Methotrexate

Background

Low-dose weekly methotrexate is the first-line glucocorticoid-sparing drug in the treatment of inflammatory diseases such as RA, PsA and PsO resistant to topical treatment and/or phototherapy.240–244 It is also used in the treatment of steroid-dependent IBD, and is often combined with biologics.150,245,246 Methotrexate use has increased in the biologic era with only a minority of patients (approximately 3%) with RA, and PsO +/− PsA prescribed biologics in European countries.237,247–249 Although effective and well tolerated, methotrexate can cause cytopenia, elevated liver enzymes and AKI, mostly during the first few months of treatment.250,251 Concern about these side effects has led to the recommendations to undertake 2- to 4-weekly blood tests for FBC, LFT, and urea electrolytes and creatinine during the initial few months of treatment followed by 3-monthly testing for everyone thereafter.13,129

We previously reported that clinically significant cytopenia, elevated liver enzyme and AKI that required methotrexate discontinuation were infrequent after the first year of methotrexate treatment.252 To better understand this risk, we developed and validated a prognostic model for estimating the probability of clinically significant methotrexate toxicity during long-term treatment and stratify those at greatest and least risk.253

Methods

Data source: Data from the Clinical Practice Research Datalink (CPRD) Gold and Aurum were used for model development and validation, respectively.254,255 CPRD is an anonymised longitudinal database of electronic health records originated during clinical care in the NHS. With almost universal coverage of UK residents, participants that contributed data to the CPRD are representative of the UK population.254 The CPRD includes information on demographic details, lifestyle factors (e.g. smoking, alcohol intake), diagnoses, results of blood tests and details of primary care prescriptions. CPRD Gold and Aurum complement each other in terms of coverage of general practices due to their use of different software for data capture. Some general practices that have contributed data to both databases were identified using a bridging file provided by CPRD and were only included in the development cohort.

Ethical approval: This study was approved by the Independent Scientific Advisory Committee of the CPRD, Reference: 19_275R, 20_000236R.253

Study design: Retrospective cohort study.

Study period: 1 January 2007–31 December 2019.

Study population: Participants aged ≥ 18 years with a new diagnosis of inflammatory condition (e.g. RA, PsO, PsA, IBD etc.) and prescribed methotrexate by their GP were eligible. Patients were required to have ≥ 1-year disease-free registration with their current general practice to be classified as having a new diagnosis.6,256 This minimised the chance of patients on long-term methotrexate treatment for established disease appearing as new users of methotrexate when they changed their general practice.6,256 Additionally, patients were required to have received their first methotrexate prescription either after the first record of disease in the CPRD or in the 90 days preceding. This 90-day period was allowed because recording of diagnosis may lag prescriptions. Patients with severe chronic liver disease, CKD stage ≥ 4, or severe haematological diseases prior to start of follow-up were excluded as these conditions contraindicate methotrexate use and would have caused substantial uncertainty in outcome ascertainment.253

Methotrexate prescriptions:253 In the UK, initial methotrexate prescription and dose escalation occurs at hospital outpatient clinics. During this period, hospital specialists organise monitoring blood tests and act on abnormalities. Once a patient is established on treatment, that is reaches a stable, well-tolerated and effective methotrexate dose, typically approximately 6 months after starting treatment, the responsibility for prescribing and monitoring with periodic blood tests is handed to the patients’ GP as per NHS shared-care protocols.257,258 During shared-care prescribing, the GP seeks advice from hospital specialists if there are side effects including abnormal blood-test results, and treatment changes are directed by the latter.

Start of follow-up: Patients were followed up from 180 days after their first primary care methotrexate prescription until the earliest of outcome, death, transfer out of practice, 90-day prescription gap, last data collection from practice, 31 December 2019 or 5 years.253

Outcome: Methotrexate-toxicity-associated drug discontinuation was the outcome of interest.253 This was defined as a prescription gap of ≥ 90 days with either an abnormal blood-test result or a diagnostic code for abnormal blood-test result within ± 60 days of the last prescription date.252 The threshold for abnormal blood-test results were: total leucocyte count < 3.5 × 10⁹/l; neutrophil count < 1.6 × 10⁹/l; platelet count < 140 × 10⁹/l; ALT/AST > 100 IU/ml; and kidney function decline defined as either CKD progression based on medical codes recorded by the GP, or a creatinine increase of > 26 μmol/l, the threshold for consideration of AKI.259,260 In our previous validation study, only 5.4% of abnormal blood-test results in this time window were potentially explained by an alternate illness.252

Predictor selection: Clinical members of the study team suggested predictors based on their clinical expertise and knowledge of the literature.253 Age, sex, BMI, alcohol intake and diabetes were included as they are associated with DILI, and current smoking was included as it associates with methotrexate non-response, potentially requiring higher doses.261,262 CKD was included as it reduces methotrexate clearance.263 Disease type was included as patients with PsO are at higher risk of elevated liver enzymes with methotrexate than those with RA.252

Methotrexate dose was included to account for potential dose-dependent toxicity.253 Statins, NSAIDs, aspirin (≥ 300 mg/day), paracetamol, proton pump inhibitor, carbamazepine, levetiracetam and valproate were included as their use is associated with methotrexate toxicity relevant to this study as per the British National Formulary. Hydroxychloroquine was included as it increases exposure to methotrexate.264 Sulfasalazine, 5-ASA and other immune-suppressing drugs were included as they can cause cytopenia, elevated liver enzymes and AKI.

Either cytopenia (neutrophil count < 2 × 10⁹/l, total leucocyte count < 4 × 10⁹/l, or platelet count < 150 × 10⁹/l) or elevated transaminase (ALT and/or AST > 35 IU/l) during the first 6 months of primary care methotrexate prescription, that is, before the start of follow-up in this study, were included as a prognostic factor because they predicted cytopenia and/or transaminitis in other studies.41,265

The latest record of demographic and lifestyle factors, diseases recorded within 2 years prior to start of follow-up, and latest primary care prescription within 6 months prior to start of follow-up were used to define predictors except for CKD stage 3 that was defined using both GP records and/or eGFR 30–59 ml/minute. GPs typically review patients with long-term conditions annually.253 A 2-year lookback was utilised to minimise the risk of missing data from those that did not attend in the last year.

Sample size: For model development, we adhered to the formulae of Riley et al.,266 which gave a minimum sample size needed to minimise model overfitting (a target shrinkage factor of 0.9) and ensure precise estimation of overall risk of 1398 participants (189 outcomes) based on a maximum of 20 parameters, Cox–Snell R² value of 0.12, estimated outcome rate of 0.057/person-year, a 5-year time horizon, and a mean follow-up period of 2.36 years using the findings from our earlier work.252 The sample size for external model validation was much larger than the typically recommended minimum sample size of 200 events.

Statistical analysis: Multiple imputation handled missing predictor data on BMI, alcohol intake and methotrexate dose using chained equations.267 We carried out 10 imputations in the development data set and 5 imputations in the validation data set – a pragmatic approach considering the larger size of CPRD Aurum. The imputation model included all candidate predictors, Nelson–Aalen cumulative hazard function and outcome variables. The methodology has been reported in our previous publication.253

Model development: Fractional polynomial regression (first-order) analysis was used to model non-linear risk relationships with continuous predictors and was found not to be better than the linear terms (p > 0.05, comparing models of linear terms with the best fitting first-order polynomials); hence continuous predictors were not transformed.253 All 11 candidate predictors (17 parameters) were included in the Cox model and coefficients of each predictor estimated and combined using Rubin’s rule across the imputed data sets.253 The risk equation for predicting an individual’s risk of methotrexate discontinuation with abnormal blood-test results by 5-year follow-up was formulated using the development data. The baseline survival function at t = 5 years, a non-parametric estimate of survival function when all predictor values are set to zero, was estimated along with the estimated regression coefficients (β) and the individual’s predictor values (X). This led to the equation for the predicted absolute risk over time:268

Predicted risk of methotrexate-toxicity-associated drug discontinuation at 5 years = 1 – S₀(t_= 5)^exp(Xβ) where S₀(t_= 5) is the baseline survival function at 5 years of follow-up and βX is the linear predictor, β₁ × ₁ + β₂ × ₂ + … + β_px_p.

