Machine Learning-Based Prediction of Lymph Node Metastasis and Volume Using Preoperative Ultrasound Features in...

A predictive model of cervical lymph node metastasis and metastasis volume was constructed based on a machine learning algorithm and ultrasound characteristics before surgery.

A retrospective analysis was conducted on 573 cases of PTC patients who underwent surgery in our institution, from 2017 to 2022. Patient demographic and clinical characteristics were systematically collected. Feature selection was performed using univariate analysis, Logistic regression (LR) analysis. Statistically significant variables were identified using a threshold of p < 0.05. Predictive models for cervical lymph node metastasis and metastatic volume in papillary thyroid carcinoma were constructed using advanced machine learning algorithms: K-Nearest Neighbors (KNN), Gradient Boosting Machine (XGBoost), and Support Vector Machine (SVM). Model performance was rigorously assessed using validation cohort data, evaluating area under the Receiver Operating Characteristic (ROC) curve, sensitivity, specificity, and accuracy.

In this retrospective study of 573 patients (320 had lymph node metastasis, 127 had small volume lymph node metastasis, and 193 had medium-volume lymph node metastasis). In the model predicting the neck lymph node metastasis, the Gradient Boosting method exhibited the best performance, with an area under the ROC curve of 0.784, sensitivity of 76.2%, specificity of 70.6%, and accuracy of 73.8%. In the model predicting the metastatic volume in neck lymph nodes for PTC, the Gradient Boosting method also demonstrated the best performance, with an area under the ROC curve of 0.779, sensitivity of 71.7%, specificity of 75.9%, and accuracy of 74.4%.

Machine learning-based predictive models integrating preoperative ultrasound features demonstrate robust performance in stratifying neck lymph node metastasis risk for PTC patients. These models optimize surgical planning by guiding lymph node dissection extent and individualizing treatment strategies, potentially reducing unnecessary extensive surgeries. The integration of advanced computational techniques with clinical imaging provides a data-driven paradigm for preoperative risk assessment in thyroid oncology.