Prediction of primary non-response to methotrexate therapy using demographic, clinical and psychosocial variables: results from the UK Rheumatoid Arthritis Medication Study (RAMS)

Methotrexate (MTX) is now the conventional synthetic disease-modifying antirheumatic drug (csDMARD) of first choice, either as monotherapy or combination therapy, for most patients with rheumatoid arthritis (RA) [1]. This is emphasised in a number of international and national guidelines [2–5]. However, response to MTX, although better than to most other csDMARDs, is not universal. In observational studies approximately 30% of patients discontinue MTX in the medium term - around half due to inefficacy and half due to adverse events [6, 7]. Patient-related factors such as female gender and current smoking are associated with MTX non-response [6, 7]. Disease related factors such as disease duration, disease activity, rheumatoid factor (RF) and anti-citrullinated protein antibody (ACPA) status are moderately predictive of inefficacy [6, 7]. Psychosocial factors may also be important but have received little attention to date [8]. Response to treatment may be influenced by the patient’s social background [9, 10], by their existing beliefs about their illness and the likely efficacy of the drug, and by whether they have actually taken the medication (adherence) [11]. Genetic or other biological factors may also influence drug response [6, 7].

Many previous studies have attempted to identify independent predictors of response to MTX, although frequently without assessing the ability to assign probabilities of response to individual patients [12–14]. Those prediction models that have been developed have used data from the restricted populations and rigid treatment regimens of clinical trials [15, 16]; have used only small numbers of participants (fewer than 100) from observational studies [17, 18]; or have analysed the outcome of treatment discontinuation rather than a broader assessment of patient condition [19]. The value of predictions from such models for “real world” patients with RA about to start MTX for the first time is uncertain. More likely to be of use is a model developed in an observational study including patients seen in routine clinical practice using readily available or easily measurable demographic, clinical and psychosocial factors. If such a model could identify those unlikely to respond to MTX prior to starting therapy, with sufficient accuracy to be clinically useful, it could enable earlier access to alternative medications such as biologic therapy and the avoidance of disease progression for some patients.

The objectives of this study were, in a large national multi-centre observational study of patients with RA or undifferentiated polyarthritis (UP) commencing MTX for the first time, to (1) describe the pattern of 6-month treatment response, (2) identify patient-specific, disease-specific and psychosocial predictors of primary non-response to MTX, (3) combine predictors of non-response in a model that could be used to assign probability of non-response at the individual patient level and (4) test the accuracy of the model.

In this large observational study investigating response to MTX among patients with RA in the current era, 43% of patients were classified as non-responders by 6 months after starting treatment, with those discontinuing MTX due to adverse events excluded from the analysis. Baseline predictors of non-response in a multivariable logistic regression model were RF negativity, higher HAQ score, higher tender joint count, higher HADS anxiety score and lower disease activity. The AUC was 0.77 (0.74 optimism-corrected). The AUC was lower in models that excluded either all those in remission or all those with LDA at baseline when attempting to address the fact that it was harder for those with lower baseline DAS28 scores to achieve the definition of response. All models included RF negativity, lower baseline DAS28-CRP and higher HADS anxiety score as predictors of non-response. This is the first study to explore potential psychological factors in prediction of individual MTX non-response and all models retained HADS anxiety score as an independent predictor. Hence this psychological predictor added significant additional predictive information once the clinical predictors in the multivariable model had been accounted for. While the design of the current study does not allow us to examine the mechanism by which anxiety is associated with non-response, and this relationship requires further research, anxiety could be considered a modifiable risk factor, suggesting that the shared decision-making between patient and rheumatologist [2] should be mindful of anxiety issues, and that patient education prior to starting MTX should address anxiety. Although the literature on the association between patient anxiety and response to treatment in RA is limited, a recent study [31] did report that depression and anxiety (without differentiating between the two) may reduce likelihood of remission in those treated with MTX or tumour necrosis factor inhibitors.

While previous attempts have been made to develop models to predict response to MTX [15–19], this is the first model to be developed using a large cohort of patients with RA starting MTX for the first time and recruited from routine clinical care with few exclusions, i.e. representative of the setting in which such a model might be applied. That is, our model is designed to be applied in the real-world population of those about to commence MTX, which includes individuals with disease activity lower than might be expected. Final sets of predictor variables vary between published models, with only some measure of patient condition at baseline, be it TJC [15], DAS [16, 17] or HAQ [19], common to most studies. While lower baseline DAS28 was found to be associated with non-response to MTX when defined primarily using the EULAR response criteria, higher baseline DAS28 was associated with failure to achieve LDA (although significant only univariably, with higher TJC being retained in the multivariable model), which matches findings elsewhere [16]. This is a reminder that models are outcome-specific and that the clinical relevance of outcomes should be considered.

