Prediction of Non-Response to First-Line Methotrexate Treatment in Rheumatoid Arthritis: A Real-World Data Analysis Using Machine Learning
Author(s)
Icten Z1, Starzyk K2, Friedman M3, Menzin J1
1Panalgo, Boston, MA, USA, 2OM1, Boston, MA, USA, 3Panalgo LLC, Boston, MA, USA
OBJECTIVES:
This study aims to identify predictors of non-response to methotrexate (MTX), the first choice among disease-modifying drugs (DMARDs) for most rheumatoid arthritis (RA) patients, using a longitudinal clinical cohort of RA patients and machine learning (ML).METHODS:
Adult RA patients initiating MTX between 01/01/2014 and 08/01/2020 without previous bDMARD/tsDMARD utilization were identified in the OM1 RA Registry (OM1, Inc; Boston, MA). First MTX date was the index date, with ≥1 Clinical Disease Activity Index (CDAI) and ≥2 healthcare encounters in the preceding 12 months. Patients initiating bDMARD/tsDMARDs within 30 days of index date or with other indication(s) for bDMARD/tsDMARDs were excluded. MTX non-response was defined as failing to attain remission or low disease activity (CDAI>10) within 4-8 months and/or the initiation of a bDMARD/tsDMARD within 8 months following index date. Features included demographics, comorbidities, baseline CDAI, body mass index, medications, procedures and healthcare resource utilization. Three-fold cross-validation was used to tune and evaluate traditional and regularized logistic regression, XGBoost, support vector machine, random forest and feed-forward neural network models. Best model was selected using the area under the ROC curve (AUC) and accuracy, recall, precision and specificity were assessed.RESULTS:
Among 6,648 eligible RA patients, 3,748 (56.4%) were classified as non-responders (mean age=59.8 years; females=79.1%). A total of 250 features were included; XGBoost model had the best AUC (AUC=76.4%; accuracy=70.9%; recall=75.1%; precision=73.9%; specificity=65.5%) with improvement over logistic regression (AUC=72.3%). Top predictors included increasing baseline CDAI, use of GABA analogs, analgesics, glucocorticoids, anxiolytics or sedatives, history of mood disorders, younger age and being female. Regardless of model type, inclusion of CDAI greatly improved model prediction.CONCLUSIONS:
Our study identified predictors of MTX non-response with strong predictive accuracy using ML. Further research is needed to better understand the potential role of comorbidities, like mood disorders, and their management on treatment success.Conference/Value in Health Info
2022-05, ISPOR 2022, Washington, DC, USA
Value in Health, Volume 25, Issue 6, S1 (June 2022)
Code
RWD88
Topic
Clinical Outcomes, Methodological & Statistical Research, Real World Data & Information Systems
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics, Clinical Outcomes Assessment, Health & Insurance Records Systems
Disease
Drugs, Musculoskeletal Disorders, Personalized and Precision Medicine, Systemic Disorders/Conditions