SUN-310 PMRA and Shapley-Based Machine Learning for Predicting Lymph Node Metastasis in Central Subregions of cN0 PTMC:A Prospective Multicenter Validation and Development of a Web Calculator

Le résumé fourni par la source

Abstract Disclosure: J. zhou: None. Background: The management of clinically node-negative(cN0) papillary thyroid microcarcinoma (PTMC) is complicated by high rates of occult lymph node metastasis (LNM). The optimal surgical extent remains controversial due to difficulties in preoperative identification of metastatic nodes.We aimed to develop and validate a comprehensive prediction model for central LNM and its subregions by evaluating both machine learning (ML) algorithms and traditional nomograms using Probability-based Ranking Model Approach (PMRA). Methods: We conducted a prospective multicenter study involving 4,882 patients across 3 hospitals (2016-2023). After applying inclusion criteria, 1,953 patients from the primary center were allocated to model train and test (7:3 ratio). External validation included prospective cohorts of 286 and 176 patients from two independent centers.13 ML algorithms and traditional nomogram models were systematically evaluated using PMRA.We compared models using preoperative features alone versus those incorporating both preoperative and intraoperative frozen section pathology data. Feature selection utilized six methods, with L1-based selection proving optimal for most predictions.Model interpretability was enhanced through SHapley Additive exPlanations (SHAP) visualization. Results: Among 2,415 total patients, 785 (32.5%) had central LNM. PMRA analysis identified XGBoost as the superior algorithm among all evaluated models. Using preoperative features alone, both traditional nomogram (AUC 0.77 train/0.717 test) and XGBoost (AUC 0.831 train/0.747 test) showed moderate performance. However, incorporating intraoperative frozen section pathology significantly improved XGBoost's predictive performance: prelaryngeal LNM (AUC 0.918 train/0.866 val2), pretracheal LNM (AUC 0.940 train/0.855 val2), paratracheal LNM (AUC 0.886 train/0.895 val2), contralateral paratracheal LNM (AUC 0.920 train/0.903 val1), and recurrent laryngeal nerve posterior LNM (AUC 0.933 train/0.921 val2). SHAP analysis revealed that paratracheal LNM ratio was the strongest predictor for prelaryngeal and pretracheal LNM, while pretracheal LNM strongly predicted paratracheal and recurrent laryngeal nerve posterior LNM. A web-based calculator (http://121.41.36.155:9006/static/html/index.html) demonstrated robust performance in prospective external validation.. Conclusions: Our study demonstrates that the XGBoost model not only outperforms traditional nomograms and other ML algorithms but also achieves significantly enhanced prediction accuracy when combining preoperative and intraoperative frozen section features.The model's high accuracy, interpretability, and successful prospective validation, combined with its accessible web interface, provide clinicians with a reliable tool for personalized surgical planning. Presentation: Sunday, July 13, 2025

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.