Development and Validation of an Interpretable Machine Learning Model for Predicting Distant Metastasis in Tongue Squamous Cell Carcinoma: A Multicentre Study

Comprehensive Summary

This retrospective cohort study aimed to develop and validate a machine learning model (MLM) to predict distant metastasis in postoperative tongue squamous cell carcinoma (TSCC). Distant metastasis in TSCC has a poor prognosis, with a median overall survival typically of about 10 months. Early identification of patients with high risk of distant metastasis is vital to optimizing postoperative surveillance and guiding treatment; however, current assessment of this risk fails to capture the multifactorial nature of tumor metastasis. Due to this gap, the authors aimed to develop MLMs that could consider the complexity of this pathology. This study analyzed 1,091 patients, with inclusion criteria being over 18 years old, confirmed TSCC, and no distant metastasis prior to surgery. Patients were excluded if they had prior head or neck treatment, received nonsurgical therapy, or if they died from non-TSCC causes within 2 years of treatment. Demographic variables and postoperative pathological features were collected as well. The primary endpoint was defined as the occurrence of distant metastasis during the two-year follow-up period after curative surgery. Six ML algorithms were tested: Light Gradient Boosting Machine (LightGBM), logistic regression (LR), random forest (RF), Extreme Gradient Boosting (XGBoost), k-Nearest Neighbours (KNN) and Elastic Net (ENet). Model performance was evaluated using the area under the receiver operating characteristic curve (ROC-AUC) and the SHAP method was used to rank importance of variables collected. Overall, the Enet model demonstrated the highest ROC-AUC (0.935) and PR-AUC (0.754) in both the internal and external validation cohorts. SHAP identified number of regional lymph node metastases, maximum tumor diameter, monocyte-to-lymphocyte ratio, neutrophil-to-lymphocyte ratio, depth of invasion, and lymphovascular invasion as key predictors of increased risk of distant metastasis. Additionally, SHAP identified lymphovascular invasion and poorly differentiated histological grade as variables that had a positive contribution to predicting distant metastasis risk, while well-differentiated tumors had a negative contribution.

Outcomes and Implications

This study showed that MLMs are effective at predicting distant metastasis risk in TSCC patients. Overall, TSCC is a very distinct subtype of head and neck squamous cell carcinoma, and risk assessment strategies designed for HSCC generally may not be precise for stratifying the risk associated with distant metastasis in TSCC. This paper addresses a very important clinical need, and can still be utilized quickly after surgery to guide treatment for these high-risk patients. Additionally, this model enables immediate risk prediction after receiving postoperative pathology details, helping physicians take a proactive approach to risk stratification instead of using conventional imaging methods like CT to identify already established metastases. Ultimately, these models demonstrate their ability to account for the multitude of factors that affect distant metastases and show true clinical utility for a high-risk patient population.