Application and validation of the machine learning-based multimodal radiomics model for preoperative prediction of...

Papillary thyroid carcinoma (PTC) frequently develops lateral lymph node metastasis (LLNM) in 12.6%-32.8% of patients, increasing recurrence risk and mortality. Current diagnostic methods show significant limitations, with occult LLNM rates of 41.0%-51.7% requiring secondary surgeries. This study aims to develop and validate a multimodal prediction model integrating clinical, ultrasound, and CT radiomics features for accurate preoperative LLNM prediction in PTC patients.

Clinical data, ultrasound and CT images from 799 PTC patients were retrospectively analyzed (524 training, 225 internal validation, 50 external validation). Clinical features were selected through logistic regression after collinearity analysis. A total of 874 ultrasound radiomics features and 1433 CT radiomics features were extracted and selected using LASSO regression. Four machine learning models were constructed and compared, with model interpretability explored using SHAP and LIME analyses.

Logistic regression identified five independent clinical risk factors: maximum tumor diameter, multiple lesions, upper pole location, decreased monocyte count, and lower lymphocyte-to-monocyte ratio (LMR). LASSO regression selected 4 key ultrasound features and 11 key CT features. The Gradient Boosting Machine (GBM) model demonstrated superior performance, with areas under the curve of 0.973, 0.803, and 0.975, and accuracies of 0.914, 0.725, and 0.900 in the training, internal validation, and external validation sets respectively. Decision curve analysis confirmed the GBM model's highest net clinical benefit. SHAP analysis identified LMR as the most important predictor.

The GBM-based multimodal prediction model accurately predicts LLNM in PTC patients preoperatively. This non-invasive, interpretable tool enables individualized risk assessment, potentially reducing missed metastases requiring secondary surgery, thereby supporting precise treatment decisions in PTC management.