Prognostication of differentiated thyroid cancer recurrence: An explainable machine learning approach

Journal article

G. M. Idroes, T. R. Noviandy, Ghalieb Mutig Idroes, Irsan Hardi, T. F. Duta, Lama MA. Hamoud, Hala T. Al-Gunaid
Narra X, 2024

Semantic Scholar DOI

Cite

APA Click to copy
Idroes, G. M., Noviandy, T. R., Idroes, G. M., Hardi, I., Duta, T. F., Hamoud, L. M. A., & Al-Gunaid, H. T. (2024). Prognostication of differentiated thyroid cancer recurrence: An explainable machine learning approach. Narra X.

Chicago/Turabian Click to copy
Idroes, G. M., T. R. Noviandy, Ghalieb Mutig Idroes, Irsan Hardi, T. F. Duta, Lama MA. Hamoud, and Hala T. Al-Gunaid. “Prognostication of Differentiated Thyroid Cancer Recurrence: An Explainable Machine Learning Approach.” Narra X (2024).

MLA Click to copy
Idroes, G. M., et al. “Prognostication of Differentiated Thyroid Cancer Recurrence: An Explainable Machine Learning Approach.” Narra X, 2024.

BibTeX Click to copy

@article{g2024a,
  title = {Prognostication of differentiated thyroid cancer recurrence: An explainable machine learning approach},
  year = {2024},
  journal = {Narra X},
  author = {Idroes, G. M. and Noviandy, T. R. and Idroes, Ghalieb Mutig and Hardi, Irsan and Duta, T. F. and Hamoud, Lama MA. and Al-Gunaid, Hala T.}
}

Abstract

Differentiated thyroid cancer (DTC) generally has a favorable prognosis, but recurrence remains a concern for a subset of patients, highlighting the need for accurate predictive tools. While traditional methods, such as the American Thyroid Association (ATA) guidelines, are widely used, they may not fully capture the complex patterns in clinical data. To address this, we developed a machine learning model using LightGBM and enhanced its interpretability with SHAP (SHapley Additive exPlanations). Our model, trained on data from 383 DTC patients, identified response to initial therapy as the most significant predictor of recurrence, alongside age and risk level. The model achieved an accuracy of 93.51%, with precision and sensitivity of 94.23% and 96.08%, respectively, using only five key features selected through Recursive Feature Elimination (RFE). SHAP analysis provided clear insights into how these features influenced predictions, offering a transparent and interpretable approach to risk stratification. These results highlight the potential of explainable machine learning to improve recurrence prediction, support personalized care, and build clinician trust, while laying the groundwork for further validation in diverse populations.