Int J Surg. 2025 Jun 5. doi: 10.1097/JS9.0000000000002641. Online ahead of print.
ABSTRACT
INTRODUCTION: Due to the high propensity for occult lymph node metastasis (OLNM) in Early-stage oral tongue squamous cell carcinoma (OTSCC), elective neck dissection has become standard practice for many patients with clinically node-negative (cT1-2 N0) disease, which may lead to overtreatment in some patients. Hence, accurate identification and prediction of OLNM are of great significance.
AIM: This study aimed to develop and validate an explainable machine learning (ML) model to predict OLNM in OTSCC.
METHODS: A total of 678 Early-stage OTSCC patients from multiple centers were enrolled and randomly classified into the derivation and external validation cohorts. The variables considered in this study primarily included clinicopathological characteristics associated with the occurrence of OLNM in OTSCC. Feature selection utilized multivariate logistic regression analysis and Lasso regression analysis. Meanwhile, 6 ML algorithms were employed to develop an OLNM diagnostic model, assessed with area under the curve (AUC), calibration curve, decision curve analysis (DCA), sensitivity, specificity, and validation cohorts. Moreover, the Shapley Additive exPlanations (SHAP) method was applied to rank the feature importance and interpret the final model.
RESULTS: In this study, 192 patients (34.7%) developed OLNM in the derivation cohort, while 38 patients (30.6%) developed OLNM in the external validation cohort. Through feature selection, 9 clinicopathological variables were identified as independent predictive factors for OLNM, and six ML models were developed based on these factors. Among the six evaluated ML models, the Random Forest (RF) model achieved the highest AUC (0.941, 95% CI: 0.907-0.975) for internal validation. External validation further confirmed the RF model’s effectiveness, yielding an AUC of 0.917 (95% CI: 0.868-0.967). The calibration curves also demonstrated a high level of concordance between the anticipated risk and the observed risk of the RF model. Additionally, this study also compared the RF model with the currently accepted traditional statistical methods, including depth of invasion (DOI) and tumor budding (TB), demonstrating superior prediction performance and greater clinical application value. Ultimately, an online computing platform (https://prediction-model-for-olnm.streamlit.app/) for this RF model is freely available to both clinicians and patients.
CONCLUSION: This study innovatively utilized 9 easily obtained clinicopathological features to construct an explainable RF model, providing a practical and reliable tool for predicting OLNM in Early-stage OTSCC. More importantly, it also provided interpretable results, thus overcoming the “impenetrable black box” of conventional ML models.
PMID:40479496 | DOI:10.1097/JS9.0000000000002641
