Marina Popović Krneta1*, Nemanja Krajčinović2, Zoran Bukumirić3, and Miljana Tanić4,5
1Department of nuclear medicine, Institute for oncology and radiology of Serbia, Pasterova 14, 11000 Belgrade, Serbia
2Faculty of Technical Sciences, University of Novi Sad, Trg Dositeja Obradovića 6, 21000 Novi Sad, Serbia
3Institute of Medical Statistics and Informatics, Faculty of Medicine, University of Belgrade, Dr Subotića 8, 11000 Belgrade, Serbia
4Department of Experimental Oncology, Institute for Oncology and Radiology of Serbia, Pasterova 14, 11000 Belgrade, Serbia
5UCL Cancer Institute, 72 Huntley St London WC1E 6DD, United Kingdom
marina.popovic1989 [at] gmail.com
Abstract
This study aimed to assess and compare four machine learning (ML) based classifiers in predicting occult cervical lymph node metastasis (LNM) in clinically node-negative (cN0), T1-T2 papillary thyroid carcinoma (PTC) patients.
The study cohort included 288 PTC patients who underwent total thyroidectomy and prophylactic central neck dissection with sentinel lymph node biopsy performed for lateral LNM identification. The classifiers, namely k-Nearest Neighbor (k-NN), Support Vector Machines, Decision Tree, and Logistic Regression were developed using patients’ demographic and clinicopathological variables. Evaluation metrics such as area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive and negative predictive values (PPV and NPV), accuracy, and F1 and F2 scores were utilized for model comparison.
The final ML classifier was selected based on achieving the highest specificity and the lowest degree of overfitting while maintaining a sensitivity of 95%. Among the evaluated models, the k-NN emerged as the best-performing, demonstrating an AUC of 0.72. The sensitivity, specificity, PPV, NPV, F1, and F2 scores were 98%, 27%, 56%, 93%, 72%, and 85%, respectively Furthermore, a web application was developed allowing users to predict the potential of cervical LNM and explore possibilities for further model development.
The k-NN classifier incorporating patients’ clinicopathological information shows potential in predicting LNM. Improved prediction models are necessary to identify patients at higher risk of LNM, guiding appropriate postsurgical treatment for high-risk individuals while minimizing unnecessary interventions for low-risk patients.
Keywords: machine learning, papillary thyroid carcinoma; lymph node metastasis
Acknowledgement: This research was supported by the Serbian Ministry of Science, Innovation and Technological development (451-03-47/2023-01/200043).
