Explainable machine learning models for endometriosis diagnosis: a retrospective development and validation study

Background This study aimed to develop and validate machine learning (ML) models for diagnosing endometriosis based on common clinical characteristics, aiming to offer guidance for clinical screening and early diagnosis of endometriosis. Methods We retrospectively collected data on female patients aged 18 years and older from a tertiary hospital in China. The data were randomly split into a training set and a validation set in a 7:3 ratio. Patients were divided into an endometriosis group and a non-endometriosis group. In the training set, nivariable and multivariable logistic regression analyses were performed to identify independent risk factors for endometriosis. (logistic regression [LR, a traditional statistical inference model with a linear baseline], random forest [RF, bagging ensemble of decision trees], support vector machine [SVM, radial-basis-function kernel], and XGBoost [gradient-boosted trees]) were constructed. These four algorithms were selected because they cover complementary learning paradigms, encompassing interpretable linear models and nonlinear ensemble methods.The optimal Youden index was used to select the cutoff probability, thereby determining the sensitivity and specificity of each model. Results A total of 233 patients were included. Among them, 120 patients had endometriosis. Our analysis showed that a history of uterine fibroids, menopausal status, dysmenorrhea, white blood cell count, CA125, and CA199 were independent influencing factors for endometriosis. Results revealed that in the validation set, the AUC, sensitivity, and specificity were 0.878 (95% confidence interval [CI]: 0.793–0.962), 0.944, and 0.706 for LR; 0.878 (95% CI: 0.795–0.960), 0.694, and 0.912 for RF; 0.853 (95% CI: 0.764–0.941), 0.667, and 0.853 for SVM; 0.860 (95% CI: 0.777–0.943), 0.722, and 0.794 for XGBoost. The detection performance of the LR model was not significantly inferior to that of the other models. Conclusion Building ML models to detect endometriosis based on basic demographic information, clinical characteristics, and laboratory indicators is a feasible approach with desirable performance. The sample size in this study was limited, and internal validation was performed using a random split of the dataset. Therefore, caution is still warranted regarding the validation performance and interpretation of this model.