Interpretable machine learning for thyroid cancer recurrence predicton: Leveraging XGBoost and SHAP analysis.

European journal of radiology [Eur J Radiol] 2025 May; Vol. 186, pp. 112049. Date of Electronic Publication: 2025 Mar 14.

Competing Interests: Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Purpose: For patients suffering from differentiated thyroid cancer (DTC), several clinical, laboratory, and pathological features (including patient age, tumor size, extrathyroidal extension, or serum thyroglobulin levels) are currently used to identify recurrence risk. Validation and potential adjustment of their individual and combined prognostic values using a large patient cohort with several years of follow-up might improve the correct identification of patients at risk.
Methods: In this retrospective study, we developed an XGBoost model using clinical and biomarker features for accurate DTC recurrence prediction using a cohort of 1228 consecutive patients (965 papillary, and 263 follicular) that were treated at the Department of Nuclear Medicine at University Hospital Augsburg between 1976 and 2010. The dataset was split into 982 patients for model training, and 246 for independent testing. From the 982 patients, 200 different random combinations of 785 training and 197 validation patients were conducted. To identify critical risk factors and understand the model's decision-making process, we conducted Shapely Additive exPlanations (SHAP) analysis.
Results: The XGBoost model achieved an AUROC of 0.84 (95 % CI: 0.84-0.86; SD: 0.08), sensitivity of 0.79 (95 % CI: 0.77-0.81; SD: 0.17), and specificity of 0.78 (95 % CI: 0.77-0.79; SD: 0.04) on the validation datasets, and an AUROC of 0.88 (sensitivity 0.83, specificity 0.80) on the independent test set. Tumor size, maximal thyroglobulin values within six months after thyroidectomy, and maximal thyroglobulin antibody levels within 12 to 24 months after thyroidectomy were the most important factors. SHAP dependence plots suggested new recurrence risk thresholds for a tumor size of 25 mm, maximal serum thyroglobulin levels of 3 and 10 ng/mL, respectively, and maximal thyroglobulin antibody levels of 120 IU/mL.
Conclusion: Our XGBoost model, supported by SHAP analysis empowers clinicians with interpretable insights and defined risk thresholds and could facilitate informed decision-making and patient-centric care.
(Copyright © 2025 The Author(s). Published by Elsevier B.V. All rights reserved.)

English

Publisher: Elsevier Science Ireland Ltd Country of Publication: Ireland NLM ID: 8106411 Publication Model: Print-Electronic Cited Medium: Internet ISSN: 1872-7727 (Electronic) Linking ISSN: 0720048X NLM ISO Abbreviation: Eur J Radiol Subsets: MEDLINE

Boosting Machine Learning Algorithms* , Neoplasm Recurrence, Local*/epidemiology , Neoplasm Recurrence, Local*/diagnosis , Thyroid Neoplasms*/epidemiology , Thyroid Neoplasms*/diagnosis , Thyroid Neoplasms*/diagnostic imaging, Adult ; Aged ; Female ; Humans ; Male ; Middle Aged ; Biomarkers, Tumor/blood ; Germany/epidemiology ; Machine Learning ; Prognosis ; Reproducibility of Results ; Retrospective Studies ; Risk Assessment ; Risk Factors ; Sensitivity and Specificity

20250515