Model internal validation and shrinkage: The performance of the model in terms of calibration (i.e. the agreement between predicted and observed risks) was assessed by plotting agreement between predicted and observed outcomes, using a smoothed non-parametric calibration curve across the whole range.253 A curve close to the 45° line is ideal (quantified by a calibration-in-the-large of zero, calibration slope of 1).

Internal validation was performed to correct model performance estimates for optimism due to overfitting by bootstrapping with replacement 500 samples of the development data.253 The full model was fitted in each bootstrap sample and its performance was quantified in the bootstraps (apparent performance) and original samples (test model performance).253 Model optimism was calculated as the difference between test performance and apparent performance.253 A uniform shrinkage factor was estimated as the average of calibration slopes from each of the bootstrap samples.253 This process was repeated for all 10 imputed data sets, and the final uniform shrinkage calculated by averaging across the estimated shrinkage estimates from each imputation.253 Optimism-adjusted estimates of performance for the original model were then calculated as the original apparent performance minus the optimism.

To account for overfitting during model development process, the original β coefficients were multiplied by the final uniform shrinkage factor and the baseline hazards re-estimated conditional on the shrunken β coefficients to ensure that overall calibration was maintained, producing a final model.253 The D statistic, a measure of discrimination (i.e. ability to discriminate between those with and those without the outcome), interpreted as a log HR, the exponential of which gives the HR comparing two groups above or below the median of the linear predictor, was calculated.269,270 R²_D, a measure of variation explained by the model and Harrell’s c-statistic, a measure of the model’s predictive accuracy, was calculated.

Model external validation: External validation was performed using data from the CPRD Aurum.253 There were no differences from the development data in setting, eligibility criteria, outcome and predictors.253 The final developed model equation was applied to the validation data set, and calibration and discrimination were examined as described above.269,270 Calibration of 5-year risks was examined by plotting agreement between estimated risk from the model and observed outcome risks. In the calibration plot, predicted and observed risks were divided into 10 equally sized groups defined by 10ths of predicted risk.253 Additionally, pseudo-observations were used to construct smooth calibration curves across these groups via a running non-parametric smoother. Separate graphs were plotted for each imputation of the validation cohort, and an example of one plot is shown in the results. R²_D and Harrell’s c-statistic were calculated for the validation cohort. Subgroup analyses considered age groups, methotrexate doses, routes of administration and disease type. Stata-MP version 16 was used for all statistical analyses. This study was reported in line with the transparent reporting of a multivariable prediction model for individual prediction or diagnosis (TRIPOD) guidelines.271

Results

Participants: Data for 13,110 and 23,999 participants that contributed 34,298 and 67,150 person-years follow-up were included in the development and validation cohorts, respectively (see Appendix 5, Figures 16 and 17; Table 19).253

FIGURE 2 Calibration of a prognostic model for methotrexate discontinuation with abnormal monitoring blood-test results at 5 years in the derivation cohort.253 Reproduced with permission from Nakafero et al.1 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) licence, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original text. (Figure not reproduced here; see the original article.)

Model development: Eleven candidate predictors (17 parameters) were included in the model. In the development data set, there were 854 outcomes in 6.5% of patients (n = 13,110) during the follow-up period at a rate (95% CI) of 24.90 (23.28 to 26.63)/1000 person-years.253

As expected, before shrinkage, the calibration slope (95% CI) in the development data was 1.00 (0.89 to 1.11).253 From the bootstrap, a uniform shrinkage factor of 0.93 was obtained and used to shrink predictor coefficients in the final model for optimism and after re-estimation, the final model’s cumulative baseline survival function (S₀) was 0.895 at 5 years of follow-up (Table 20).253

Model performance in development cohort: Figure 2 and Appendix 5, Figure 18, show a calibration plot of the final (i.e. after shrinkage) model at 5 years, with average model predictions closely matching the average observed outcome probabilities across all 10 groups of patients, with CIs overlapping the 45° line (perfect prediction line). Data from a single imputed data set; So(t = 5) 0.895 was used.

The optimism-corrected R²_D was 0.13. Royston D statistic (95% CI) was 0.89 (0.78 to 1.00), corresponding to a HR (95% CI) of 2.44 (2.18 to 2.72) comparing the risk groups above and below the median of linear predictor. The optimism-adjusted Royston D statistic (95% CI) was 0.79 (0.69 to 0.90) corresponding to HR (95% CI) 2.20 (1.99 to 2.61) (Table 21).

Model performance in validation cohort: There were 1486 outcomes in 6.2% patients (n = 23,999) at a rate (95% CI) of 22.13 (21.03 to 23.28)/1000 person-years in the validation cohort. The calibration slope (95% CI) across the 5-year follow-up period was 0.94 (0.85 to 1.02). The calibration plot showed reasonable correspondence between observed and predicted risk at 5 years across the groups defined by tenths of risk (see Appendix 5, Figure 19). The smoothed calibration curve also showed good agreement of the predicted risk to the observed risk. [Figure 3; data from a single imputed data set were used for illustration; S₀(t = 5) 0.895. Solid black line reflects perfect prediction.] Model performance was also tested at years 1, 2, 3 and 4 (see Appendix 5, Figures 20–23) and showed a similar pattern.

FIGURE 3 Calibration of a prognostic model for methotrexate discontinuation with abnormal monitoring blood-test results at 5 years in the validation cohort.253 Reproduced with permission from Nakafero et al.1 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) licence, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original text. (Figure not reproduced here; see the original article.)

Model discrimination in the development and validation data was broadly similar (see Table 21). The Royston D statistic (95% CI) in the validation data was 0.75 (0.67 to 0.83), corresponding to a HR (95% CI) of 2.12 (1.95 to 2.29). The optimism-corrected R²_D (95% CI), and Harrell’s c-statistic (95% CI) were 0.12 (0.10 to 0.14) and 0.64 (0.62 to 0.65), respectively. The model calibration slope and discrimination were similar across subgroups defined by age (< 60 years, and ≥ 60 years), methotrexate dose (≤ 15 mg/week and > 15 mg/week), route of administration (oral or subcutaneous) and disease type (RA, and others) (see Appendix 5, Figures 24–27).

Worked examples: Ten anonymised patient profiles, 1 from the middle of each of the 10 groups defined by deciles of predicted risk, were selected from the derivation cohort and the risk equation was applied to them (see Appendix 4, Table 115). The cumulative probability of outcome over 5 years ranged from 7.1% in the middle of first group to 9.4% in the middle of seventh group and 23.7% in the middle of 10th group.

Thiopurines

Introduction

Thiopurines are a cornerstone glucocorticoid-sparing drug for the treatment of IBD, widely recommended globally for maintenance of remission in UC and Crohn’s disease (CD).12,272–274 They are among first-line glucocorticoid-sparing drugs for the treatment of SLE, and are used in the management RA and PsO +/− PsA.275,276

Thiopurines can cause myelotoxicity, hepatotoxicity and nephrotoxicity, mostly during the first few months of treatment, and dose reduction is considered in severe renal impairment.277–282 Concern about these issues led to recommendations to undertake 2- to 4-weekly monitoring blood tests during the first few months of treatment followed by testing at 3-monthly intervals thereafter.12,13,276 Regular monitoring of blood tests during established long-term thiopurine treatment, a period when blood-test abnormalities are uncommon, seems unduly cautious and unnecessary.

Whether routinely available clinical data can be used to ascertain those at low, medium or high risk of toxicity in this period, and be used to inform risk-stratified monitoring has not been evaluated. To inform such risk-stratified monitoring, we developed and validated a prognostic model for clinically significant thiopurine toxicity detectable on monitoring blood tests.

Methods

Methods were same as for methotrexate, except for the fact that data from the CPRD Aurum and Gold were used for model development and validation, respectively. This approach was chosen as CPRD Aurum has larger sample size than CPRD Gold.

