CO:05:4 | Natural language processing reveals demographic features, comorbidities, and clinical domains to predict treatment longevity of b/tsDMARDs in psoriatic arthritis
Antonio Tonutti1|2, Pierandrea Morandini3, Cosimo Faeti3, Nicoletta Luciano1|2, Saverio D'Amico3, Victor Savevski3, Carlo Selmi1|2 | 1Department of Biomedical Sciences, Humanitas University Pieve Emanuele, Milan, Italy; 2Rheumatology and Clinical Immunology, IRCCS Humanitas Research Hospital Rozzano, Milan, Italy; 3Artificial Intelligence Unit, IRCCS Humanitas Research Hospital Rozzano, Milan, Italy
Abstract
Background. Evidence on treatment trajectories and predictors of retention is limited in psoriatic arthritis (PsA). This study aims to identify factors associated with the 12-month retention of newly prescribed b/tsDMARDs in PsA using natural language processing (NLP).
Methods. For PsA patients attending our center between Jan2019 and Feb2024, the first visit in which a new b/tsDMARD was prescribed was retrieved, regardless of the treatment line. A BERT-based NLP model was developed to extract demographic, clinical, and treatment data. Treatment data were matched with dispensing records from the Hospital Pharmacy to verify which drugs were prescribed, discontinued, or switched during a 12-month follow-up. Model outputs were used to build a logistic regression model to predict drug retention. The model was trained/tested (k-fold cross-validation); performance metrics were computed, and odds ratios for predictors of retention and discontinuation were estimated.
Results. We included 413 PsA cases (52% female; age 52±12) initiating a new b/tsDMARD: 90% had peripheral arthritis, 15% axial disease, 18% dactylitis, 18% enthesitis, 61% skin psoriasis, and 2% inflammatory bowel disease. Comorbidities included depression-anxiety (7%), diabetes (10%), obesity (9%), and smoking (14%). New b/tsDMARDs included anti-TNF (53%), -IL17 (22%), -IL23 (6%), -JAK (2%), and -PDE4 (19%). At 12 months, 73% patients retained the initial treatment: rates were similar for anti-TNF (78%), -IL17 (79%), -IL23 (83%), and JAK-inhibitors (86%), but lower for anti-PDE4 (53%; p<0.001). Females were more frequent among patients discontinuing therapy compared to retainers (65% vs 48%; p=0.003), also when focusing on anti-TNFs only, but not anti-IL17 or -IL23. On multivariable analysis (Figure), female sex was confirmed as a risk factor for discontinuation (OR 2.5, CI 2.0-4.7), whereas older age (OR 0.8, CI 0.6-0.97) and diabetes (OR 0.35, 95% CI 0.1-0.7) were associated with persistence. Discontinuation was higher with anti-PDE4 (OR 10, CI 6.4-109), anti-TNF (OR 5, 95%CI 2.6-43), and anti-IL17 (OR 4, 95%CI 1.8-34). Stratifying for drug classes, female sex was associated with anti-TNF discontinuation (OR 2.3, CI 1.2-4.3), the opposite was for older age (OR 0.6, CI 0.4-0.97) and sacroiliitis (OR 0.15, CI 0.0-0.5). Concerning anti-IL17, dactylitis (OR 0.1, CI 0.0-0.6), obesity (OR 0.05, CI 0.0-0.3), and anxiety-depression (OR 0.05, CI 0.0-0.4) were protective against discontinuation.
Conclusions. The overall 12-month b/tsDMARDs retention in PsA is influenced by sex, age, and comorbidities, while selected clinical domains may impact the retention of specific drug classes. Our findings, generated through a NLP analytical framework, underscore the potential of real-world data extraction from unstructured records to discover patterns of predictors of treatment outcomes.
Methods. For PsA patients attending our center between Jan2019 and Feb2024, the first visit in which a new b/tsDMARD was prescribed was retrieved, regardless of the treatment line. A BERT-based NLP model was developed to extract demographic, clinical, and treatment data. Treatment data were matched with dispensing records from the Hospital Pharmacy to verify which drugs were prescribed, discontinued, or switched during a 12-month follow-up. Model outputs were used to build a logistic regression model to predict drug retention. The model was trained/tested (k-fold cross-validation); performance metrics were computed, and odds ratios for predictors of retention and discontinuation were estimated.
Results. We included 413 PsA cases (52% female; age 52±12) initiating a new b/tsDMARD: 90% had peripheral arthritis, 15% axial disease, 18% dactylitis, 18% enthesitis, 61% skin psoriasis, and 2% inflammatory bowel disease. Comorbidities included depression-anxiety (7%), diabetes (10%), obesity (9%), and smoking (14%). New b/tsDMARDs included anti-TNF (53%), -IL17 (22%), -IL23 (6%), -JAK (2%), and -PDE4 (19%). At 12 months, 73% patients retained the initial treatment: rates were similar for anti-TNF (78%), -IL17 (79%), -IL23 (83%), and JAK-inhibitors (86%), but lower for anti-PDE4 (53%; p<0.001). Females were more frequent among patients discontinuing therapy compared to retainers (65% vs 48%; p=0.003), also when focusing on anti-TNFs only, but not anti-IL17 or -IL23. On multivariable analysis (Figure), female sex was confirmed as a risk factor for discontinuation (OR 2.5, CI 2.0-4.7), whereas older age (OR 0.8, CI 0.6-0.97) and diabetes (OR 0.35, 95% CI 0.1-0.7) were associated with persistence. Discontinuation was higher with anti-PDE4 (OR 10, CI 6.4-109), anti-TNF (OR 5, 95%CI 2.6-43), and anti-IL17 (OR 4, 95%CI 1.8-34). Stratifying for drug classes, female sex was associated with anti-TNF discontinuation (OR 2.3, CI 1.2-4.3), the opposite was for older age (OR 0.6, CI 0.4-0.97) and sacroiliitis (OR 0.15, CI 0.0-0.5). Concerning anti-IL17, dactylitis (OR 0.1, CI 0.0-0.6), obesity (OR 0.05, CI 0.0-0.3), and anxiety-depression (OR 0.05, CI 0.0-0.4) were protective against discontinuation.
Conclusions. The overall 12-month b/tsDMARDs retention in PsA is influenced by sex, age, and comorbidities, while selected clinical domains may impact the retention of specific drug classes. Our findings, generated through a NLP analytical framework, underscore the potential of real-world data extraction from unstructured records to discover patterns of predictors of treatment outcomes.
Metrics
Citations and online attention from Dimensions, Altmetric and PlumX.
License
Copyright (c) 2026 The Author(s)
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
PAGEPress has chosen to apply the Creative Commons Attribution NonCommercial 4.0 International License (CC BY-NC 4.0) to all manuscripts to be published.