If a predicted probability of 0.9 or above from the model was used to identify patients at high risk of non-response, 100% of those meeting this criterion would go on to be non-responders. However, only 4% of non-responders would be identified in this way. Reducing the cut-off to 0.8 would identify 14% of non-responders, but at the expense of 2% of those labelled as high risk actually being patients who would respond to MTX. The trade-off between the delay in accessing alternative medications for those who will not respond to MTX but are predicted to do so, and the over-treatment with alternative medications of those who would respond to MTX but are predicted not to, is unlikely to be an equally weighted one. Deciding where to draw an appropriate threshold for a label of high risk for clinical practice requires consideration of the treatment options available in a particular setting, their benefits and risks for individual patients, and their health economic implications. Of course, even with perfect prediction of non-response to MTX therapy, there is no guarantee that a better response would be achieved with alternative treatments. Truly informed decision-making by clinicians and their patients would require personalised predictions of patient outcomes for a range of treatment options, a scenario which is still some way off.

The strengths of this analysis include the large sample size and the fact that it reflects use of MTX according to current guidelines and so supersedes earlier work [12–19]. The definition of non-response at 6 months embraced those who remained on the drug but had not exhibited enough improvement to be classified as moderate or good EULAR responders and also those who had discontinued the drug due to inefficacy or started a bDMARD. Other strengths were the inclusion of potential psychological predictors and using multiple imputation to provide reassurance of the robustness of the results to missing predictor data.

This study also has some limitations. The non-response rate was high. This may, in part, be due to suboptimal dosing or route of administration. Although the study did not dictate the treatment protocol, all patients should have been managed according to national guidelines [3] in which escalation of MTX is permitted and combination therapy is encouraged. We do not know the reasons for any deviations from treatment guidelines, but this prediction model may help guide clinicians as to which patients are less likely to respond as they are currently practicing and encourage them to consider treating more intensively. As we have attempted to predict non-response using only information available before the commencement of MTX, we have not considered the relationship between time-varying MTX dose and non-response. Titration is highly influenced by starting dose and patient response to treatment. The goal of the current work is to try and predict response prior to the start of MTX and, although maximum MTX dose and rate of titration may also be associated with response, that information would not be available pre-treatment. Further research is required to specifically investigate how rate and characteristics of MTX titration may also influence response, taking early response and adverse events into account. It seems likely that the observed association between lower DAS28 at baseline and subsequent non-response is explained by less scope for improvement (the key component of response) for those with lower disease activity. Since patients were recruited from routine clinical care, a high proportion of patients (41%) received oral or intramuscular steroids between the decision to prescribe MTX and the baseline assessment. We therefore performed sensitivity analysis, stratifying by steroid use, and the results were very similar (results not shown). The aim of this study was to use demographic, clinical and psychosocial variables that are readily available or easily measurable, to predict non-response. In future we aim to add genetic and metabolic predictors with the hope of improving the accuracy and clinical applicability of the model. Finally, these models were only validated internally. RAMS continues to recruit new patients so there will be an opportunity for further internal validation (“temporal validation”) in the future. However, external validation in an independent dataset, which includes information on the relevant predictor variables that would be needed before the models could be considered for clinical use.

It might be reasonable to consider by-passing MTX therapy altogether in patients predicted to be very unlikely to respond. This model assigns ≥ 90% probability of non-response to only a tiny proportion (2%) of patients. Would it be reasonable to accept a lower probability of non-response as a guide to MTX prescription - or a guide to starting combination therapy? This depends to some extent on the alternative forms of treatment and their efficacy in individuals predicted not to respond to MTX, and their cost. In the current situation it seems reasonable to continue to prescribe MTX for most patients in whom it is not contra-indicated but with a low threshold to move on to stronger combination or biological therapy if there is non-response at 6 months.

We have developed a model to predict non-response to MTX using data from a large contemporary observational study of patients with RA and UP commencing MTX for the first time. This is the first such model to consider patient-specific, disease-specific and psychosocial predictors. Using a high predicted probability to classify patients as at high risk of non-response would identify a small proportion of such individuals with perfect specificity. Patient anxiety was a multivariable predictor of non-response to MTX, a relationship that requires further research and which could be addressed prior to treatment commencement.