P-333 Machine learning for endometriosis prediction: analyzing self-reported data from the Lucy health app

Abstract Study question Can machine learning models trained on self-reported data from the Lucy mobile health app predict endometriosis? Summary answer Machine learning models trained on self-reported data from Lucy app accurately predicted endometriosis, with XGBoost and Random Forest achieving accuracies of 89% and 84%, respectively. What is known already Endometriosis affects millions worldwide, however it is poorly recognized, underfunded, and under-researched, leading to diagnostic delays of 4 to 11 years. Mobile health tools can potentially transform chronic disease management by improving symptom tracking and assessment—offering clinical benefits and cost savings. Despite these benefits, their application in endometriosis remains limited, and research is scarce. This study aims to bridge this gap by applying machine learning (ML) classification models to analyze real-world, self-reported data from the Lucy app for early detection of endometriosis. Study design, size, duration This prospective study utilized self-reported data from the Lucy app, which allows users to log their menstrual cycles, medical records, dietary details, and pain symptoms. Women aged 18 years and older, both with and without a diagnosis of endometriosis, were included in the study. Data were collected from 4,812 users (n = 1,212 controls, n = 3,600 endometriosis cases) from eight European countries, including Hungary, Denmark, Sweden, Germany, Austria, Italy, Romania, and Poland. Participants/materials, setting, methods After filtering, 520,000 records were analyzed using correlation methods and visualization tools to identify symptom associations with endometriosis. Machine learning models (XGBoost, Random Forest) were trained on a subset of 4,812 users to predict endometriosis. To handle missing data, we applied imputation techniques, replacing missing values with statistically estimated ones to maintain data integrity and enhance model performance. Main results and the role of chance The first step was to filter the data to address issues such as unrealistic data and inconsistencies typical of real-world datasets. Strong associations between known endometriosis-associated symptoms—including pelvic pain, pelvic cramps, dysmenorrhea, and lower back pain—were shown by records from patients diagnosed with endometriosis but not yet treated. Our machine learning models effectively distinguished between endometriosis and control cases. The XGBoost model achieved an overall accuracy of 89%, with an F1-score of 0.92. The Random Forest model achieved an accuracy of 84%, with an F1-score of 0.78. These metrics were obtained as means from 5-fold cross-validation, supporting the robustness of the model. Feature importance analysis identified dysmenorrhoea and pelvic pain as the most important predictors of classification. Limitations, reasons for caution While our models demonstrated strong predictive performance, limitations include reliance on self-reported data and lack of clinical validation. When working with patient symptom data, missing values are a common challenge and may affect accuracy. Future studies should focus on external validation and clinical implementation to improve diagnostic reliability. Wider implications of the findings Our preliminary analysis showed that real-world, self-reported data are consistent with known endometriosis symptoms, highlighting that using mobile apps like Lucy for endometriosis monitoring is a promising strategy. These findings may pave the way for improving early detection and transforming endometriosis management, ultimately improving patient care. Trial registration number Yes