Related Experiment Video
Updated: Jun 13, 2026
Noninvasive Monitoring of Lesion Size in a Heterologous Mouse Model of Endometriosis
Published on: February 26, 2019
Machine learning models for non-invasive endometriosis triage using a laparoscopically and histologically verified
Marina Paula Andres1, Clarisse da Costa Rufino2, Ramanathan Anantharama Krishnan2
1HerHealth.AI, Orlando, FL, USA; Disciplina de Ginecologia, Departamento de Obstetrícia e Ginecologia, Faculdade de Medicina FMUSP, Universidade de Sao Paulo, Sao Paulo, BR, Brazil; Gynecologic Division, Beneficência Portuguesa de São Paulo, Sao Paulo, SP, Brazil.
Objective:
To develop and internally validate machine-learning models for non-invasive triage of women at risk for endometriosis using structured clinical variables in a laparoscopically and histologically verified cohort.
Study Design:
This retrospective study included 2546 women who underwent laparoscopic surgery between 2008 and 2023 at two tertiary referral centers in São Paulo, Brazil. Endometriosis was confirmed in 1983 patients and absent in 563 controls, corresponding to an enriched tertiary-care case prevalence of 77.9%. Two feature-selection strategies were compared: clinician-guided selection and statistically optimized selection. Preprocessing, feature selection, MinMax scaling, and Synthetic Minority Oversampling Technique (SMOTE) were performed within the training workflow of stratified 10-fold cross-validation to minimize information leakage. Primary performance measures were F1-score, recall, positive predictive value (PPV), negative predictive value (NPV), and AUC-ROC; accuracy was reported only as a secondary metric.
Results:
Statistically optimized feature selection produced modest but consistent improvements in discrimination and F1-score across most models. XGBoost achieved the highest F1-score in the statistically optimized analysis (0.916, 95% CI 0.909-0.922), with recall of 0.932 (95% CI 0.921-0.943), PPV of 0.900 (95% CI 0.894-0.906), NPV of 0.730 (95% CI 0.700-0.761), and AUC-ROC of 0.895 (95% CI 0.887-0.904). The ensemble model achieved the highest PPV (0.924, 95% CI 0.899-0.947). The most informative variables included infertility, dysmenorrhea, pain level, cyclic intestinal pain, abnormal vaginal examination, number of diseases reported, and menstrual-flow characteristics.
Conclusion:
Machine-learning models based on structured clinical variables may support non-invasive triage of women at risk for endometriosis. The contribution of the revised framework is the use of a clinically verified cohort, transparent feature selection, and leakage-aware internal validation rather than removal of diagnostically challenging records. Because this retrospective study was internally validated in tertiary referral centers with enriched disease prevalence, external validation, local calibration, and prospective clinical utility assessment are required before implementation.
