Risk of Postoperative Recurrence in Patients With Unusual Smooth Muscle Neoplasia and Construction of a Prediction Model
Weiwei Hu, Shiying Wu, Jingxi Zhang, Yueran Li, Xiangyang Zeng, Songshu XiaoABSTRACT
Objective
To identify the risk factors associated with postoperative recurrence of unusual (special‐type) uterine smooth muscle neoplasia (USMN) and to construct and interpret a machine‐learning model for recurrence risk.
Methods
We carried out a retrospective review of patients who had myomectomy for uterine masses at the Third Xiangya Hospital between January 2015 and December 2024, zeroing in on those with paraffin‐confirmed special variants—specifically cellular, atypical/bizarre, and vascular leiomyomas. After a negative transvaginal ultrasound at the 3‐month mark, we tracked recurrence over the next 12 months, assigning patients to either the recurrence or non‐recurrence groups. The full cohort was randomly split 7:3 into a training set and a hold‐out validation set. For missing data within the training portion, we turned to random forest imputation. To pin down the most relevant predictors, we first ran LASSO regression for screening, then solidified our choices using multivariable logistic regression. On the modeling front, we put eight distinct algorithms head‐to‐head: logistic regression, support vector machine, random forest, gradient boosting machine, XGBoost, k‐nearest neighbors, AdaBoost, and a neural network. All were compared across the validation set, with a close eye on discrimination, calibration, and net benefit. The winner among them was later unpacked via SHAP to see exactly how each feature drove predictions. As for the surgical approach, we examined its link to recurrence using two separate strategies—inverse probability of treatment weighting and pairwise propensity‐score matching—to ensure the association wasn't driven by confounding.
Results
After screening 285 patients, we analyzed 248, among whom 87 (35.1%) had a recurrence within the first year. Multivariable logistic regression singled out six independent predictors: intramural location, T 2 WI hyperintensity, tumor multiplicity (≥ 2), size ≥ 6 cm, raised BMI, and coexisting adenomyosis (all p < 0.05). No significant difference in recurrence rates was observed across subtypes (cellular 35.6%, atypical 38.5%, vascular 31.0%; p = 0.793). Neither IPTW nor propensity‐score matching revealed an independent association between surgical approach and recurrence. Of the eight models tested, GBM yielded the best performance on validation (AUC 0.785, sensitivity 0.750, F1 0.679). SHAP analysis further pinpointed those same six variables as top contributors and corroborated their effect directions.
Conclusion
Taken together, special‐type USMN entails a notable risk of early recurrence. Our GBM model, which relies on six easily obtainable preoperative factors, demonstrated reasonable discriminative ability; coupled with SHAP explanation, it could aid in personalized patient counseling and surveillance planning.