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

Area of Science:

• Artificial Intelligence
• Oncology
• Medical Diagnostics

Background:

• Accurate early-stage tumour diagnosis is vital for patient prognosis.
• Machine learning (ML) offers advanced tools for diagnostic support.
• Morphometric data analysis is key for tumour characterization.

Purpose of the Study:

• To compare base classifiers and ensemble models for tumour type prediction using morphometric data.
• To evaluate the performance and interpretability of CatBoost, Voting, and Stacking classifiers.
• To leverage the SHAP framework for feature importance analysis in tumour diagnosis.

Main Methods:

• Comparison of base and ensemble ML models (CatBoost, Voting, Stacking).
• Evaluation using standard diagnostic metrics and confusion matrices.
• Application of the SHAP framework for model interpretability and feature importance.

Main Results:

• CatBoost provided explainable results, highlighting tumour size and border irregularity.
• Voting Classifier enhanced stability, reducing critical false negative errors.
• Stacking Classifier achieved superior performance by minimizing both false positive and false negative classifications.

Conclusions:

• Interpretable ensemble ML methods are valuable for neuro-oncology diagnostics.
• The methodology enhances AI reliability and transparency in medical applications.
• Potential applications include treatment monitoring and predicting tumour recurrence risk.