Using Natural Language Processing to Support Early Detection of Endometriosis

Endometriosis is a chronic inflammatory condition that affects approximately 190 million women who are assigned female at birth. Endometriosis lesions cause severe, debilitating pain, yet remains one of the most misdiagnosed conditions in women’s health. As a result, clinically relevant information exists primarily in patient narratives. By using natural language processing (NLP) techniques to parse patient narratives, supervised machine learning models can be used to differentiate control cases and cases of diagnosed endometriosis. This demonstrates the potential of using computational methods to combat diagnostic delays in common, underdiagnosed diseases.

Figure 1. Endometriosis misdiagnosis rate (Bontempo and Mikesell 2020). Dataset & Preprocessing: a publicly available dataset was used (Jackson 2025). It contains 10,000 synthetic but realistic data with clinical indicators of endometriosis. Data was cleaned by handling missing values and standardizing categorical variables. Patient narratives were normalized through tokenization, lowercasing, and removal of stopwords. Impact & Applications This work demonstrates the potential of NLP-driven models to support an early identification of endometriosis using patient- reported narratives. An earlier detection could reduce diagnostic delays, improve patient outcomes, and enable more timely intervention in a condition that is underdiagnosed. The proposed approach can serve as a non-invasive clinical decision support tool, which can assist healthcare providers in identifying high-risk patients based on symptom language patterns. This has value in cases where traditional diagnostic pathways rely heavily on subjective symptom reporting. Because this model operates on text data, it has potential to be integrated into digital health platforms, online symptom checkers, or

Bontempo A, Mikesell L. 2020. Patient Perceptions of Misdiagnosis of Endometriosis: Results from an Online National Survey. Diagnosis 7(2): 97-106. https://doi.org/10.1515/dx-2019-0020. Google. 2025. Machine Learning: Text Classification. https://developers.google.com/machine-learning/guides/text-classification/step-3 Horne A. W., Missmer S. A. 2022. Pathophysiology, Diagnosis, and Management ofEndometriosis. BMJ: British Medical Journal, 379, 1–19. https://www.jstor.org/stable/27424071. Jackson A. 2025. Endometriosis Dataset. Kaggle. https://www.kaggle.com/datasets/michaelanietie/endometriosis-dataset/data. Ranney, B. 1952. Endometriosis. The American Journal of Nursing, 52(12), 1465–1467.https://doi.org/10.2307/3459462. Teh D. 2022. Not Just Period Pain. HealthMatch. https://healthmatch.io/blog/not-just-period-pain-the-endometriosis-crisis-that-affects-1-in-10- women. Figure 3. Total time taken from symptoms to medical visits (Teh, 2022). Natural Language Features & Model Development: features were extracted from symptom descriptions using tokenization and vectorization, and frequency-based keyword extraction. The supervised machine learning classifier was trained to predict the presence of endometriosis. Logistic regression and random forest models were evaluated. Accuracy, precision, recall, and F1-score was used to measure the effectiveness of NLP-based detection. Figure 2. An example of preparing data for machine learning through tokenization and vectorization (Google 2025). electronic health records to provide scalable, low-cost preliminary screening for endometriosis risk. This could vastly reduce the time taken between onset symptoms to consultation to diagnosis. This model also provides a possible framework to support other underdiagnosed diseases.