Model predicts thyroid cancer in hard-to-reach lymph nodes

A machine learning (ML) model predicted metastasis to lymph nodes around the entrance point of the recurrent laryngeal nerve in patients with clinically lymph node-negative papillary thyroid carcinoma, according to a study published in Frontiers in Endocrinology.

Papillary thyroid carcinoma (PTC) accounts for approximately 90% of thyroid cancers and frequently metastasizes to lymph nodes. Reported metastasis rates range from 20% to 90%, and recurrence rates approach 30%.

Central lymph node dissection (CLND) is commonly performed in patients with clinical lymph node-negative (cN0) PTC. Within the central compartment, the lymph nodes around the entrance point of the recurrent laryngeal nerve (LN-epRLN) represent a specific anatomical subregion. Dissection in this area can be technically challenging because of its proximity to the recurrent laryngeal nerve (RLN), which increases the risk of complications such as RLN injury and hypoparathyroidism.

Analysis From 1,800 Patients

Researchers conducted a retrospective study to develop ML models that predict metastasis to LN-epRLN (LNM-epRLN). The analysis included records from 1,800 patients with cN0 PTC treated between June 2023 and September 2024 who underwent CLND with systematic LN-epRLN sampling. Histopathologic evaluation was used to determine metastatic status.

Among the 1,800 patients initially reviewed, LN-epRLN tissue was identified in 149 patients (8%). Metastasis in this region occurred in 29 patients, corresponding to a metastasis rate of about 19%. The cohort included 115 female and 34 male patients, with a mean age of about 43 years.

Univariate analysis identified several clinicopathologic factors associated with LNM-epRLN. Compared with patients without metastasis, those with LNM-epRLN had larger tumors and were more frequently aged 45 years or younger. Patients with LNM-epRLN also had higher postoperative drainage volumes.

LNM-epRLN was associated with metastasis in several lymph node regions. These included total central lymph node metastasis, lateral lymph node metastasis, and contralateral paratracheal lymph node metastasis. Associations were also observed with metastasis in the prelaryngeal, pretracheal, and ipsilateral paratracheal lymph nodes and in lymph nodes posterior to the right recurrent laryngeal nerve.

The investigators evaluated 9 supervised ML models to predict LNM-epRLN: Logistic Regression, Decision Tree, Random Forest, Extra Trees, XGBoost, Support Vector Machine, k-Nearest Neighbors, Neural Network, and Gaussian Naive Bayes. Models were optimized using grid search and 10-fold cross-validation.

Among these approaches, the Random Forest model showed the strongest predictive performance. In the testing set, the model achieved an area under the receiver operating characteristic curve of 0.92.

Researchers applied Shapley Additive Explanations (SHAP) analysis to interpret the model. Feature importance ranking identified several predictors of LNM-epRLN, including the ratio and number of total central lymph node metastases, the number and ratio of pretracheal lymph node metastases, tumor size, age, and measures related to paratracheal lymph node metastasis.

SHAP analysis showed that the ratio of total central lymph node metastasis was the most influential predictor of LNM-epRLN.

Researchers also developed a simplified model using 7 predictors: total central lymph node metastasis number and ratio, pretracheal lymph node metastasis number and ratio, tumor size, age, and paratracheal lymph node metastasis number. The simplified model retained strong predictive performance.

The study had several limitations. It was conducted at a single center and used retrospective data, which may limit generalizability. In addition, LNM-epRLN occurred in a relatively small number of patients (n = 29). The model’s sensitivity in the test set was 0.50.

The researchers concluded that the interpretable Random Forest–based ML model may assist surgeons in predicting LNM-epRLN and support individualized intraoperative decision-making about LN-epRLN dissection while balancing oncologic completeness with preservation of RLN function.

The researchers reported no commercial or financial conflicts of interest.