According Rheumatic and musculoskeletal diseases.1
“With the increasing amount of data available in medicine, new tools are needed to extract information,” the researchers explained. “Machine learning algorithms learn patterns from data and assume they will recur in the future. These algorithms identify patterns and rules without being explicitly programmed to do so, enabling unbiased discoveries. This is particularly interesting in medicine to identify markers or combinations of markers hitherto unknown to physicians.
Data from the RA ESPOIR, a French multicentre, longitudinal and prospective cohort on early arthritis, were used to train the models. Patients included in the study received 1 or more injections of TNFi and met the 2019 American College of Rheumatology (ACR)/European Alliance of Associations for Rheumatology (EULAR) criteria for RA.
Endpoints included a good or moderate EULAR response, assessed 12 months after TNFi initiation, and a change in erythrocyte sedimentation rate Disease Activity Score (DAS28) at 12 months. The investigators compared the performance of the linear regression, random forest, CatBoost, and XGBoost models on the training set, then validated them using either the root mean square error or the area under the characteristic curve of operation of the receiver (AUROC). Among the models, the best performing model was further analyzed in a replication cohort, ABIRISK, a prospective study used to investigate predictors of anti-drug antibody development in RA patients treated with TNFi.
A total of 161 HOPE patients (95 receiving etanercept and 96 receiving anti-TNF monoclonal antibody) and 118 ABIRISK patients (68 receiving etanercept and 50 receiving either adalimumab or infliximab) were included in the study. In the ESPOIR cohort, 59% reported a response to TNFi treatment and 61% responded in the validation set (ABIRISK). Main characteristics focusing on DAS28, aspartate aminotransferase (ALT), neutrophils, lymphocytes, age, weight and smoking status.
Of the 4 models tested, CatBoost was able to predict the best EULAR response by achieving an AUROC of 0.72 (0.68–0.73) on the HOPE cohort train. Better results were obtained when etanercept and monoclonal antibodies were analyzed separately. In the ABIRISK cohort, models achieved an AUROC of 0.70 (0.57–0.82) and 0.71 (0.55–0.86), respectively.
In the end, CatBoost and random forest had the best performance. Two decision thresholds were tested, the first prioritizing high confidence in identifying responders and yielded up to 90% confidence in predicting response to TNFi. The second focused on high confidence in identifying inadequate responders, which gave up to 70% confidence in predicting non-response to TNFi. Changes in DAS28 were predicted with an average error of 1.1 DAS28 points.
Limitations included the relatively small sample size compared to the amount of data typically used in machine learning, which limits the accuracy of the algorithms and potentially prevents investigators from drawing strong conclusions. In addition, the use of data between the 2 cohorts resulted in variable delays between follow-ups. However, their performance was similar for both cohorts, highlighting the strengths of the study design. Finally, the limitations of the modeling may have resulted in an immortal temporal bias. Recurrent neural networks can help improve modeling of longitudinal data.
“Focusing on clinical use, we developed a model and evaluated its performance in two scenarios, having high confidence in identifying TNFi responders or in identifying TNFi inadequate responders,” the researchers concluded. “Both demonstrated interesting results compared to current clinical practice and these algorithms pave the way for a personalized treatment strategy in RA.”
Bouget V, Duquesne J, Hassler S, et al. Machine learning predicts the response to anti-TNFs in rheumatoid arthritis: results on the ESPOIR and ABIRISK cohorts. Open MDM. 2022;8(2):e002442. doi:10.1136/rmdopen-2022-002442