Improving appendix cancer prediction with SHAP-based feature engineering for machine learning models: a prediction
1Ewha Womans University College of Medicine, Seoul, Korea.
Purpose:
This study aimed to leverage Shapley additive explanation (SHAP)-based feature engineering to predict appendix cancer. Traditional models often lack transparency, hindering clinical adoption. We propose a framework that integrates SHAP for feature selection, construction, and weighting to enhance accuracy and clinical relevance.
Methods:
Data from the Kaggle Appendix Cancer Prediction dataset (260,000 samples, 21 features) were used in this prediction study conducted from January through March 2025, in accordance with TRIPOD-AI guidelines. Preprocessing involved label encoding, SMOTE (synthetic minority over-sampling technique) to address class imbalance, and an 80:20 train-test split. Baseline models (random forest, XGBoost, LightGBM) were compared; LightGBM was selected for its superior performance (accuracy=0.8794). SHAP analysis identified key features and guided 3 engineering steps: selection of the top 15 features, construction of interaction-based features (e.g., chronic severity), and feature weighting based on SHAP values. Performance was evaluated using accuracy, precision, recall, and F1-score.
Results:
Four LightGBM model configurations were evaluated: baseline (accuracy=0.8794, F1-score=0.8691), feature selection (accuracy=0.8968, F1-score=0.8860), feature construction (accuracy=0.8980, F1-score=0.8872), and feature weighting (accuracy=0.8986, F1-score=0.8877). SHAP-based engineering yielded performance improvements, with feature weighting achieving the highest precision (0.9940). Key features (e.g., red blood cell count and chronic severity) contributed to predictions while maintaining interpretability.
Conclusion:
The SHAP-based framework substantially improved the accuracy and transparency of appendix cancer predictions using LightGBM (F1-score=0.8877). This approach bridges the gap between predictive power and clinical interpretability, offering a scalable model for rare disease prediction. Future validation with real-world data is recommended to ensure generalizability.
