A Personalized Decision-Support Framework for Predicting Immune Checkpoint Blockade Efficacy Using Ensemble Learning and Explainable AI

Introduction Predicting the efficacy of immune checkpoint blockade (ICB) therapies is crucial for advancing personalized medicine in oncology. As ICBs become a prominent treatment for various cancers, accurate prediction models that guide individualized decision-making are essential to improving patient outcomes. This study builds on the work presented in "Improved prediction of immune checkpoint blockade efficacy across multiple cancer types" by applying advanced machine learning techniques to enhance the predictive accuracy of ICB response. Our proposed decision-support framework incorporates ensemble learning and explainable AI (XAI) tools to offer more personalized predictions, aiding clinicians in making data-driven, patient-specific treatment decisions. Methods We employed a dataset consisting of ۱,۴۷۹ patients spanning ۱۶ different cancer types, as described in the original study. Four machine learning models—RandomForest, Logistic Regression, GradientBoosting classifier, and XGBoost—were evaluated for their predictive performance. Among these methods, XGBoost achieved the highest individual accuracy, recording an accuracy of ۸۶.۲%, significantly outperforming the authors’ reported accuracy. To further refine the predictions, we utilized a Tsetlin machine-based ensemble learning framework. In this approach, classifier outputs (labels) served as input to the Tsetlin machine, which aggregated the individual classifiers’ predictions, improving the robustness of the model. Feature importance and model interpretability were analyzed using SHAP (Shapley Additive Explanations) and LIME (Local Interpretable Model-agnostic Explanations), allowing us to identify critical variables driving predictions. Results Our proposed ensemble learning framework achieved a weighted average accuracy of ۹۲.۵%, representing a substantial improvement over the best-reported accuracy of ۷۵.۷۶% in the original study. The XGBoost model, as the most effective single classifier, was enhanced by the ensemble method, which balanced the strengths of all models for more consistent predictions. The integration of SHAP and LIME tools revealed key features that strongly influenced the predicted efficacy of ICB therapies, offering actionable insights for clinical applications. Importantly, the framework’s personalized predictions enable oncologists to assess the likelihood of treatment success based on individual patient profiles. Conclusion Our study demonstrates the potential of an ensemble learning framework, combined with explainable AI tools, to significantly improve the accuracy of ICB efficacy predictions across diverse cancer types. By leveraging patient-specific data, our framework provides personalized decision-making support that can be crucial in tailoring immunotherapy treatments to individual patients. The ۸۶.۲?curacy achieved with XGBoost, alongside the Tsetlin machine’s ability to optimize ensemble predictions to ۹۲.۵%, illustrates the power of combining machine learning with personalized medicine approaches. This decision-support framework has the potential to contribute significantly to clinical oncology by improving treatment outcomes and optimizing immunotherapy strategies.

کلیدواژه ها:

Immune Checkpoint Blockade ، Personalized Medicine ، Tsetlin machine ، XAI ، Immunotherapy