Interpretable artificial intelligence (AI) for cervical cancer risk analysis leveraging stacking ensemble and expert knowledge

Objectives This study develops a machine learning (ML)-based cervical cancer prediction system emphasizing explainability. A hybrid feature selection method is proposed to enhance predictive accuracy and stability, alongside evaluation of multiple classification algorithms. The integration of explainable artificial intelligence (XAI) techniques ensures transparency and interpretability in model decisions. Methods A hybrid feature selection approach combining correlation-based selection and recursive feature elimination is introduced. An ensemble model integrating random forest, extreme gradient boosting, and logistic regression is compared against eight classical ML algorithms. Generative artificial intelligence methods, such as variational autoencoders and generative teaching networks, were evaluated but showed suboptimal performance. The research integrates global and local XAI techniques, including individual feature contributions and tree-based explanations, to interpret model decisions. The effects of feature selection and data balancing on classification performance are examined to stabilize precision, recall, and F1 scores. Classical ML models without preprocessing achieve 95-96% accuracy but exhibit instability. Results The proposed feature selection and data balancing strategies significantly enhance classification stability, creating a robust predictive model. The ensemble model achieves 98% accuracy with an area under the curve of 99.50%, outperforming other models. Domain experts validate critical contributing features, confirming practical relevance. Incorporating domain knowledge with XAI techniques significantly increases transparency, making predictions interpretable and trustworthy for clinical use. Conclusion Hybrid feature selection combined with ensemble learning substantially improves cervical cancer prediction accuracy and reliability. The integration of XAI techniques ensures transparency, supporting interpretability and trustworthiness, demonstrating significant potential in clinical decision-making.