(T) Predictors of Treatment Changes in Patients with Rheumatoid Arthritis Using Machine Learning

Publication Number: 310

Background: Given treatment guidelines and options, understanding patterns of drug use in rheumatoid arthritis (RA) is critical.

Objectives: To evaluate the application of machine learning (ML) to real-world data from electronic health records (EHRs) to predict factors associated with treatment changes among patients with RA.

Methods: We used Optum® EHR data from Jan 2011-Mar 2019 that contained records on >80M lives to identify a cohort of patients diagnosed with RA and receiving 1st line treatment with a tumor necrosis factor inhibitor (TNFi) (not previously treated with a biologic disease modifying antirheumatic drug [bDMARD] or targeted synthetic [ts]DMARD). Changes in treatment were defined as 1) “cycling” from one TNFi to another and 2) “switching” from a TNFi to a non-TNFi bDMARD or tsDMARD within 12-months after TNFi initiation. A Bayesian methodology with an ensemble approach examined a network of at least 128 predictive models. 80% of patients were used to train our model; 20% were used to validate the model.

Results: Among 24,871 patients with RA who initiated a TNFi as their 1st line b/tsDMARD, mean age was 55.5 years, 77% were female, and the most common initiated TNFi was adalimumab (42%). During the 1-year follow-up, 3,870 patients (16%) cycled to another TNFi and 2,426 patients (10%) switched to another drug type. At baseline, compared to switchers, cyclers were 2.5 years younger on average, slightly more likely to be male and Asian, and more likely to have ≥1 comorbidity, but less likely to have ≥4 comorbidities. At baseline, cyclers also were more likely to be treated with an antimetabolite, folate, or a soluble TNF receptor agent, and less likely to be treated with an inflammatory bowel agent. Cyclers and switchers had similar use of procedures, hospitalizations, ER visits and outpatient visits. Factors that increased the likelihood of switching to another drug type were high non-rheumatic Charlson Comorbidity Index score, receiving cancer chemotherapy, and C-Reactive Protein being abnormal or not performed. Factors that increased the likelihood of TNFi cycling included having psoriasis. Model AUC on training data was 0.624; AUC on the validation dataset was 0.611.

Conclusions: ML methods can be leveraged to address questions in pharmacoepidemiology using real-world data. It provides improved sampling of the modelling space (including interactions among variables) and increases model predictive power. The approach also amplifies relationships best supported by the data and downplays spurious associations. However, predicting treatment changes remains challenging even with novel methodological approaches, as unmeasured factors (such as prescriber’s preference or patient’s beliefs) remain key determinants of change.