Preoperative prediction of lymph node metastasis risk in papillary thyroid carcinoma based on multiple model comparisons

Abstract

The clinical necessity of lymph node dissection in papillary thyroid carcinoma (PTC) surgery remains contentious. This study compared four logistic regression (LR) models (with distinct feature selection strategies) and four machine learning (ML) models to preoperatively predict lymph node metastasis (LNM) risk in PTC patients, with emphasis on multidimensional evaluation and cross-populational generalizability. Data from 3,175 PTC patients (2021 cohort) were randomly split into training (70%) and testing (30%) subsets, with external validation performed using a Chinese (2024, n = 104) and a Canadian (2019-2022, n = 412) cohort. Twelve predictors were screened, and models were evaluated using metrics of discrimination (AUC), calibration (Brier Score), classification accuracy, and clinical utility. The prevalence of LNM was 34.48%, 36.54%, and 30.10% in the internal, Chinese, and Canadian cohorts, respectively. Among ML models, Random Forest achieved the highest internal AUC (0.767), whereas XGBoost demonstrated superior generalization (external AUCs: 0.785 and 0.725). LR models, particularly BestSubset-GLM, outperformed these ML models with an internal AUC of 0.770 and external AUCs of 0.831 and 0.785. Notably, BestSubset-GLM exhibited high specificity (0.86), precision (0.59), favorable calibration (Brier Score < 0.20), and robust clinical utility across the approximately 15-90% threshold probability range. Extrathyroidal extension, tumor size above 1.00 cm, younger age, and male gender were identified as key LNM risk factors. Bethesda classification and molecular aberrations were integrated into models. BestSubset-GLM balanced parsimony, interpretability, and generalizability, thereby supporting clinical decision-making through dynamic nomograms. Comprehensive evaluation beyond AUC is crucial.

Ähnliche Arbeiten