Patient-Level Risk Characterization of Drug-Associated Hidradenitis Suppurativa Using Machine Learning.

Importance: Hidradenitis suppurativa (HS) has been reported with several medications, but the spectrum of implicated drug classes and patient-level risk factors remains poorly characterized. Objective: To characterize drug-associated HS pharmacovigilance signals across drug classes, distinguish new-onset from worsened disease, and evaluate whether machine learning (ML) approaches identify reproducible patient-level predictors. Design: Disproportionality and ML of spontaneous adverse event reports submitted from 2004 through 2023. Setting: The US Food and Drug Administration Adverse Event Reporting System. Participants: Reports containing a preferred term for hidradenitis, stratified by whether HS was a listed indication (worsened) or an adverse event only (new-onset). A parallel cohort comprised all tumor necrosis factor inhibitor (TNFi)-exposed reports. Exposures: Primary suspect and concomitant medications. Main Outcomes and Measures: Disproportionate reporting quantified by the reporting odds ratio; agent- and indication-specific hidradenitis suppurativa reporting rates; and discrimination, calibration, feature concordance, and risk enrichment of supervised machine-learning models estimating the probability of HS. Results: Among 5529 HS reports (3725 with and 1804 without an indication), patients were predominantly female (3511 [63.5%]), with a mean (SD) age of 41 (14) years. Among reports without a hidradenitis suppurativa indication, significant signals included adalimumab (reporting odds ratio, 12.6; 95% CI, 11.3-14.0), infliximab (8.2; 6.7-9.9), secukinumab (6.6; 5.2-8.2), and isotretinoin (6.2; 4.2-8.9). Among TNFi exposed reports, reporting rates were highest for adalimumab (1.08 per 1000; OR, 2.75; 95% CI, 2.41-3.14) and lowest for etanercept (0.11 per 1000; OR, 0.12; 95% CI, 0.09-0.15), and the association was strongest in inflammatory bowel disease (OR 3.06; 2.20-4.25); all reported associations were significant at P < .001. Supervised ML models demonstrated similar discrimination while identifying consistent patient-level predictors of HS. Tree-based models showed greater enrichment of high-risk reports than penalized regression. Conclusions and Relevance: HS reporting signals spanned biologic, hormonal, retinoid, and immunomodulatory drug classes. Multiple supervised ML approaches consistently identified TNFi agent (adalimumab), age, and a history of arthropathy as associated with paradoxical HS, suggesting that pharmacovigilance data contain reproducible information on heterogeneity of HS reporting. These findings support indication-aware monitoring when initiating these therapies, while recognizing that spontaneous reports reflect disproportionate reporting rather than incidence or causality.