Predictors were selected by clinical members of the team. Age, sex, BMI, alcohol intake and diabetes were included as they associate with DILI.261,262 Smoking was included as it associates with active RA, CD and, SLE that may require the use of higher thiopurine doses.262,283,284 CKD was included as it reduces thiopurine clearance.285 The mercaptopurine equivalent dose was included in the model as thiopurine toxicity is dose dependent.286,287 Statins and angiotensin-converting enzyme (ACE) inhibitors were included as their use is associated with thiopurine toxicity.282 Allopurinol was included as it inhibits thiopurine metabolism by inhibiting xanthine oxidase and coprescription can cause cytopenia if the dose of thiopurine is not reduced substantially, for example to 25% of the regular dose.282 Additionally, low-dose allopurinol may be combined with low-dose azathioprine in patients with prior thiopurine-induced hepatotoxicity to minimise the future risk of hepatotoxicity and myelotoxicity in hyper-methylators and this approach was shown to have greater likelihood of achieving remission in a clinical trial.288–292 Sulfasalazine, 5-ASA and other immune-suppressing drugs were included as they can cause cytopenia, elevated liver enzymes and AKI. Additionally, 5-ASA drugs are associated with thiopurine-induced myelotoxicity as they increase 6-thioguanine nucleotide (6-TGN) levels.293,294 Either cytopenia (neutrophil count < 2 × 10⁹/l, total leucocyte count < 4 × 10⁹/l, or platelet count < 150 × 10⁹/l) or elevated transaminase (ALT and/or AST > 35 IU/l) during the first 6 months of primary care prescription was included as they predicted cytopenia and/or transaminitis in other studies.41,265

Sample size

For model development, assuming an event rate of 17 per 1000 person-years from previously published studies and an average follow-up of 3.19 years, the minimum sample size needed to minimise model overfitting (a target shrinkage factor of 0.9) and ensure precise estimation of overall risk was 1748 participants (95 outcomes) for a maximum of 25 parameters, Cox–Snell R² value of 0.12, a 5-year time horizon, using the formulae of Riley et al.190,266 The sample size for external model validation was much larger than the typically recommended minimum sample size of 200 events.

Statistical analysis was as per the methotrexate study. Age group, inflammatory disease type and whether the patient was commenced on thiopurine after the year 2010 formed the basis of subgroup analyses. This study was reported in line with the TRIPOD guidelines.271

Results

Data for 5982 and 3573 participants that contributed 16,117 and 9075 person-years follow-up were included in the development and validation cohorts, respectively (see Appendix 5, Figures 28 and 29). The distribution of disease and demographic factors were similar between cohorts (Table 22).

Model development: Fourteen candidate predictors (21 parameters) were included. Continuous predictors were not transformed as first-degree non-linear risk relationships were no better than linear terms (p > 0.05). Four hundred and five outcomes occurred during follow-up at an incidence (95% CI) of 25.13 (22.80 to 27.70)/1000 person-years. From the bootstrap, a uniform shrinkage factor of 0.80 was applied to all predictor coefficients. The final model’s cumulative baseline survival function (S₀) was 0.938 at 5 years (Table 23).

The average model predictions matched the average observed outcome probabilities across all 10 groups of patients, with CIs overlapping the 45° line. (Figure 4; data from a single imputed data set were used for illustration; S₀o(t = 5) 0.938. Solid black line reflects perfect prediction.)

FIGURE 4 Calibration of a prognostic model for thiopurine discontinuation with abnormal monitoring blood test results at 5 years in the derivation cohort. Reproduced with permission from Nakafero et al.1 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) licence, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original text. (Figure not reproduced here; see the original article.)

The calibration curve at 5 years showed some mis-calibration at the individual level in the higher risk patients; however, there were less data at the higher risk probabilities with wide 95% CIs (see Figure 4, Appendix 5, Figure 30). Royston D statistic (95% CI) was 0.91 (0.75 to 1.07), corresponding to HR (95% CI) 2.48 (2.12 to 2.92). The optimism-adjusted Royston D statistic was 0.76 corresponding to HR 2.13 (Table 24).

Model performance in the validation cohort: There were 269 outcomes incidence (95% CI) of 29.64 (26.30 to 33.40)/1000 person-years. The calibration slope (95% CI) across the 5-year follow-up period was 1.10 (0.84 to 1.36). The calibration plot showed reasonable correspondence between observed and predicted risk at 5 years across the tenths of risk (see Appendix 5, Figure 31). The smoothed calibration curve showed alignment of the predicted risk to the observed risk at low risk. [Figure 5; data from a single imputed data set were used for illustration; S₀(t = 5) 0.938. Solid black line reflects perfect prediction.]

FIGURE 5 Calibration of a prognostic model for thiopurine discontinuation with abnormal monitoring blood-test results at 5 years in the validation cohort. Reproduced with permission from Nakafero et al.1 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) licence, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original text. (Figure not reproduced here; see the original article.)

Model performance at years 1, 2, 3 and 4 showed a similar pattern (see Appendix 5, Figures 32–35). Model discrimination in the validation cohort was broadly like that in derivation cohort (see Table 24). The Royston D statistic (95% CI) was 0.72 (0.52 to 0.92), corresponding to HR (95% CI) 2.05 (1.68 to 2.51). The model performed well across age groups, in those with IBD or other inflammatory diseases, and in patients starting thiopurines before 2010 or in 2010 or later (see Appendix 5, Figures 36 and 37). Worked examples across deciles are illustrated in Appendix 4, Table 116.

Sulfasalazine

Introduction

Sulfasalazine is commonly used in the treatment of inflammatory diseases such as RA, PsA, axial spondylarthritis, ReA and infrequently in the management of IBD (the latter is mostly treated with 5-ASAs due to a better safety profile).295–297 Although effective, sulfasalazine can cause cytopenia and elevated liver enzymes typically in the first 3–6 months of treatment, although late-onset toxicity is reported.95,298–309 Sulfasalazine can also cause crystalluria and interstitial nephritis, and is not recommended in those with severe renal impairment.282 Cautious use is recommended in those with mild to moderate renal impairment.282

There is considerable inconsistency in guidance on how to monitor patients on long-term sulfasalazine treatment. The British Society for Rheumatology (BSR) guidelines recommend 2- to 4-weekly blood tests for FBC, LFT and UEC for the first 3 months of treatment followed by 3-monthly testing in the first year and no further monitoring blood tests thereafter.13 On the contrary, the American College of Rheumatology (ACR) guidelines recommend close monitoring for the first 3 months of treatment, followed by 3-monthly blood testing for FBC, UEC and LFT during the entire duration of treatment.310 The summary of product characteristics for sulfasalazine recommends monitoring with FBC, LFT and UEC at 3-monthly intervals during long-term treatment.311 To predict clinically significant laboratory abnormalities during established sulfasalazine treatment and to inform the frequency of testing, we have developed and validated a prognostic model for clinically significant myelotoxicity, hepatotoxicity and/or nephrotoxicity due to sulfasalazine.

Methods

The methodology was same as that used for methotrexate, except that data from CPRD Aurum and Gold were used for model development and validation, respectively.254,255 This has been published in our previous manuscript.3 Age, sex, BMI, alcohol intake and diabetes were included as they associate with DILI.261,262 Individual inflammatory diseases were considered separately because sulfasalazine toxicity is reported to be less common in people with IBD than in those with RA.297 CKD stage 3 was included as it reduces sulfasalazine clearance.285 Statins, carbamazepine, valproate and paracetamol were included as their use is associated with sulfasalazine toxicity as per the British National Formulary. Methotrexate, leflunomide and thiopurines were included as they can cause cytopenia, elevated liver enzymes and AKI. Either cytopenia (neutrophil count < 2 × 10⁹/l, total leucocyte count < 4 × 10⁹/l, or platelet count < 150 10⁹/l) or elevated transaminase (ALT and/or AST > 35 IU/l) during the first 6 months of primary care prescription was included as they predicted cytopenia and/or transaminitis in other studies.41,265

Sample size: In a previously published cohort of 1321 RA patients, 85 stopped sulfasalazine with neutropenia, thrombocytopenia, or elevated liver enzymes during a mean follow-up of 2.39 years.95 Assuming a similar incidence of treatment discontinuation for model development, the minimum sample size needed to minimise model overfitting (a target shrinkage factor of 0.9) and ensure precise estimation of overall risk was 1748 participants (113 outcomes) based on a maximum of 25 parameters, Cox–Snell R² value of 0.12, outcome rate of 0.027/person-year,95 a 5-year time horizon and a mean follow-up period of 2.39 years using the formulae of Riley et al.266 The sample size for external model validation was much larger than the typically recommended minimum sample size of 200 events.312

Statistical analyses were as for methotrexate, except for subgroup analyses considered age group and inflammatory disease type (RA vs. others).

