Construction of predictive models for contralateral occult thyroid carcinoma and central lymph node metastasis in unilateral papillary thyroid carcinoma using machine learning.

Background: This study aimed to develop predictive models based on preoperative clinicopathological and imaging features to accurately assess the individual risk of contralateral occult thyroid carcinoma (OTC) and determine the number of central lymph node metastasis (CLNM) in patients with unilateral papillary thyroid carcinoma, thereby providing actionable guidance for surgical planning. Methods: Seven widely-used machine learning algorithms were employed to develop predictive models. Hyperparameter tuning was performed via cross-validation in combination with grid search. The models were subsequently trained and evaluated by using the optimal hyperparameter combinations. To facilitate comparative analysis, ROC curves, calibration curves were generated and DCA was performed. The optimal model was then selected on the basis of this comprehensive evaluation. Furthermore, a clinical prediction model was constructed utilizing the significant predictors identified. Results: The logistic regression model was identified to be the optimal predictive model. For the clinical prediction model of OTC, the following independent variables were incorporated: body mass index, and ultrasonographic findings, including capsular disruption, number of malignant nodules within a unilateral lobe, sum of the longest diameter (SLD) of tumors, and the presence of isthmic malignant nodule(s). This model yielded an area under the ROC curve (AUC) of 0.74 and 0.70 in the training and validation cohorts, respectively. For the clinical prediction model of ≥5 CLNM, the incorporated independent variables included: age, sex, chronic lymphocytic thyroiditis, and ultrasonographic features covering malignant nodules located near the isthmus, SLD, capsular disruption, and calcification. This model produced an AUC of 0.75 and 0.71 in the training and validation cohorts, respectively. Decision curve analysis indicated that clinical interventions guided by the two models could provide net benefit within threshold probability ranges of 10% to 90% and 10% to 70% for patients with PTC. And the calibration curves demonstrated a good agreement between model predictions and actual observations. Conclusion: This study developed and validated clinical prediction models to estimate the risk of contralateral OTC and the presence of ≥5 CLNM in patients with unilateral PTC. These models were designed to prevent overtreatment in low-risk patients while providing evidence-based guidance for decision-making about treatment choice in high-risk patients.