Construction and Validation of a Machine Learning Model Integrating Ultrasound Features and Inflammatory Markers (OVART-ML) for Predicting Ovarian Torsion and Ischemic Necrosis Risk in Children

Zhifei Zhao¹Yubing Wang¹Binyi Yang¹Jiaxiang Tang¹Jinbin Wang¹Shujie Song¹Yuezhen Zhang²Hongting Lu^1*

• ¹Qingdao University, Qingdao, China
• ²Linyi People's Hospital, Linyi, China

Objective: To construct a machine learning (ML) model (OVART-ML) using multimodal clinical data for predicting the risk of ovarian torsion (OT) and secondary ischemic necrosis (IN) in children and to identify key factors to assist clinical decision-making. Methods: A retrospective analysis was conducted on data (demographic characteristics, symptoms, ultrasonic findings, and laboratory indicators) of 112 children with ovarian space-occupying lesions admitted to Qingdao Women and Children's Hospital and Linyi People's Hospital between January 2012 and December 2024. After preprocessing (data standardization and LASSO feature selection), 11 ML algorithms (including Support Vector Machine [SVM], K-Nearest Neighbors [KNN], and Random Forest [RF]) were used to construct predictive models. Model performance was evaluated using indicators such as the Area Under the Curve (AUC), accuracy, and specificity. Key risk factors were identified using SHapley Additive exPlanations (SHAP). Results: Among 112 children, 60 (53.6%) developed OT and 23 (20.5%) developed IN. The SVM model exhibited the optimal performance: in the test set, its AUC was 0.911 (95% Confidence Interval [95% CI]: 0.809-1.000), accuracy was 0.882, sensitivity was 0.900, and specificity was 0.857. SHAP analysis identified 8 key factors: the follicular edema ring sign, vomiting, pelvic effusion, eosinophil (EOS) count, white blood cell (WBC) count, hemoglobin (Hb) level, Neutrophil-to-Eosinophil Ratio (NER), and Systemic Immune-Inflammatory Index (SII). Among these, the follicular edema ring sign (mean |SHAP value| = 0.12) and EOS count (mean |SHAP value| = 0.08) had the highest predictive weights. Conclusion: This study developed an interpretable ML model that could accurately predict the risks of OT and IN in children. Key factors such as the follicular edema ring sign and vomiting provide important references for early diagnosis and intervention. This tool may assist clinicians in making timely surgical decisions to preserve ovarian function in children.