Results

Data for 8936 and 5203 participants contributing 23,299 and 12,867 person-years follow-up were included in the derivation and validation cohorts, respectively (see Appendix 5, Figures 38 and 39) (Table 25).

Model development: Nine candidate predictors (21 parameters) were included. In the derivation data set, 473 outcome events occurred during the follow-up period at a rate (95% CI) of 20.30 (18.55 to 22.22) per 1000 person-years. Of these, 256, 131 and 113 patients respectively stopped treatment due to cytopenia, renal function decline and elevated liver enzymes. These events occurred throughout the 5-year follow-up period when the entire cohort was considered (see Appendix 5, Figure 40) and when patients coprescribed either methotrexate or leflunomide or thiopurine with sulfasalazine were excluded (see Appendix 5, Figure 41). From the bootstrap, a uniform shrinkage factor of 0.84 was obtained and used to shrink predictor coefficients in the final model for optimism and after re-estimation, the final model’s cumulative baseline survival function (S₀) was 0.940 at 5 years of follow-up (Table 26).

Model performance in the development cohort: As expected, the calibration slope (95% CI) in the development data was 1.00 (95% CI 0.85 to 1.15). Calibration plot of the final (i.e. after shrinkage) model at 5 years showed that the average model predictions matched the average observed outcome probabilities across 10 groups of patients, with CIs overlapping the 45° line (perfect prediction line) [Figure 6; data from a single imputed data set was used; S₀(t = 5) 0.940].

FIGURE 6 Calibration of a prognostic model for sulfasalazine discontinuation with abnormal monitoring blood-test results at 5 years in the development cohort. Reproduced with permission from Abhishek et al.3 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) licence, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original text. (Figure not reproduced here; see the original article.)

As most patients had a low risk of outcome, most of the deciles clustered at the bottom left of the calibration plot (see Appendix 5, Figure 42). The smoothed calibration curve at 5 years showed alignment of observed risk to the predicted risk with wide CIs at high-risk probabilities (see Figure 6). The Royston D statistic was 0.91 (95% CI 0.77 to 1.05), corresponding to a HR (95% CI) of 2.48 (2.16 to 2.86) comparing the risk of participants who were above the median of linear predictor to that below the median. The optimism adjusted Royston D statistic was 0.79, corresponding to a HR of 2.20 (Table 27).

Model performance in the validation cohort: There were 280 outcomes at a rate (95% CI) of 21.76 (19.36 to 24.47)/1000 person-years in the validation cohort. The calibration slope (95% CI) across the 5-year follow-up period was 1.19 (0.96 to 1.43). [Figure 7; data from a single imputed data set was used; S₀(t = 5) 0.940].

FIGURE 7 Calibration of a prognostic model for sulfasalazine discontinuation with abnormal monitoring blood-test results at 5 years in the validation cohort. Reproduced with permission from Abhishek et al.3 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) licence, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: https://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original text. (Figure not reproduced here; see the original article.)

The calibration plot showed reasonable correspondence between observed and predicted risk at 5 years across the tenths of risk (see Appendix 5, Figure 43). Most of the deciles clustered at the bottom left of the calibration plot due to a low risk of outcome for most patients (see Appendix 5, Figure 43). When individual risks were plotted, the smoothed calibration curve showed alignment of the predicted risk to the observed risk at low risk and wide CIs overlapping the perfect prediction line at high-risk probabilities (see Figure 7). Model performance was also tested at years 1, 2, 3 and 4 (see Appendix 5, Figures 44–47) and showed a similar pattern except for overprediction of risk at 1 year. The Royston D statistic in the validation data was 0.87 (0.67 to 1.07), corresponding to a HR (95% CI) of 2.39 (1.95 to 2.92). Model discrimination in the derivation and validation data was broadly similar (see Table 27). The model performed well in those younger or older than 60 years, in those with RA or other conditions (see Appendix 5, Figures 48 and 49).

Worked examples: Ten anonymised patient profiles, 1 from the middle of each of the 10 groups defined by deciles of predicted risk, were selected from the development cohort, the higher the decile group the higher the risk, and the risk equation was applied to each. The cumulative probability of outcome over 5 years ranged from 5.3% in the middle of the first group to 9.3% in the middle of the 7th group, and 19.0% in the middle of the 10th group (see Appendix 4, Table 117).

5-aminosalicylates

Introduction

5-aminosalicylate are the mainstay of treatment for mild to moderate UC and are often combined with biologics in the treatment of severe disease.12,313,314 Although effective, they can cause side effects such as interstitial nephritis,297,315–319 hepatitis and blood dyscrasias.297,320–324 These side effects are common early during the treatment325–327 but they also occur later. For instance, the median time to development of interstitial nephritis ranged between 2.3 and 3 years in previous studies.318,319 Given the rarity of interstitial nephritis (1.2–1.7 per 1000 person-years315,316) and the lack of data on incidence of cytopenia and hepatotoxicity there is substantial variation in recommendations on monitoring during established 5-ASA treatment. For instance, the summary of product characteristics recommends 3-monthly FBC, LFT and UEC during established treatment,327 whereas the European consensus is to monitor with FBC and UEC at 3- to 6-monthly intervals,314 while the American Gastroenterology Association and the British Society of Gastroenterology recommend monitoring UEC periodically or annually respectively during established treatment.12,313

It would be beneficial to be able to predict clinically significant toxicity during established 5-ASA treatment to allow risk-stratified monitoring for cytopenia, hepatotoxicity and nephrotoxicity. This is particularly important because 5-ASA drugs only uncommonly cause pre-specified hepatic, renal or haematological side effects. This study developed and validated a prognostic model for clinically significant 5-ASA toxicity during long-term treatment.

Methods

Methodology in this study was the same as that used for methotrexate model development and validation except that data from the CPRD Aurum and Gold were used for model development and validation, respectively,254,255 and only participants aged ≥ 18 years with a new diagnosis of IBD newly prescribed 5-ASA by their GP for at least 6 months were eligible.

Predictors: Age, sex, BMI, alcohol intake, and diabetes were included as they associate with DILI.261,262 CKD was included as it reduces 5-ASA clearance.282 Statins, carbamazepine, valproate and paracetamol were included as their use is associated with 5-ASA toxicity relevant to this study.282 Methotrexate, leflunomide and thiopurines were included as they can cause cytopenia, elevated liver enzymes and AKI. Either cytopenia (neutrophil count < 2 × 10⁹/l, or total leucocyte count < 4 × 10⁹/l, or platelet count < 150 10⁹/l) or elevated transaminase (ALT and/or AST > 35 IU/l) during the first 6 months of primary care prescription were included as they predicted cytopenia and/or transaminitis in other studies.41,265

Sample size: The incidence of toxicity from 5-ASA ranged between 1.2 and 1.7 per 1000 person-years for interstitial nephritis.315,316 To minimise model overfitting and ensure precise estimation of overall risk, the minimum sample size required for new model development was 2590 participants (39 events) based on a maximum of 20 parameters, Cox–Snell R² value of 0.12, estimated event rate of 0.005/person-year, a 5-year time horizon, and a mean follow-up period of 3 years using the formulae of Riley et al.266 The sample size for external model validation was larger than the typically recommended minimum sample size of 200 events.312

Results

Data for 14,109 and 7523 participants that contributed 41,146 and 21,070 person-years follow-up were included in the development and validation cohorts, respectively (see Appendix 5, Figures 50 and 51; Table 28).

Model development: In the derivation data set, 401 outcomes occurred in 2.8% patients (n = 14,109) during follow-up at a rate (95% CI) of 9.75 (8.84 to 10.75) per 1000 person-years. Cytopenia, increase in serum creatinine and elevated liver enzymes caused 5-ASA discontinuation in 185, 172 and 64 participants, respectively (see Appendix 5, Figure 52). Twelve candidate predictors (17 predictor parameters) were included in the model.

