Poster #51: Lannawill Caruth

Multimodal models integrating rare variants and unstructured EHR features identify endometriosis with AUROC ~0.80 in a chart-reviewed biobank cohort

Lannawill Caruth (Department of Pathology and Laboratory Medicine, University of Pennsylvania), Lindsay Guare (Genomics and Computational Biology, University of Pennsylvania), Jagyashila Das (Department of Pathology and Laboratory Medicine, University of Pennsylvania), Anaya Kapoor ( Biomedical Informatics, University of Pennsylvania), Ananya Rajagopalan (Genomics and Computational Biology, University of Pennsylvania), Penn Medicine BioBank, Regeneron Genetics Center, Shefali Setia-Verma (Department of Pathology and Laboratory Medicine, University of Pennsylvania)

Endometriosis affects ~10% of reproductive-age women, yet diagnosis is frequently delayed by years, prolonging pain, impairing fertility, and increasing healthcare burden. Although large-scale genomic studies (GWAS and ExWAS) have advanced understanding of its genetic architecture and enabled polygenic risk scores (PRS), these approaches capture only a limited fraction of disease risk. Current PRS frameworks are largely restricted to common variants and linear modeling, limiting their ability to account for rare genetic variation and nonlinear interactions that are central to this heterogeneous disease. Critically, clinically relevant signals including symptoms such as pelvic pain and dysmenorrhea are often embedded in unstructured electronic health record (EHR) data and are not incorporated into existing risk models. To address these limitations, we present,to our knowledge, the first multimodal machine learning framework that integrates common and rare genetic variation with clinically derived features to improve risk stratification for endometriosis, adenomyosis, both conditions, or neither. Using the richly phenotyped Penn Medicine BioBank, we analyze a chart-reviewed cohort of 1,400 patients with endometriosis status derived from unstructured data and validated through detailed clinical note review, enabling inclusion of high-resolution clinical variables typically inaccessible to structured analyses. To determine which features and modeling approaches are most informative, we systematically evaluate the combinations of common variant scores, rare variant scores, and clinical variables, alongside baseline covariates including age and principal components, across four model architectures: random forest, XGBoost, a feedforward neural network, and a Gaussian Naïve Bayes probabilistic graphical model. Outcomes are defined as endometriosis only, adenomyosis only, both conditions, or neither condition. Across models, AUROC performance varied by outcome: the strongest performance was observed for predicting endo and the co-occurrence of endo and adeno, and weaker performance was observed for distinguishing adeno alone from individuals with neither condition. Models integrating clinical features with genetic data consistently outperformed genetics-only approaches, highlighting the added predictive value of clinically derived signals. In particular, feature sets combining clinical variables with rare variants, with or without common variant scores, achieved the strongest performance, with AUROCs of ~0.75–0.80 across multiple outcomes. Model performance was comparable across architectures, with the feedforward neural network and Gaussian Naïve Bayes performing on par with, and in some cases slightly exceeding, Random Forest and XGBoost. In contrast, models relying solely on rare variants or baseline covariates showed poor discrimination, with AUROCs frequently near or below ~0.60. Together, these results demonstrate that integrating genomic and clinically derived features enables robust, non-invasive risk stratification for endometriosis and adenomyosis. In the absence of clinically deployed predictive models, this approach demonstrates feasibility for earlier identification, with the potential to reduce diagnostic delay and improve patient outcomes.