Title : Development and validation of a clinical decision-support model for preoperative individualized assessment of high-volume central lymph node metastasis in older adults (over 55 years) with papillary thyroid carcinoma
Abstract:
Background: Elderly patients with papillary thyroid carcinoma (PTC) constitute a unique subgroup in clinical practice, often exhibiting more aggressive tumor characteristics. Although current guidelines offer few specific recommendations for patients aged ≥55 years, it is vital to identify high-volume central lymph node metastasis (HVCLNM) before surgery to guide personalized treatment. This study focused on creating and validating a machine learning (ML) model based on real-world clinical data to predict HVCLNM preoperatively in this understudied group, thereby aiding clinical decision-making.
Methods: We retrospectively examined electronic medical records from a real-world cohort of 644 elderly PTC patients. To address class imbalance, we applied the synthetic minority oversampling technique for nominal and continuous features (SMOTE-NC) during data preprocessing. Eleven critical risk factors were identified by combining univariate logistic regression, LASSO, Boruta, mRMR, and Random Forest selection methods. We then developed and systematically compared ten machine learning algorithms. Model performance was assessed using the area under the curve (AUC), and SHAP analysis was used to interpret and visualize feature importance. The final model was integrated into an interactive web-based application.
Results: Among the ten ML models evaluated, the CatBoost classifier achieved the highest discriminative performance, with AUCs of 0.864 (95% CI: 0.834–0.894) in the training set, 0.721 (95% CI: 0.627–0.815) in the internal test set, and 0.737 (95% CI: 0.646– 0.828) in the external validation set. SHAP analysis identified the most influential preoperative predictors and their interactions, improving model interpretability. A clinical heatmap further demonstrated the model’s ability to stratify patients effectively. The web tool allows clinicians to input patient variables for personalized risk assessment.
Conclusions: Using real-world data, this study developed and validated an interpretable machine learning framework to predict preoperative HVCLNM in elderly PTC patients. The model showed stable performance across internal and external cohorts and was converted into an easy-to-use web application for clinical use. This tool provides a valuable resource for preoperative assessment, supporting personalized clinical decisions in a population with limited evidence-based guidance.