Before shrinkage, the calibration slope in the derivation data was 1.00 (95% CI 0.82 to 1.18). From the bootstrap, a uniform shrinkage factor of 0.83 was obtained and used to shrink predictor coefficients in the final model for optimism and after re-estimation, the final model’s cumulative baseline survival function (S₀) was 0.971 at 5 years of follow-up (Table 29).

Autoimmune rheumatic diseases included RA, SLE, AS and ReA.

Model performance in the derivation cohort: Calibration plot of the final (i.e. after shrinkage) model at 5 years showed that the average model predictions matched the average observed outcome probabilities across 10 groups of patients, with CIs overlapping the 45° line (perfect prediction line) [Figure 8; data from a single imputed data set was used; S₀(t = 5) = 0.971]. The smoothed calibration curve at 5 years showed alignment of observed risk to the predicted risk with wide CIs at > 0.1 risk probabilities (see Figure 8, see Appendix 4, Figure 53).

FIGURE 8 Calibration of a prognostic model for 5-ASA discontinuation with abnormal monitoring blood-test results at 5 years in the development cohort. (Figure not reproduced here; see the original article.)

The Royston D statistic (95% CI) was 0.75 (0.59 to 0.91) corresponding to HR 2.12 (1.80 to 2.48) when comparing the risk groups above and below the median of linear predictor. The optimism adjusted Royston D statistic was 0.60 corresponding to HR 1.82, obtained by exponentiating the D statistic (Table 30).

Model performance in the validation cohort: There were 243 outcomes in 3.2% (n = 7523) patients at a rate (95% CI) of 11.53 (10.17 to 13.08) per 1000 person-years in the validation cohort. The calibration slope (95% CI) across the 5-year follow-up period was 0.90 (0.61 to 1.19). The calibration plot showed reasonable correspondence between observed and predicted risk at 5 years across the tenths of risk with CIs crossing the perfect prediction line (see Appendix 5, Figure 54). When individual risk was plotted, the smoothed calibration curve showed instability in the model predictions with a tendency to underpredict low risk by a small amount [Figure 9; data from a single imputed data set were used; So(t= 5) = 0.971].

FIGURE 9 Calibration of a prognostic model for 5-ASA discontinuation with abnormal monitoring blood-test results at 5 years in the validation cohort. (Figure not reproduced here; see the original article.)

Model performance was also tested at years 1, 2, 3 and 4 and showed a similar pattern (see Appendix 5, Figures 55–58). Model discrimination in the derivation and validation data was broadly similar (see Table 30). The Royston D statistic in the validation data was 0.57 (0.37 to 0.77), corresponding to HR (95% CI) 1.77 (1.45 to 2.16). The R² score was 0.07 (95% CI 0.03 to 0.12).

Worked examples: Ten anonymised patient profiles, one from the middle of each decile of predicted risk, were selected from the development cohort (see Appendix 4, Table 118). The cumulative probability of outcome over 5 years ranged from 3.1% in the middle of the first group to 4.6% in the middle of the 7th group, and 8.6% in the middle of the 10th group.

Mycophenolate

Background

Mycophenolate is used in the management of anti-neutrophil cytoplasmic antibody (ANCA) vasculitis, SLE, diverse skin conditions including atopic dermatitis, PsO and autoimmune blistering disorders, and to prevent transplant rejection.328–331 Its efficacy and safety have been evaluated in several clinical trials in SLE, ANCA vasculitis, RA, uveitis and vitiligo.200,328,329,332–334 Several of these studies reported high cumulative incidence of cytopenia (5–23%) and elevated liver enzymes (7%).329,332,334 Consequently, indefinite monitoring with 1- to 3-monthly blood tests is recommended after an initial period of closer monitoring.13,335

Thus, as discussed in our previous publication,229 the objectives of this study were to examine the incidence of mycophenolate discontinuation with abnormal and severely abnormal blood-test results in inflammatory conditions, ascertain the pattern of abnormal blood-test results leading to mycophenolate discontinuation, and explore risk factors of stopping mycophenolate with abnormal monitoring blood-test results.

Methods

Data source: Data from the CPRD Aurum were used.

Approvals: Independent Scientific Advisory Committee of CPRD Reference: 20_000236.

Study design: Retrospective cohort study.

Study duration: 1 January 2007–31 December 2019.

Inclusion criteria:

1. diagnosed with RA, SLE, PsO, PsA, ReA, AS, SLE, or IBD at age ≥ 18 years during the study period
2. ≥ 1 GP prescription of mycophenolate after the first record of above conditions in CPRD Aurum, and
3. continuous registration for ≥ 1 year in a GP practice contributing data to CPRD Aurum before the first record of any of the above conditions or prescription of mycophenolate. The above two criteria prevent prevalent cases on long-term treatment that have recently changed GP surgeries from entering the cohort as incident cases and new mycophenolate users.

Exclusion criteria: As for methotrexate study.

Cohort entry: First primary care prescription of mycophenolate.

In the UK, immune-suppressing drugs are initiated in hospital outpatient clinics and dose escalation with monitoring overseen by specialists. However, once a stable well-tolerated dose is reached with acceptable monitoring blood-test results (typically after 3 months of the first prescription), the responsibility for prescribing and monitoring, including with periodic blood tests, is often handed over to the patients’ GP under shared-care agreements. Any decisions to change the dose, interrupt or stop treatment are guided by the hospital specialists.

Cohort exit: The earliest of date of outcome, death, transfer out of the GP practice, last data collection from the GP practice, 5-year follow-up, or 31 December 2019.

Outcomes: Drug discontinuation associated with abnormal, and severely abnormal blood-test results, defined as a prescription gap of ≥ 90 days, with an abnormal or severely abnormal blood-test result or Snomed code indicating such a result within ± 60 days of the last prescription date.13,260,336

Any abnormal blood test defined as cytopenia (WBC < 3.5 × 10⁹/L, neutrophil < 1.6 × 10⁹/l or platelet < 140 × 10⁹/l) or elevated liver enzymes (ALT or AST > 100 IU/l) or increase in serum creatinine by > 26 μmol/l.

Severely abnormal blood test defined as cytopenia (WBC < 2.5 × 10⁹/l, neutrophil < 1.0 × 10⁹/l, or platelet < 50 × 10⁹/l), elevated liver enzymes (ALT or AST > 200 IU/l) or doubling of serum creatinine.

Any drug discontinuation, defined as a prescription gap of ≥ 90 days between the last prescription date and the earliest of date of death, transfer out of the GP practice, last data collection from GP practice or 31 December 2019.

Protocol amendment: An initial feasibility count of the number of patients eligible in the CPRD Aurum indicated that it would not be possible to undertake prognostic modelling and so we performed a descriptive study to explore the incidence of outcomes of interest and their risk factors.

Covariates: Age at first prescription was defined using date of birth and date of first primary care mycophenolate prescription. Sex, smoking, BMI (kg/m²) (classified according to WHO categorisation), alcohol intake status [non-user, ex-user, low (1–14 units/week), medium (15–21 units/week) and hazardous (> 21 units/week) user], inflammatory condition, and CKD were defined using the latest CPRD record prior to cohort entry. CKD stage-3 or higher was defined using Snomed codes or latest eGFR < 60 ml/min before cohort entry. Concurrent immune-suppressive prescriptions were defined using GP prescriptions in first 6 months of follow-up, provided such prescriptions were followed by a mycophenolate prescription.

Outcome validation. All mycophenolate discontinuations with a blood-test abnormality were selected. Data for all consultations within ± 60 days of the abnormal blood-test result were extracted. A.A. (Consultant Rheumatologist trained in General Medicine and Rheumatology) screened all codes to draw up a list of other diagnoses that could potentially cause abnormal blood-test results. All clinical experts in the study team (one rheumatologist, one nephrologist, one hepatologist, one gastroenterologist, one haematologist, one dermatologist and one academic GP) reviewed these codes. A code was removed from the list if all experts agreed that it will not cause blood, liver, or kidney injury. The proportion of mycophenolate discontinuations with abnormal blood-test results potentially explained by an alternate illness was calculated.

