Date: Monday, October 22, 2018
Session Type: ACR Poster Session B
Session Time: 9:00AM-11:00AM
Rheumatoid arthritis (RA) is a chronic inflammatory disease associated with increased mortality and disability. Although different factors have been associated with prognosis, it is still difficult to predict the evolution of a specific patient. Therefore, our objective was to train and validate a predictive model of disease severity using radiological damage as a surrogate marker, based on Artificial Intelligence techniques, and using clinical and genetic data as predictors.
Four independent cohorts were included (959 patients with 1902 hand X-rays). Radiological damage was measured using the hands and wrists Sharp / van-der-Heijde score (SvdH). The variables to be predicted [total value of SvdH, erosion component (ES) and joint narrowing (NS)] were logarithmically transformed before analysis. As clinical predictors, age at onset of symptoms, sex, duration of the disease at the time of each radiograph, year of onset of symptoms and presence of rheumatoid factor were used. As genetic variables, the single nucleotide polymorphism data obtained from the Immunochip genotyping platform (Illumina) were used. In addition to an additive effect of the genetic variants, an interaction between each polymorphism and the duration of the disease was introduced in the analysis. Three cohorts were used for the selection of variables, generation of predictive models and internal validation. The fourth cohort was used to perform the external validation of the models. Regression trees with random effects (RTRE) were generated using the R package “REEMtree”. The goodness of fit of the models was measured using the root mean squared error (RMSE) and the intraclass correlation coefficient (ICC).
After the variable selection step, for the prediction of total SvdH, ES and NSLS, the RTRE selected 253, 235, and 192 unique sets of variables composed of a median (interquartile) 31 (26-38), 21 (17-26), and 34 (28-38) elements, respectively. Regarding interval validation, the lowest RMSEs were 3.16, 1.25 and 2.43 units of the Sharp / van-der-Heijde score, for total SvdH, ES and NSLS, respectively. The highest ICCs were 0.91, 0.88 and 0.92, respectively. Regarding external validation, the lowest RMSEs were 5.79, 3.34 and 4.09 units of the Sharp / van-der-Heijde score, respectively. The highest ICCs were 0.90, 0.77 and 0.89, respectively. We selected those sets of variables located in the lowest deciles of the RMSE for both the internal and the external validation cohorts: 4 data sets were selected for Total SvdH, 16 for ES, and 14 for NSLS, with 88 polymorphisms in combination.
It is possible to generate predictive models of radiological damage of great precision using Artificial Intelligence techniques. This could allow early stratification of patients according to prognosis. It is necessary to validate these models in other populations.
To cite this abstract in AMA style:Lezcano JM, Ivorra-Cortes J, Madrid A, Lopez-Mejías R, Martin J, Fernández-Gutiérrez B, González-Gay MA, Balsa A, Gonzalez-Alvaro I, Salazar F, Alcazar LA, Rodriguez-Rodriguez L. Development of a Predictive Model of Radiological Damage in Patients with Rheumatoid Arthritis Based on Artificial Intelligence [abstract]. Arthritis Rheumatol. 2018; 70 (suppl 10). https://acrabstracts.org/abstract/development-of-a-predictive-model-of-radiological-damage-in-patients-with-rheumatoid-arthritis-based-on-artificial-intelligence/. Accessed December 11, 2019.
« Back to 2018 ACR/ARHP Annual Meeting
ACR Meeting Abstracts - https://acrabstracts.org/abstract/development-of-a-predictive-model-of-radiological-damage-in-patients-with-rheumatoid-arthritis-based-on-artificial-intelligence/