high dimensions and hidden biological information in genes. When analyzing gene expression data with many genes and few samples, the main problem is to separate disease-related information from a vast quantity of redundant gene expression data and their noise. Clinicians are interested in identifying the specific genes responsible for HCC in individual patients. These responsible genes may differ between patients, leading to variability in gene selection. Moreover, ML approaches, such as classification algorithms, are similar to black boxes, and it is important to interpret the ML model outcomes. In this paper, we use a reliable pipeline to determine important genes for discovering HCC from microarray analysis. We eliminate redundant and unnecessary genes through gene selection using principal component analysis (PCA). Moreover, we detect responsible genes with the random forest algorithm through variable importance ranking calculated from the Gini index. Classification algorithms, such as random forest (RF), naïve Bayes classifier (NBC), logistic regression, and k-nearest neighbor (kNN) are used to classify HCC from responsible genes. However, classification algorithms produce outcomes based on selected genes for a large group of patients rather than for specific patients. Thus, we apply the local interpretable model-agnostic explanations (LIME) method to uncover the AI-generated forecasts as well as recommendations for patient-specific responsible genes. Moreover, we show our pathway analysis and a dendrogram of the pathway through hierarchical clustering of the responsible genes. There are 16 responsible genes found using the Gini index, and CCT3 and KPNA2 show the highest mean decrease in Gini values. Among four classification algorithms, random forest showed 96.53% accuracy with a precision of 97.30%. Five-fold cross-validation was used in order to collect multiple estimates and assess the variability for the RF model with a mean ROC of 0.95±0.2. LIME outcomes were interpreted for two random patients with positive and negative effects. Therefore, we identified 16 responsible genes that can be used to improve HCC diagnosis or treatment. The proposed framework using machine-learning-classification algorithms with the LIME method can be applied to find responsible genes to diagnose and treat HCC patients.