Develop and evaluate a machine learning–based prognostic model to support individualized management decisions in ovarian torsion.
Approach:
Study Design: Retrospective cohort study of 219 patients diagnosed with ovarian torsion.
Data Analysis: Analyzed demographic characteristics, reproductive and surgical history, comorbidities, presenting symptoms, laboratory findings, and imaging features.
Machine Learning Models: Trained and compared multiple supervised machine learning models including Logistic Regression, Decision Tree, Random Forest, Gradient Boosting, and Neural Network classifiers.
Class Imbalance Handling: Addressed class imbalance using class weighting and oversampling techniques (ROSE, SMOTE, SMOTE-ENN).
Model Evaluation: Evaluated model performance using area under the receiver operating characteristic curve (AUC), sensitivity, and specificity.
Dimensionality Reduction: Performed principal component analysis (PCA) and unsupervised clustering to identify latent clinical domains and phenotypic subgroups.
Key Findings:
The class-weighted Decision Tree model demonstrated the best performance (AUC = 0.76; sensitivity = 0.75; specificity = 0.73).
Key predictors included Doppler findings, BMI, blood group (ABO), ethnicity, clinical symptoms, pelvic mass presence, menopause, history of OCP, infertility, and previous surgery.
PCA identified eight clinical domains, and clustering revealed two distinct patient profiles.
Interpretation:
Machine learning models can assist in individualized risk stratification for ovarian torsion management.
Limitations:
Retrospective design may introduce bias.
Need for prospective multicenter validation before clinical implementation.
Conclusion:
A transparent, non-overfitted Decision Tree model offers a clinically interpretable framework for management decisions in ovarian torsion.
