Application of Decision Tree Algorithms to Predict Central Lymph Node Metastasis in Well-differentiated Papillary Thyroid Carcinoma Based on Multimodal Ultrasound Parameters: a Retrospective Study

Background: Prophylactic central neck dissection (pCND) in patients with well-differentiated primary papillary thyroid carcinoma (PTC) has become controversial. Several attempts have been made to predict central compartment lymph node metastasis (CLNM) based on clinical and conventional ultrasonic parameters. This study aimed to develop a decision tree (DT) model for predicting the risk of CLNM in patients with PTC based on clinical and preoperative multimodal ultrasound (US) characteristics.

Methods: A total of 148 PTC nodules confirmed by surgical pathology at Beijing Tiantan Hospital were retrospectively analyzed. All nodules underwent multimodal US examinations preoperatively from January 2020 to September 2021. Correlation analysis of CLNM with clinical characteristics as well as multimodal US parameters of PTC lesions based on gray-scale US, color Doppler flow imaging (CDFI), superb microvascular imaging (SMI), contrast-enhanced ultrasound (CEUS), and shear wave elastography (SWE) technology was carried out. Finally, the chi-squared automatic interaction detector (CHAID) with a 10-fold cross-validation was used to establish DTs for CLNM prediction. The area under the curve was calculated to compare the predictive performance.

Results: Univariate analysis indicated that CLNM was positively correlated with thyroglobulin level, maximum size, taller-than-wide, the number of microcalcifications greater than or equal to 5, contact capsule, abnormal cervical lymph node on conventional US, noncentripetal perfusion, delayed clearance, the average shear wave velocity (SWV mean), and the SWV ratio (P<0.05). The multimodal US DT based on taller-than-wide, contact capsule, abnormal cervical lymph node on conventional US, and centripetal enhancement as independent variables showed good discrimination: the sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve were 80.0%, 76.7%, 78.4%, and 0.837 [95% confidence interval (CI): 0.771-0.902]. There was a significant difference between the multimodal and conventional US DTs (P=0.009).

Conclusions: Our results indicated that the DT based on the preoperative multimodal US characteristics of PTCs has a reasonable predictive ability for CLNM and can be conveniently used for clinical decision-making of individualized treatment in patients with well-differentiated PTC.