Andreas Schindele, Anne Krebold, Ursula Heiß, Kerstin Nimptsch, Elisabeth Pfaehler, Christina Berr, Ralph A. Bundschuh, Thomas Wendler, Olivia Kertels, Johannes Tran-Gia, Christian H. Pfob, Constantin Lapa
- 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 197Purpose
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.…
