Z41 - Machine Learning Assessment of Patient Symptoms as Indicators of Endometriosis Presence and Severity

Friday, October 10, 2025

10:00 AM - 11:00 AM PDT

Location: Exhibit Hall F, G & H

Presenting Author(s)

Yvette Wang

Undergraduate Student
University of Pittsburgh, United States

Primary Investigator(s)

Yvette Wang

Undergraduate Student
University of Pittsburgh, United States

Co-Author(s)

IC

Isabelle Chickanosky

Graduate Student
University of Pittsburgh, United States
ND

Nicole Donnellan, MD

Associate Professor, Department of Obstetrics, Gynecology and Reproductive Sciences Director
University of Pittsburgh, United States
Timothy K. Chung, PhD

Research Assistant Professor
University of Pittsburgh
Pittsburgh, Pennsylvania, United States

Last Author(s)

David A. Vorp, PhD

John A. Swanson Professor of Bioengineering and the Senior Associate Dean of Research and Facilities
University of Pittsburgh, United States

Introduction: : Endometriosis (EM) is a gynecological condition characterized by the growth of endometrium-like tissue outside of the uterus. Affecting at least 11% of American women of reproductive age, endometriosis can cause chronic, debilitating pain and leads to infertility. Despite endometriosis’ prevalence, the diagnostic process is invasive and lengthy– taking up to 10 years on average after symptom onset, with its only confirmation through a laparoscopic surgery. Out of the patients who undergo this surgery, only ~40% have pathologically diagnosed endometriosis. This project seeks to improve EM diagnosis among symptomatic EM and symptomatic non-EM patients by combining three types of patient data: (1) symptom and quality of life surveys, (2) clinical, and (3) biological. The utilization of all three types of data may contribute to a holistic EM diagnosis tool through statistical analyses and machine learning (ML) to identify less noticeable connections within the data.

Materials and

Methods: : Patient data and biospecimens were collected under an IRB protocol Pitt STUDY23020127 at UPMC Magee-Womens Hospital. Symptom survey data was collected pre-operatively using demographic questions, the Endometriosis Health Profile-30 (EHP), and the Visual Analog Scale (VAS). Clinical data was chart reviewed by Dr. Donnellan. Patient serum was isolated from peripheral blood samples to quantify total protein using bicinchoninic acid assays (BCA). Two patient datasets were evaluated: (1) 276 patients with survey and clinical data and (2) 74 patients with survey, clinical, and available biological data (total protein levels). The platforms GraphPad Prism (Boston, MA) and RStudio (Boston, MA) were used to quantify and statistically analyze assay readings before using MATLAB (Boston, MA) Classification Learner. ML models for the prediction of EM presence and severity (low stage versus high stage) were developed using different combinations of survey, clinical, and biomarker data, resulting in eight models total. For the second dataset, no severity model was developed due to the small sample size with EM (n=46). Feature importance for each model was analyzed using the analysis of variance function.

Results, Conclusions, and Discussions:: From the dataset of 276 patients, four ML models were developed using survey and/or clinical data. For survey inputs alone, EM presence detection had an AUC of 0.628 and an accuracy of 67.0%, while severity prediction had an AUC of 0.730 and an accuracy of 70.2%. When using survey and clinical inputs, EM presence detection had an AUC of 0.652 and an accuracy of 73.2%, while severity prediction had an AUC of 0.731 and an accuracy of 69.3%. Feature importance was analyzed (Fig. 1). Using data from 74 patients, the inclusion of total protein analysis showed minimal improvement to presence prediction. Four ML models were developed using a combination of survey, clinical, and biological data. Likely due to the small sample size, model performance was lower, with accuracies ranging from 59.5-62.2%.

For the dataset of 276 patients, prediction of disease presence improved when both clinical and survey data were used, despite clinical data not ranking as one of the top five features. This demonstrates that the usefulness of clinical data may be understated by examining feature importance alone. However, for stage prediction, feature importance demonstrated that patient survey data was most critical. This could be driven by the close tie between symptoms and quality of life decline that EM could cause. The models from the dataset of 74 patients showed minimal predictive value due to the small sample size of this dataset and non-specific analysis of protein.

This study relates patient symptoms, clinical data, biomarkers, and EM diagnosis. A total of eight models were created from two patient datasets using different combinations of survey, clinical, and biological data to predict disease presence and severity. For presence detection, inputs of both clinical and survey data resulted in the best accuracy. Severity prediction was most reliant on survey data and showed negligible improvement in prediction with the inclusion of clinical data. Future work is needed to examine the effects of adding biological data to these models. We will work to increase sample size and to measure specific protein levels of inflammatory biomarkers known to be elevated in the serum of EM patients.

Acknowledgements and/or References (Optional): : 1 Horne, A. W. & Missmer, S. A. Pathophysiology, diagnosis, and management of endometriosis. Bmj 379, e070750 (2022). https://doi.org/10.1136/bmj-2022-070750
2 Anchan, M. M. et al. Unveiling the fibrotic puzzle of endometriosis: An overlooked concern calling for prompt action. F1000Res 13, 721 (2024). https://doi.org/10.12688/f1000research.152368.3
3 Agarwal, S. K. et al. Clinical diagnosis of endometriosis: a call to action. Am J Obstet Gynecol 220, 354.e351-354.e312 (2019). https://doi.org/10.1016/j.ajog.2018.12.039
4 Buck Louis, G. M. et al. Incidence of endometriosis by study population and diagnostic method: the ENDO study. Fertil Steril 96, 360-365 (2011). https://doi.org/10.1016/j.fertnstert.2011.05.087