[Predictive value of a multimodal radiomics model for central lymph node metastasis in clinically node-negative papillary thyroid microcarcinoma based on machine learning].

OBJECTIVES

To develop and validate a machine learning-based multimodal radiomics model for predicting central lymph node metastasis (CLNM) in patients with clinically node-negative (cN0) papillary thyroid microcarcinoma (PTMC).

METHODS

A retrospective study was conducted on the clinical data of 532 consecutive cN0 PTMC patients who underwent surgery at the Department of Thyroid Surgery of the First People's Hospital of Changzhou and the Department of Thyroid and Breast Surgery of Suzhou Municipal Hospital between January 2022 and June 2024. Among them, 487 patients from the First People's Hospital of Changzhou were randomly assigned to a training set (n=352) or an internal validation set (n=135), while 45 patients from Suzhou Municipal Hospital served as an external validation set. Clinical feature screening involved collinearity analysis using variance inflation factors, followed by logistic regression to identify independent risk factors for CLNM. Radiomics features were extracted from ultrasound and CT images. An initial feature screening was performed using statistical tests (t-test or Mann-Whitney U test, P<0.05) along with mutual information analysis (score >0.015), followed by least absolute shrinkage and selection operator (LASSO) regression for key feature selection. Using the optimized feature set, four machine learning models were constructed: random forest, gradient boosting machine (GBM), support vector machine, and K-nearest neighbors. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), decision curve analysis, and SHapley Additive exPlanations (SHAP) method.

RESULTS

Logistic regression identified five clinical features independently associated with CLNM: age <55 years (OR=2.391, 95%CI: 1.072-5.334, P<0.05), coexisting Hashimoto's thyroiditis (OR=3.084, 95%CI: 1.474-6.453, P<0.01), maximum tumor diameter (OR=11.086, 95%CI: 2.881-48.378, P<0.01), monocyte count (OR=0.005, 95%CI: 0.001-0.044, P<0.01), and the lymphocyte-to-monocyte ratio (OR=0.564, 95%CI: 0.486-0.654, P<0.01). LASSO regression selected two key ultrasound and six key CT radiomics features. Among the four models, the GBM model based on multimodal feature fusion performed best, with AUC values of 0.975, 0.833, and 0.916, accuracies of 0.925, 0.748, and 0.863, specificities of 0.950, 0.800, and 0.881, and sensitivities of 0.900, 0.720, and 0.804 in the training, internal validation, and external validation sets, respectively. Decision curve analysis showed that the GBM model provided the highest net clinical benefit within the threshold probability range of 0.1-0.8. SHAP feature importance analysis revealed that the lymphocyte-to-monocyte ratio and monocyte count contributed most to CLNM prediction, followed by maximum tumor diameter and radiomics texture features.

CONCLUSIONS

The GBM-based multimodal radiomics model can accurately predict the risk of CLNM in patients with cN0 PTMC, which may facilitate individualized preoperative risk stratification and clinical descision-making.