Machine Learning‐Assisted Prediction of Varicocele Grade Using Multidimensional Spermatic Vein Reflux Time Analysis
Ningning Liu, Bo Zhang, Xiaoyi Wang, Youhe Zuo, Jing LiABSTRACT
Background
Varicocele (VC) grading has long relied on qualitative assessments of venous diameter and reflux signals, lacking standardization.
Objectives
This study aimed to explore the value of spermatic vein reflux time (SVRT) as a quantitative indicator for VC grading and to enhance diagnostic accuracy using machine learning (ML) models.
Materials and Methods
A total of 3855 VC patients were enrolled, with color Doppler ultrasound used to measure SVRT, venous diameter, and testicular hemodynamic parameters (peak systolic velocity [PSV], end‐diastolic velocity [EDV], and resistance index [RI]). VC grades were defined using an institutional clinical grading protocol based on a dual‐index grading standard (venous diameter + SVRT). Statistical analyses included one‐way ANOVA, Spearman correlation, and receiver operating characteristic (ROC) curve analysis. Three ML models (Random Forest, XGBoost, and Logistic Regression) were developed using clinical and ultrasound features; two XGBoost versions (with/without SVRT) were designed to rule out circular reasoning. SHAP analysis was applied to interpret feature contributions.
Results
SVRT increased significantly with VC grade (left side: Grade I, 2.85 ± 1.12 s; Grade II, 5.28 ± 0.92 s; Grade III, 7.61 ± 1.32 s; all p < 0.001) and showed significant correlations with venous diameter ( r = 0.389–0.426) and reflux velocity ( r = 0.478–0.512, all p < 0.001). ROC analysis demonstrated SVRT's diagnostic efficacy (AUC = 0.826–0.935) and derived grade‐specific cutoffs (Grade I: 2.30 s; Grade II: 4.50 s; Grade III: 6.20 s). XGBoost (with SVRT) achieved the highest performance: overall accuracy 89.2% (95% CI: 0.881–0.903), Macro‐F1 0.882, and AUC = 0.941 (95% CI: 0.925–0.957) for Grade III, outperforming the traditional SVRT‐based ROC (AUC = 0.935, 95% CI: 0.922–0.948; DeLong test: Z = 2.13, p = 0.03) and the SVRT‐excluded model (accuracy 80.4%, 95% CI: 0.783–0.825; Grade III AUC = 0.872, 95% CI: 0.844–0.899). SHAP analysis identified left SVRT (mean absolute SHAP value = 0.23), left venous diameter (0.19), and left EDV (0.15) as key predictive features. For SVRT regression, the Random Forest model showed high accuracy ( R 2 = 0.820, 95% CI: 0.802–0.838; RMSE = 0.76 ± 0.12 s; MAE = 0.58 ± 0.09 s), with minimal error for severe VC cases (SVRT > 6.2 s, MAE = 0.41 ± 0.07 s).
Discussion and Conclusion
These findings support SVRT as a quantitative reference for VC severity grading, and the developed ML models may assist in standardizing VC grading by integrating multidimensional features, with the potential to reduce inter‐observer variability and borderline case misclassification. However, the absence of clinical outcome data means these tools have been validated only against grading criteria, not against patient‐relevant endpoints; prospective studies with outcome data are needed before clinical adoption.