Abstract / Summary
Abstract Reliable predictive models for ovarian cancer are crucial for early diagnosis and improved clinical decision making. Traditional diagnostic methods frequently lack sensitivity and specificity, requiring advanced machine learning approaches for better results. In this study, a novel model is developed based on the combination of a multi-objective genetic algorithm with pareto optimization and an enhanced light gradient boosting model, referred as GAPO-Enhanced LightGBM, to improve the prediction of ovarian cancer. Human epididymis protein 4, cancer antigen 125, carcinoembryonic antigen, neutrophil count, alphafeto protein, aspartate aminotransferase, lymphocyte count, mean platelet volume, globulin, and menopause were found to emerge through GAPO-driven feature selection. The performance of GAPO-Enhanced LightGBM surpasses traditional models, accomplishing an accuracy of 0.92, precision of 0.90, recall of 0.96, specificity of 0.88 and F1 score 0.93. Uncertainty analysis reveals that feature selection significantly impacts model performance, with AUC decreasing from 0.98 to 0.74 as key biomarkers are excluded. An explainable artificial intelligence technique, shapley additive explanation is used to make feature selection more understandable. The findings establish GAPO-Enhanced LightGBM as a scalable, interpretable and clinically viable artificial intelligence-driven framework to assess the risk of ovarian cancer, supporting early intervention and personalized treatment strategies.