Machine Learning Approach Effectively Predicts Remission in RA Following TNF Treatment


Lasso demonstrated advantages in predicting clinical remission in patients with RA treated with TNF, with a specificity of 69.9% and sensitivity of 61.7%.

Machine Learning Approach Effectively Predicts Remission in RA Following TNF Treatment

Koshiro Sonomoto, PhD

Credit: Lupus KCR 2023

A low-cost predictive machine learning model successfully predicted achievement of remission based on Clinical Disease Activity (CDAI) measures among a cohort of patients with rheumatoid arthritis (RA) after 6 months of treatment with tumor necrosis factor (TNF) inhibitors, according to research published in Rheumatology and Therapy.1

As investigators only used CDAI measures at baseline at month 6, they believe this method demonstrates the achievability of creating regional cohorts to create low-cost models tailored to specific institutions or geographical regions.

Currently, treatment guidelines suggest initiating therapy with methotrexate, followed by either biologic disease-modifying antirheumatic drugs (bDMARDs) or Janus kinase (JAK) inhibitors if methotrexate is ineffective. However, recent studies have favored bDMARDs over JAKs and, therefore, treatment with a TNF is becoming an increasingly popular option for this patient population. Despite these encouraging findings, only 70% of patients initiating a TNF show a favorable response.2

“Numerous efforts have been reported to predict the efficacy of TNF in advance,” wrote a team of investigators led by Koshiro Sonomoto, PhD, associated with the University of Occupational and Environmental Health in Japan. “However, access to these advanced technologies may be limited to certain countries and advanced facilities due to cost, labor requirements, and the need for process standardization. Conversely, predictive models based on routine clinical data are more accessible.”

Patients with RA beginning treatment with a TNF as the first targeted synthetic (ts)/bDMARD after inadequate response to methotrexate were included in the analysis. Data were collected from the FIRST registry between August 2003 and October 2022. The analysis of baseline characteristics and 6-month CDAI used a variety of machine learning approaches, such as lasso logistic regression (Lasso), logistic regression with stepwise variable selection, support vector machine, and decision tree, along with 48 factors available in routine clinical practice for the prediction model.

A total of 4706 patients in the FIRST registry initiated treatment with b/tsDMARDs during the study period, of which 2223 were receiving methotrexate. Of these patients, 1630 received a TNF and 79 received a JAK. The average age of patients was 59.2 years, they had an average body mass index of 22.1, a mean disease duration of 75.7 months, and all were Asian. The mean dose of methotrexate was 11.3 mg/week and the mean CDAI score was 26.1.

The models exclusively relied on patient-reported outcomes and quantitative parameters, as opposed to subjective physician input.

Of the approaches, Lasso demonstrated advantages in predicting CDAI remission, with a specificity of 69.9%, sensitivity of 61.7%, and a mean area under the curve of .704. Patients who were predicted to respond to TNF achieved CDAI remission at an average rate of 53.2%, compared with only 26.4% of predicted non-responders.

These results could also help to identify an alternative b/tsDMARDs class for patients who were predicted to be TNF non-responders.

Investigators mentioned limitations, including the decreasing accuracy of the Lasso-generated remission predictive model among a Calendar cohort. In this group, which was split into a 9:1 ratio with a cutoff of October 2019, an increase in censoring was reported in the validation cohort. They suggest this could be due to COVID-19 complications.

“This approach holds the potential to improve rheumatoid arthritis management by reducing the need for trial-and-error approaches and facilitating more personalized and effective treatment strategies,” investigators concluded. “While further validation is necessary, the study also suggests that creating cost-effective models tailored to specific regions or institutions is possible.”


  1. Sonomoto K, Fujino Y, Tanaka H, Nagayasu A, Nakayamada S, Tanaka Y. A Machine Learning Approach for Prediction of CDAI Remission with TNF Inhibitors: A Concept of Precision Medicine from the FIRST Registry. Rheumatol Ther. Published online April 18, 2024. doi:10.1007/s40744-024-00668-z
  2. Wijbrandts CA, Tak PP. Prediction of response to targeted treatment in rheumatoid arthritis. Mayo Clin Proc. 2017;92(7):1129–43
Related Videos
Ralph DeFronzo, MD | Credit: UT San Antonio
Timothy Garvey, MD | Credit: University of Alabama at Birmingham
3 KOLs are featured in this program.
3 KOLs are featured in this program.
Atul Malhotra, MD | Credit: Kyle Dykes; UC San Diego Health
Edward V Loftus, Jr, MD | Credit: Mayo Clinic
Katie Ross-Driscoll, PhD, MPH | Credit: Regenstrief Institute
Video 3 -  4 KOLs are featured in, "Collaborating with Specialists: Primary Care Clinician’s Evolving Role in HE Management"
Video 3 -  4 KOLs are featured in, "HE Experts on Symptoms, Risk Factors, Diagnosis Challenges, and Severity Grading"
© 2024 MJH Life Sciences

All rights reserved.