Statistical analyses: As described in our previous publication,229 survival analysis was undertaken to calculate the incidence (95% CIs) of outcomes per 1000 person-years for the entire follow-up period, and then separately for the first 12 months and the subsequent period. Cumulative hazards were plotted using Nelson–Aalen graphs. The incidence of mycophenolate discontinuation with abnormal blood-test results was calculated separately for SLE and other inflammatory conditions separately. Cox proportional HRs and 95% CI were calculated to determine the factors associated with mycophenolate discontinuation with any blood-test result abnormality. We used fractional polynomials to model potential non-linear relations between primary outcome and continuous covariates. Missing data for BMI and alcohol were handled by multiple imputation using chained equations. Ten imputations were carried out and the imputation model included all covariates, Nelson–Aalen cumulative hazard function and mycophenolate discontinuation with blood-test result abnormality as the outcome variable. Results from the imputed data sets were combined using Rubin’s rule. Data management and analysis were performed in Stata v16.

Results

Data for 1969 participants were ascertained; (see Appendix 5, Figure 59) of whom, 992 participants contributing 1885 person-years of follow-up were included (Table 31).

There were 389, 118, 20 mycophenolate discontinuations due to any reason, with any abnormal, and with any severely abnormal monitoring blood-test results (Table 32).

The incidence of mycophenolate discontinuation with any blood-test abnormality, and with any severe blood-test abnormality, was higher in the first 12 months of prescriptions than in subsequent years (see Table 32).

Also, 11.5% (1.7%) patients prescribed mycophenolate discontinued treatment with any (severely) abnormal blood-test results in the first year of prescription. After this period, mean 2.6% (0.7%) patients discontinued treatment with any (severely) abnormal blood-test results per year. The incidence of drug discontinuation for any blood-test abnormality or severely abnormal blood-test results were comparable in those with and without SLE (see Table 32). Increased serum creatinine and cytopenia were the commonest reasons for mycophenolate discontinuation (Table 33).

On multivariate analysis, CKD ≥ stage 3 significantly increased the risk of stopping mycophenolate associated with abnormal blood-test results (see Table 32).

Anti tumour necrosis factor-alpha

Introduction

Anti-TNF-alpha drugs are used in the treatment of RA, PsA, PsO and AS resistant to first-line systemic treatments.337 Although effective, they can cause cytopenia and elevated liver enzymes. As discussed in the published manuscript,4 concern about these side effects led the British Association of Dermatology (BAD) to recommend monitoring with FBC, LFT and UEC, at least 6-monthly during long-term established anti-TNF-alpha treatment.337 On the contrary, the American Academy of Dermatology do not recommend a fixed monitoring schedule but leave the decision on whether to monitor with blood tests, and the frequency of testing to the patients’ dermatologist for all anti-TNF-alpha drugs except for infliximab for which 6-monthly to annual blood-test monitoring was recommended by them.338 This is most likely to be due to higher reported risk of thrombocytopenia, neutropenia and hepatotoxicity with infliximab.261,339

Previous studies reported low rates of hepatotoxicity and cytopenia with anti-TNF-alpha drugs, and that the risk of hepatotoxicity was increased in those concomitantly treated with methotrexate.77,340

Protocol amendment: A feasibility assessment of the available data indicated that a prognostic model would not be able to be developed or validated due to the availability of a single data set and low event rates, and this step was not performed for anti-TNF alpha drugs. To support evidence-based practice and better utilisation of healthcare resources, we conducted a retrospective cohort study that evaluated the incidence of anti-TNF-alpha discontinuation due to either cytopenia, transaminitis, or acute kidney, ascertained the risk factors for this outcome.

Methods

Data source: Data from the British Association of Dermatology Biologics Register (BADBIR) were used in this study.341 The BAD established a web-based pharmacovigilance register BADBIR to assess the long-term safety of biologics prescribed for patients with moderate to severe PsO. Following baseline data acquisition, clinicians in hospitals recorded changes in therapy, disease activity, changes in treatment, including reasons for stopping treatment, and AEs. Data on death and malignancy are obtained from linkages with the National Health Service Information Centre system. We planned to use data from the BSR Biologics Registers for RA and AS. However, these data sets did not always record reasons for stopping biologics. There were only a handful of these events and many drug discontinuations where no cause was recorded. Consequently, data from these registers could not be used in the current study.

Approvals: This study protocol was approved by the BADBIR steering committee.

Study design: Retrospective cohort study.

Study period: 1 October 2007–31 July 2021.

Study population: Adults (age ≥ 18 years) with PsO +/− PsA prescribed their first anti-TNF-alpha drugs were eligible. There were no exclusion criteria in terms of comorbidities or concurrent drug prescriptions. In the UK and Republic of Ireland, anti-TNF drugs are prescribed from hospital outpatient clinics. The hospital team also organises monitoring blood tests and acts on any abnormalities.

Follow-up: Patients were followed up from their first anti-TNF-alpha prescription until the earliest of outcome, death, treatment discontinuation, 31/7/2021 or 5 years.

Outcome: Anti-TNF-alpha discontinuation due to monitoring blood-test abnormality was the outcome of interest. This was ascertained using the reason for stopping treatment specified in the BADBIR database. Where the reason for stopping treatment was recorded only as an ‘adverse-event’, ‘adverse-event and inefficacy’, ‘other’, or ‘death’, the descriptions of AE in free text and Medical Dictionary for Regulatory Activities (MedDRA) coding were manually reviewed to ascertain outcomes related to side effects of interest.

Values of blood tests from within 6 months before stopping anti-TNF-alpha treatment were used to ascertain whether the treatment was stopped for leucopenia, thrombocytopenia, or elevated liver enzymes. The eGFR was calculated using the 2021 CKD-EPI equation that does not consider race.342

Risk factors: Age at registration (years), sex, BMI (kg/m²), alcohol intake [not-currently drinking alcohol, and self-reporting low (1–14 units/week), medium (15–21 units/week) and high alcohol intake (≥ 21 units/week)], diabetes, high cholesterol and a recorded clinical diagnosis of liver disease before date of first anti-TNF-alpha prescription were the risk factors. The record of height and weight at registration were used to calculate the BMI. Age, sex, BMI, alcohol intake and diabetes were included as they are associated with DILI.261

Statistical analysis: As described in our previous publication,4 the incidence rate (95% CI) of outcome per 1000 person-years was calculated for the entire follow-up period, for the first 6 months, for the first 12 months, and for the follow-up period after the first 12 months. Multivariable Cox regression was used to examine the association between covariates at baseline and outcome of interest. Missing data for BMI and alcohol were handled by multiple imputation using chained equations. Ten imputations were carried out, and the imputation model included all covariates and the outcome variable. Results from the imputed data sets were combined using Rubin’s rule. Data management and analysis were performed in STATA v.16.

Results

Data for 8819 participants with PsO (42.1% female) prescribed their first anti-TNF-alpha agent that contributed 25,058 person-years of follow-up were included in this study. Their mean (SD) age and BMI were 44.76 (13.20) years and 31.15 (9.44) kg/m², respectively. Of the included participants 6744 (76.47%), 1718 (19.50%), 281 (3.19%) and 76 (0.86%) were prescribed adalimumab, etanercept, infliximab and certolizumab or their biosimilars, respectively. Diabetes, high cholesterol, liver disease and renal disease were present in 738 (8.37%), 742 (8.41%), 419 (4.75%) and 131 (1.49%) participants, respectively.

There were 125 anti-TNF-alpha treatment discontinuations for abnormal monitoring blood test results at an incidence rate (95% CI) of 5.85 (4.91 to 6.97)/1000 person-years. The incidence rate (95% CI) of discontinuation of etanercept, infliximab, certolizumab and adalimumab or their biosimilars for abnormal monitoring blood tests were 8.04 (5.55 to 11.65), 18.55 (10.53 to 32.66), 9.79 (1.38 to 69.53) and 4.90 (3.96 to 6.07)/1000 person-years, respectively.

Forty-one, 64 and 61 discontinuations for abnormal blood-test results occurred in the first 6 months, in the first year, and after the first year of first prescription, over 4111, 7428, and 17,751 person-years, at an incidence rate (95% CI) of 9.97 (7.34 to 13.54), 8.62 (6.74 to 11.01) and 3.44 (2.67 to 4.42)/1000 person-years, respectively. The incidence rate (95% CI) for discontinuation of etanercept, infliximab, certolizumab and adalimumab or their biosimilars for abnormal blood-test results after the first year of treatment were 4.63 (2.63 to 8.16), 12.30 (5.87 to 25.81), 23.60 (3.32 to 167.55) and 2.82 (2.07 to 3.83) per 1000 person-years, respectively.

