Objective:

To develop and validate an interpretable machine learning model for predicting distant metastasis in patients with papillary thyroid carcinoma (PTC), addressing the limitations of traditional prediction methods.

Key Findings:

• LightGBM model achieved a test-set AUC of 0.886 and accuracy of 0.887, indicating strong predictive capability.
• External validation AUC was 0.758, suggesting moderate generalizability.
• Extrathyroidal invasion and thyroglobulin antibody were identified as top predictors, highlighting their clinical relevance.
• The model demonstrated a 93.6% negative predictive value for excluding low-risk patients, enhancing clinical decision-making.

Interpretation:

The interpretable machine learning model significantly outperforms traditional predictors, enhancing clinical risk stratification and personalized treatment for PTC patients, and offers a promising alternative to existing methods.

Limitations:

• The model's performance may vary across different populations and settings, potentially limiting its applicability.
• Potential biases in the patient selection process could affect generalizability, necessitating careful consideration in diverse clinical environments.

Conclusion:

This study presents a robust machine learning approach for predicting distant metastasis in PTC, with implications for improving clinical decision-making and patient outcomes, and highlights the need for further research to validate its effectiveness in broader populations.