Interpretable ensemble learning for tumor-type prediction with a SHAP-based evaluation of CatBoost and voting classifiers

Accurate early-stage diagnosis of tumours is crucial for improving patient prognosis. Modern machine learning techniques provide advanced and effective tools to support this process. In this study, both base classifiers and ensemble models were compared in the task of predicting tumour type from morphometric data. Particular attention was given to the CatBoost model as well as the Voting and Stacking classifiers. Model performance was evaluated comprehensively using standard diagnostic metrics and confusion matrices. To enhance interpretability, the SHAP framework was employed to assess the contribution of individual features to model predictions. The CatBoost model stood out due to its ability to provide explainable results through feature importance analysis and SHAP values, which highlighted the key role of parameters related to tumour size and border irregularity. The Voting Classifier improved stability and reduced variance, significantly lowering the number of false negative errors - a factor of particular clinical importance in oncological diagnostics. The Stacking Classifier achieved the highest overall predictive performance, minimising both false positive and false negative classifications by integrating heterogeneous base learners with a meta-classifier. The findings of this study confirm that interpretable ensemble methods represent a valuable approach to supporting the diagnostic process in neuro-oncology. The methodology applied here not only increases the reliability and transparency of artificial intelligence systems but also demonstrates potential for application in treatment monitoring and in predicting the risk of tumour recurrence.