Phenotyping Adolescent Endometriosis: Characterizing Symptom Heterogeneity Through Note- and Patient-Level Clustering

IntroductionPelvic pain (dysmenorrhea and non-menstrual) is the most common presentation of adolescent endometriosis, but symptoms vary between and within patients. Other presentations, such as gastrointestinal (GI) symptoms, are often misattributed, leading to diagnostic delays. Patients incur frequent primary and specialty care visits, generating multiple and diverse clinical notes. These offer insights into disease trajectory and symptom heterogeneity, which can be rigorously investigated using clustering methods. This study aims to 1) evaluate phenotypes using electronic health records (EHRs) and 2) compare two clustering models (note-vs patient-level) for their ability to identify symptom patterns. MethodsWe queried the Mount Sinai Data Warehouse for clinical notes from patients aged 13-19 years with a SNOMED endometriosis diagnosis, yielding an initial sample of 7,221 notes. A randomly selected subsample was annotated with 12 disease-relevant labels, including symptoms, hormone use, and medications. The final analytic sample included 695 notes from 26 unique patients. Pelvic pain, dysmenorrhea, chronic pain, and GI symptoms were selected as model predictors based on principal component analysis. Two unsupervised machine learning (ML) methods were then applied for note-vs patient-level analyses: Partitioning Around Medoid (PAM) and Multivariate Mixture Models (MGM). ResultsThe PAM model identified K=3 clusters with average silhouette width of 0.76, indicating strong between-cluster separation. The "feature-absent" (abs) phenotype (76%) was distinct for absence of all 4 features. The "classic" phenotype (8%) exhibited pelvic pain, dysmenorrhea, and chronic pain. The "GI" phenotype (16%) was dominated by GI symptoms. The MGM identified K=2 stable patient-level clusters ({Delta} weighted model deviance = -224.93 from K=2 to 3) with a mean cluster membership probability of 0.97: A "classic" phenotype (50%), characterized by pelvic pain and chronic pain, and a "non-classic" phenotype (50%), defined by the absence of these features. PAM-based classic phenotype had significantly higher rates of hormonal intervention (78% vs 26% abs, 49% GI) and pain medication (68% vs 9% abs, 14% GI). For the patient-level, the classic phenotype also had higher average rates per person of hormonal therapy (26% vs 7%) and prescription pain medications (27% % vs 9%) (p<0.01 for all). ConclusionsBoth methods captured classic and non-classic phenotypes, with the note-level model uniquely identifying a feature-absent group. The classic phenotypes link to higher hormonal and pain intervention underscores the importance of recognizing non-classic symptoms. This study, the first to directly compare note-and patient-level clustering of EHR notes in endometriosis, demonstrates the ability to detect the less clinically recognizable phenotypes. This proof-of-concept can be applied to larger datasets to refine phenotype identification, aiding in earlier diagnosis.