[Interpretable machine learning models for preoperative precision prediction of perineural invasion in cervical
Mengxin Zhu1,2, Xiao Liu3, Shan Zhao4
1School of Public Health, North China University of Science and Technology, Tangshan 063210, China.
Objectives:
To develop an interpretable machine learning model and web-based prediction tool for preoperative risk assessment of perineural invasion (PNI) in cervical cancer.
Methods:
A total of 845 cervical cancer patients undergoing radical surgery at Fourth Hospital of Hebei Medical University were retrospectively enrolled and divided into training and testing sets in a 7:3 ratio, with another 223 cervical cancer patients at Hebei Medical University Second Hospital during the same period serving as the external validation cohort. LASSO regression identified 13 preoperative predictors, which were incorporated into 7 machine learning algorithms. Model performance was evaluated using AUC and decision curve analysis. The optimal model was interpreted using SHAP values and deployed as a web-based prediction tool.
Results:
Of the total of 1068 patients enrolled, 192 (17.98%) were diagnosed to have PNI. Thirteen preoperative features, namely lymphovascular space invasion (LVSI), depth of stromal invasion, lymph node metastasis (LNM), colposcopy-directed biopsy (CDB), tumor maximum diameter, carcinoembryonic antigen, SCC-Ag, platelet-to-lymphocyte ratio (PLR), neutrophil-to-lymphocyte ratio (NLR), lymphocyte-albumin-neutrophil ratio (LANR), menopausal status, age, and histological type were selected. Comparison of model performance revealed that the Extreme Gradient Boosting (XGBoost) model resulted in the best efficacy in both the training and testing datasets with AUC of 0.962 and 0.923 (95% CI: 0.942-0.979 and 0.874-0.960), sensitivity of 0.873 and 0.767, and specificity of 0.939 and 0.942, respectively. Decision curve analysis demonstrated greater net benefit of the XGBoost model across a broader threshold range. The SHAP-XGBoost model showed excellent performance in external validation with an AUC of 0.924 and an accuracy of 0.933.
Conclusions:
The interpretable SHAP-XGBoost model effectively predicts PNI risk preoperatively. The predictive website derived from this model provides an useful tool to facilitate clinical decision-making in cervical cancer treatment.