Blood-test results from within 6 months prior to the stop date were available for 75 patients that stopped treatment. Among them, 51 had ALT > 40 IU/l, 29 had ALT > 80 IU/l and 7 had ALT > 200 IU/l; 5 had total WBC count < 4.0 × 10³/ml and none had WBC count < 3.0 × 10³/ml; 8 had platelet count < 150 × 10³/ml, 4 had platelet count < 100 × 10³/ml, and none had platelet count < 50 × 10³/ml; 2 had eGFR < 60 ml/minute, with none having eGFR < 30 ml/minute.

Increasing age, diabetes and liver disease associated with anti-TNF-alpha discontinuation with abnormal blood tests in the model adjusted for age, sex, BMI, alcohol intake, pre-existing liver disease, diabetes and hypercholesterolaemia with adjusted HR (95%CI) of 1.02 (1.01 to 1.04), 1.68 (1.00 to 2.81), and 2.27 (1.26 to 4.07), respectively (Table 34).

Infliximab was statistically significantly associated with stopping treatment due to abnormal monitoring blood tests than etanercept with unadjusted and adjusted HR (95% CI) of 2.38 (1.21 to 4.67) and 2.21 (1.12 to 4.34), respectively. Adalimumab had significantly lower risk of stopping treatment with abnormal monitoring blood tests when compared to etanercept on univariate analysis [unadjusted HR (95% CI) 0.63 (0.41 to 0.97)] which became non-significant on adjustment for covariates [adjusted HR (95% CI) 0.66 (0.43 to 1.03)].

Discussion

We have developed and externally validated a prognostic model for predicting the likelihood of treatment discontinuation due to abnormal blood-test results that utilises disease and demographic factors ascertained during routine clinic visits. Our prognostic model performed well in predicting outcomes by 5 years and did so in clinically relevant subgroups. Cytopenia and/or elevated liver enzymes in the 6 months immediately before start of follow-up, CKD stage 3 and diabetes were strong independent predictors for methotrexate toxicity. The latter two associations could be due to reduced methotrexate clearance in CKD, and diabetes being associated with DILI and non-alcoholic fatty-liver disease (NAFLD).343 As reported previously, increasing age and prior blood-test abnormalities were strong predictors of thiopurine toxicity in the current study.38,204 The reduced thiopurine clearance in CKD and occurrence of renal involvement in SLE could also explain their being strong predictors of thiopurine toxicity in this study.343 We did not set out to establish an etiological model but a prognostic model in which known predictors are included in the model to assess an individual’s risk of developing an outcome of interest. The beta-coefficients and 95% CIs should not be interpreted to assess causality.

Side effects can occur several years after starting on thiopurines, for example hepatotoxicity due to late metabolic shunting to 6-methylmercaptopurine ribonucleotide (6-MMPR) production occurred in 1.7% patients prescribed thiopurines after median 644 days of treatment.344 Another cohort study with long follow-up showed that adverse effects occur later in therapy, but are far less frequent in this period.281 It is in this later period that we propose our risk score has its utility.

As reviewed in our previous publication,1 genomic studies have established TPMT activity prior to thiopurine prescription as a useful and cost-effective predictor of myelotoxicity, and though it has a clear place, TPMT testing cannot predict all myelosuppression.12,42,207,345 Unfortunately, the results of TPMT testing are not available in the CPRD and we were unable to include them in the prognostic model. For therapeutic drug monitoring (TDM), there is uncertainty on how well it predicts future toxicity with 6-TGN and 6-MMPR. At treatment start, TDM did not perform well in predicting hepatotoxicity and myelotoxicity in two studies but was a strong predictor of myelotoxicity in the TPOIC trial cohort.192,346,347 Another study reported more cases with leucopenia in the presence of high 6-TGN levels compared to low 6-TGN levels, although the differences were not statistically significant.348 Reactive TDM during thiopurine treatment aids accurate ascertainment of toxicity, facilitates dose reduction for dose-dependent side effects, and allows personalised dosing, thereby improving persistence on treatment and disease outcomes.349,350 TDM for thiopurines is not commonly performed in the UK and their results are not available in the CPRD. Further research is required to explore if the inclusion of TPMT and TDM biomarkers at baseline would improve model performance. Future validation studies in Asian and Hispanic populations should also consider including polymorphisms in the NUDT 15 in the model as these polymorphisms are common in the Asian and Hispanic populations with a prevalence of 27%.351

Among patients prescribed thiopurines, due vigilance needs to be maintained for hepatic nodular regenerative hyperplasia which is indicated by thrombocytopenia and/or elevated liver enzymes and non-cirrhotic portal hypertension which may also manifest with anaemia, splenomegaly, variceal bleed and/or ascites.352,353 The median time from starting azathioprine to hepatic nodular regenerative hyperplasia was 48 months in a large case-series.352

Strengths of this study included the use of a large, real-world and nationally representative data set for model development and a similar independent data set for external validation. The study population included patients with a range of immune-mediated inflammatory diseasess, and the results have broad generalisability. The prognostic factors were selected by an expert multidisciplinary team based on their scientific knowledge and clinical experience. Outcome definition required the abnormal blood-test result to be associated with methotrexate discontinuation, thus allowing the model to predict clinically relevant outcomes. Finally, information on factors included in this prognostic model are easily ascertainable during routine care, making the model easy to use and allowing it to be incorporated into electronic health records.

Some limitations of this study need further consideration. First, we did not have access to the date when the patient was first prescribed treatment in the hospital clinic. However, we had accurate information on the date of first primary care prescription, and, therefore, our model is fit to risk-stratify monitoring from 6 months after first primary care prescription. We also did not have data on the concurrent use of biologics as these are hospital prescribed. However, blood-test abnormalities during long-term treatment with biologics are uncommon, and evidence summarised in the BSR monitoring guidelines shows no evidence to suggest that concurrently prescribed biologics increase toxicity.354 This study assessed the risk of drug discontinuation with blood-test abnormalities, only some of which could be due to a serious illness. In clinical practice, it is rare to get serious side effects during established treatment with immune-suppressing drugs; although we were interested in studying serious illnesses as an outcome of drug side effects, these are not adequately coded in the CPRD. Consequently, we did not assess these as outcomes separately. We did not have data on disease activity as these are not recorded in the CPRD. Disease activity is not expected to cause abnormal monitoring blood-test results directly but may result in the use of higher doses of therapy, which was included in the model. It is possible the abnormal blood test could be due to a different illness and not be related to treatment. In our previous validation study, only 5.4% of abnormal blood-test results could be explained by an alternative illness.252 Although the external validation data set was distinct from the model development data set, it also originated from UK general practice. We recommend that our model be validated in a data set from another country. Our imputation strategy is congenial with all predictors being available at deployment, and further research would need to consider how to handle missing predictor values at deployment.

We did not use dose reduction to define outcomes because of missing dose data between consecutive prescriptions. Despite not considering such events, the study had sufficient power for the parameters included. We excluded patients with a prior diagnosis of serious liver, haematological or renal diseases as these conditions often contraindicate drug use and may additionally cause uncertainty with outcome ascertainment. The results of TPMT testing are not available in the CPRD. We do not believe that this is a major limitation as approximately 90% of Whites are not deficient in TPMT, and the vast majority of myelotoxicity due to TPMT deficiency occurs in the first few months of treatment.355,356 Participants included in our study were already prescribed thiopurines for at least 6 months from primary care after an initial period of prescription and monitoring from hospital outpatient clinic and most instances of myelotoxicity from thiopurine deficiency would have occurred before follow-up started and this prognostic model should not be used to replace the need for TPMT testing. Furthermore, the model performance was comparable up to 2010 and in the years after that when TPMT testing became more widespread in the UK. Finally, it should be noted that the model is derived from a primary care data set and may not be applicable to patients with complex comorbidities prescribed and monitored exclusively in secondary care. However, this is unlikely to be a significant limitation because such patients are very uncommon nowadays due to the availability of other treatments.

