Explainable Thyroid Cancer Diagnosis Through Two-Level Machine Learning Optimization with an Improved Naked Mole-Rat...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Oncology
• Bio-inspired Computing

Background:

• Machine learning (ML) shows promise for cancer diagnostics.
• Enhancing ML models is crucial for improving diagnostic accuracy.
• Thyroid tumor malignancy assessment requires robust diagnostic tools.

Purpose of the Study:

• To improve thyroid tumor malignancy classification using ML.
• To apply the naked mole-rat algorithm (NMRA) for ML model optimization.
• To evaluate the performance of NMRA-enhanced ML classifiers.

Main Methods:

• Utilized a 2022 dataset from Shengjing Hospital (1232 records, 19 features).
• Applied 10 ML classifiers (e.g., XGBoost, LightGBM, random forest).
• Employed NMRA for parameter optimization and feature selection in classifiers.

Main Results:

• The NMRA-optimized LightGBM model achieved the highest accuracy (81.82%) and F1-score (86.62%).
• NMRA effectively optimized parameters and selected relevant features for improved classification.
• SHAP values provided explainability for the LightGBM model's decisions.

Conclusions:

• NMRA significantly enhances ML model performance for thyroid tumor malignancy diagnosis.
• Optimized LightGBM offers a powerful tool for accurate cancer diagnostics.
• Explainable AI (SHAP) is valuable for understanding diagnostic model behavior.