EVALUATION AND APPLICATION OF K-NEAREST NEIGHBORS ALGORITHM FOR PREDICTING CERVICAL LYMPH NODE METASTASIS IN PAPILLARY THYROID CARCINOMA

Abstract

There was no universally accepted surgical approach for treating clinically node-negative (cN0) papillary thyroid carcinoma (PTC) patients staged as T1-T2. Due to the low sensitivity (Sn) of ultrasound in preoperative lymph-node evaluation, cervical lymph-node metastases (LNM) often went undetected, potentially leading to persistent or recurrent disease requiring further treatment. These challenges underscored the need for accurate predictive tools to identify patients at higher LNM risk. To address this gap, a k-Nearest Neighbors (KNN)-model was constructed based on clinicopathological characteristics of 288 cN0 T1-T2 PTC patients who underwent total thyroidectomy, prophylactic central neck dissection, and sentinel lymph-node biopsy performed to identify lateral LNM. The KNN model exhibited promising performance with an area under the receiver operating characteristic curve of 0.72 and Sn, specificity (Sp), positive and negative predictive values (PPV, NPV), F1, and F2 scores of 98%, 27%, 56%, 93%, 72%, and 85%, respectively. These results demonstrated that the KNN model, with its high NPV, accurately identified patients without LNM, suggesting that additional therapy may not been necessary for these patients. Based on these results, a user-friendly web-application was developed utilizing the KNN model to enhance its usability in clinical practice. The application was then tested on 15 completely new PTC patients to validate its performance in real-world scenarios, demonstrating a Sn, Sp, PPV, NPV, accuracy, F1 and F2 score of 100%, 27%, 34%, 100%, 47%, 50%, 72%. These results confirm that the KNN-model maintains a high NPV in clinical setting, indicating its potential for adequate planning of future therapeutic interventions.