Explainable machine learning model for predicting paratracheal lymph node metastasis in cN0 papillary thyroid cancer

Area of Science:

• Oncology
• Surgical Oncology
• Medical Informatics

Background:

• Prophylactic paratracheal lymph node dissection in clinically node-negative (cN0) papillary thyroid carcinoma (PTC) is debated.
• Accurate prediction of paratracheal lymph node metastasis (PLNM) is crucial for surgical decision-making.

Purpose of the Study:

• To develop and validate an interpretable machine learning (ML) model for predicting PLNM in cN0 PTC patients.
• To compare the performance of ML models against traditional nomograms.
• To create a clinically applicable tool for surgical planning.

Main Methods:

• Retrospective analysis of 3213 cN0 PTC patients from a primary hospital and 533 from an external hospital.
• Development and validation of nine ML models using 10-fold cross-validation and hyperparameter tuning.
• Comparison of the best ML model (XGBoost) with a logistic regression-based nomogram using ROC curves, DCA, and calibration curves.

Main Results:

• The XGBoost model demonstrated superior predictive performance with AUCs of 0.935 (training), 0.857 (validation), and 0.775 (test set), outperforming the nomogram (AUCs 0.85, 0.844, 0.769).
• SHapley Additive exPlanations (SHAP) identified key predictive features.
• A web-based calculator was developed based on the top predictive features.

Conclusions:

• Machine learning offers a reliable method for predicting PLNM in cN0 PTC.
• The SHAP-based XGBoost model provides interpretable insights.
• The developed web-based calculator can assist surgeons in planning paratracheal lymph node dissection.