Аннотация
A comparative evaluation of machine learning models for predicting liver cancer risk was performed using an open structured clinical dataset. A Kaggle dataset was used in the study (N=5000): target variable Liver_cancer, 13 features (9 numerical, 4 categorical), with class imbalance (0: 78.22%, 1: 21.78%). Data preprocessing, model training and evaluation were carried out within a unified experimental protocol. Numerical variables were normalized via standardization; categorical variables were encoded in binary format. The comparative analysis included Logistic Regression, Support Vector Machine with an RBF kernel, k-Nearest Neighbors, Random Forest, and Gradient Boosting. All models were evaluated using accuracy, precision, recall (sensitivity), specificity, F1-score, and the area under the ROC curve (ROC-AUC). Gradient Boosting achieved the highest overall performance: ROC-AUC=0.999498, Accuracy=0.9754, Precision=0.998967, Sensitivity=0.887971, Specificity=0.999744, F1=0.940204. Although SVM (RBF) and Random Forest demonstrated stable discrimination, they underperformed compared to Gradient Boosting in terms of sensitivity and F1. The results indicate that selecting a model based on structured clinical features requires a comprehensive multi-metric assessment; key limitations include class imbalance and the absence of external validation.