Machine learning identifies a profile of inadequate responder to methotrexate in rheumatoid arthritis.

OBJECTIVES: Around 30% of patients with RA have an inadequate response to MTX. We aimed to use routine clinical and biological data to build machine learning models predicting EULAR inadequate response to MTX and to identify simple predictive biomarkers. METHODS: Models were trained on RA patients fulfilling the 2010 ACR/EULAR criteria from the ESPOIR and Leiden EAC cohorts to predict the EULAR response at 9 months (± 6 months). Several models were compared on the training set using the AUROC. The best model was evaluated on an external validation cohort (tREACH). The model's predictions were explained using Shapley values to extract a biomarker of inadequate response. RESULTS: We included 493 therapeutic sequences from ESPOIR, 239 from EAC and 138 from tREACH. The model selected DAS28, Lymphocytes, Creatininemia, Leucocytes, AST, ALT, swollen joint count and corticosteroid co-treatment as predictors. The model reached an AUROC of 0.72 [95% CI (0.63, 0.80)] on the external validation set, where 70% of patients were responders to MTX. Patients predicted as inadequate responders had only 38% [95% CI (20%, 58%)] chance to respond and using the algorithm to decide to initiate MTX would decrease inadequate-response rate from 30% to 23% [95% CI: (17%, 29%)]. A biomarker was identified in patients with moderate or high activity (DAS28 > 3.2): patients with a lymphocyte count superior to 2000 cells/mm3 are significantly less likely to respond. CONCLUSION: Our study highlights the usefulness of machine learning in unveiling subgroups of inadequate responders to MTX to guide new therapeutic strategies. Further work is needed to validate this approach.

Machine learning predicts response to TNF inhibitors in rheumatoid arthritis: results on the ESPOIR and ABIRISK cohorts.

OBJECTIVES: Around 30% of patients with rheumatoid arthritis (RA) do not respond to tumour necrosis factor inhibitors (TNFi). We aimed to predict patient response to TNFi using machine learning on simple clinical and biological data. METHODS: We used data from the RA ESPOIR cohort to train our models. The endpoints were the EULAR response and the change in Disease Activity Score (DAS28). We compared the performances of multiple models (linear regression, random forest, XGBoost and CatBoost) on the training set and cross-validated them using the area under the receiver operating characteristic curve (AUROC) or the mean squared error. The best model was then evaluated on a replication cohort (ABIRISK). RESULTS: We included 161 patients from ESPOIR and 118 patients from ABIRISK. The key selected features were DAS28, lymphocytes, ALT (aspartate aminotransferase), neutrophils, age, weight, and smoking status. When predicting EULAR response, CatBoost achieved the best performances of the four tested models. It reached an AUROC of 0.72 (0.68-0.73) on the train set (ESPOIR). Better results were obtained on the train set when etanercept and monoclonal antibodies were analysed separately. On the test set (ABIRISK), these models respectively achieved on AUROC of 0.70 (0.57-0.82) and 0.71 (0.55-0.86). Two decision thresholds were tested. The first prioritised a high confidence in identifying responders and yielded a confidence up to 90% for predicting response. The second prioritised a high confidence in identifying inadequate responders and yielded a confidence up to 70% for predicting non-response. The change in DAS28 was predicted with an average error of 1.1 DAS28 points. CONCLUSION: The machine learning models developed allowed predicting patient response to TNFi exclusively using data available in clinical routine.