Previous studies have variably reported NAT-2 acetylator status to be associated with sulfasalazine toxicity.309,357 However, these studies evaluated all side effects and did not separately assess either myelotoxicity, hepatotoxicity, or nephrotoxicity as evaluated in the current study. Although generally perceived to be safe, sulfasalazine use carries a risk of myelotoxicity and hepatotoxicity comparable to that observed with methotrexate in people with RA.358 CKD stage 3, diabetes and concomitant methotrexate or leflunomide therapy were strong independent predictors of sulfasalazine discontinuation with abnormal monitoring blood-test results in this study. These associations may be due to reduced sulfasalazine clearance in CKD and DILI being associated with diabetes.343 Abnormal blood-test results during the first 6 months of therapy were strong independent predictors of discontinuing sulfasalazine with abnormal monitoring blood-test results, like findings for methotrexate and leflunomide.228,253 Elevated liver enzymes and cytopenia before starting treatment have previously been associated with abnormal blood-test results in patients treated with methotrexate and biologics, respectively.41,52,215,222,227,231,232

5-aminosalicylates were stopped for cytopenia, elevated creatinine and elevated liver enzymes, suggesting that monitoring blood tests should include FBC, UEC and LFT. Numerically, more patients discontinued 5-ASA due to cytopenia than due to nephrotoxicity, with far fewer participants stopping treatment due to elevated liver enzymes. In a previous study, there were numerically greater number of blood dyscrasias than AKI reported with relatively few reports of elevated liver enzymes.297 In another study, the mean WBC count reduced significantly with 5-ASA treatment, while the mean levels of creatinine and liver enzymes remained unchanged.101

Chronic kidney disease-3, hazardous and prior-alcohol use were strong independent predictors of 5-ASA discontinuation. These are novel findings. Nevertheless, these are plausible predictors. Coprescription of a thiopurine was a strong risk factor for toxicity consistent with other studies.212,359,360 Abnormal blood-test results during the first 6 months of therapy were strong independent predictors of discontinuing 5-ASA with abnormal monitoring blood-test results, similar to previous reports about other drugs.228,253

The risk-score output from the prognostic model may be used to decide individual monitoring strategies after the first 6 months of primary care prescription, that is patients at low risk of toxicity could be advised to undergo less frequent monitoring blood tests, while those at high risk of toxicity undergo more frequent testing. This requires shared decision-making taking account of the individualised risk scores, patient preferences, health-professional recommendations and updated guidelines. It is beyond our scope to recommend specific risk thresholds at which current clinical practice may be changed.

The frequency of monitoring blood tests is decided according to recommendations of specialist societies such as the BSR, British Society for Gastroenterology and BAD, and these thresholds are best decided by their guideline writing groups that are external to the research team and include broader clinical and patient representation.

There is other widely used prognostic models where the risk of future events is calculated as a continuous score and the threshold at which clinical practice is changed is recommended in specialist society guidelines. For instance, fracture risk algorithm calculates the risk of future osteoporotic fractures as a continuous score, and the threshold at which either lifestyle interventions, bone density measurement, or drug treats is recommended is decided according to guidance from the National Osteoporosis Guideline Group.

Lengthening the time between monitoring blood tests would save patient and health-professional time, minimise discomfort from unnecessary venepunctures, and conserve healthcare resources because patients with well-controlled diseases undergo quarterly monitoring blood tests at their GP’s office or at hospital or community phlebotomy services and are only followed up in specialist clinics annually. Patients taking drugs are adherent to monitoring recommendations,237 and once implemented the results of this study will reduce the volume of monitoring blood tests. It is important that the results of this study are not used to risk-stratify monitoring in patients newly started on treatment because we did not use data from this period in our study. In the UK, it typically takes 6 months to stabilise patients’ dose before prescription and monitoring is handed over to the GP.

For mycophenolate, this is the largest study to evaluate the incidence and pattern of mycophenolate discontinuation with abnormal monitoring blood tests. It reports that mycophenolate is frequently discontinued associated with abnormal blood-test results in the first 12 months of shared-care prescription. However, discontinuations associated with this reason became approximately fivefold less frequent thereafter. Mycophenolate discontinuation associated with severe blood-test abnormalities occurred uncommonly in the first 12 months and became approximately threefold less frequent after this. Similar findings were observed in SLE and other inflammatory conditions.

Chronic kidney disease stage ≥ 3 was associated with mycophenolate discontinuation associated with abnormal blood-test results. This is expected as mycophenolate is renally excreted. Age was not associated with mycophenolate discontinuation with abnormal blood-test result, consistent with previous studies.361,362 Elevated liver enzymes were a significantly less common cause of mycophenolate discontinuation than cytopenia in our study as reported in most previous trials.328,329,334,363 However, a single previous trial reported a higher incidence of elevated liver enzymes than of cytopenia with mycophenolate, but at 6 months follow-up.332

Using a similar study design, we previously reported on incidence of methotrexate and leflunomide discontinuation with abnormal and severely abnormal blood-test results.252 The incidence of mycophenolate discontinuation was higher than leflunomide or methotrexate for abnormal (and severely abnormal) blood-test results across the entire study period with crude incidence rates of 76.20 (12.65), 58.22 (6.16) and 27.78 (3.66)/1000 person-years for mycophenolate, leflunomide and methotrexate, respectively. These results suggest that methotrexate may be preferred over mycophenolate where there is no evidence for a superior efficacy of the latter.

Strengths of this study included inclusion of a broad range of inflammatory conditions, and the use of real-world data, thus increasing generalisability. Outcomes were stratified according to their severity and time course, adding detail to the results. However, this study has several limitations. First, patients included in this study had been commenced on mycophenolate in secondary care, were stabilised on treatment, and prescribing responsibilities had been handed to primary care. Consequently, patients with severe, unstable, or uncommon diseases that are managed in specialised services or those at high risk of side effects that may be prescribed treatment from secondary care were excluded from this study. Similarly, some services where shared-care prescription and monitoring do not extend to mycophenolate were excluded. The patient population in this study did not include those with small-vessel vasculitis or myositis, further limiting generalisability. Next, due to missing data on the dose of mycophenolate provided by CPRD, missing information on number of tablets prescribed and/or length of prescription, and large dose range (0.5–4.0 g/day), we were unable to calculate daily dose of mycophenolate. Therefore, we did not evaluate the incidence of dose reduction with abnormal monitoring blood-test results as an outcome, because incomplete information on dosing meant it was difficult to establish when this occurred. Additionally, increase in serum creatinine by > 26 μmol/l is the minimum change required to consider presence of AKI according to guidelines, and was used to ascertain drug discontinuation with renal function decline.336 However, the guidelines require that this increase occurs within 48 hours. We were unable to apply this part of the definition due to monitoring blood tests being performed 3 monthly. This may have overestimated the incidence of mycophenolate discontinuation with elevated serum creatinine. Our results therefore represent a worst-case scenario for kidney function. Some of the abnormal blood-test results could be due to concurrent prescription of other drugs. This can potentially elevate the outcome rate. However, this is unlikely to play a large part as our outcome definition required a prescription gap of ≥ 90 days, and it can reasonably be expected that in this period the drug responsible for the blood-test abnormality will easily be ascertained. Additionally, some treatment discontinuations in SLE may be due to treatment escalation to address increased disease activity, for example, cytopenia. However, a stratified analysis of cases with and without SLE reported similar event rates, implying that the event rates reported in this study are a reasonably accurate reflection of pre-specified haematological, hepatic and renal side effects due to mycophenolate. We are unable to attribute causality with certainty to abnormal blood tests for the decisions to cease MMF. However, this in no way interferes with our conclusion that abnormal blood tests are not causing frequent cessations of MMF after 1 year of use in general practice. Reassuringly, in our validation exercise only 11% outcomes were potentially explained by alternate medical condition.252 Gastrointestinal intolerance, a common reason for mycophenolate discontinuation, was not an outcome in this study as the study focused on asymptomatic renal, hepatic or haematological side effects. MMF often remains a hospital prescription. It is possible, therefore, that the patients in the study are those considered lower risk by the secondary care team. Similarly, it is possible that these patients have a much longer period of secondary care prescription before being transferred to primary care under a shared-care agreement. This is another limitation to our study. Information on disease severity and disease activity are not included in the CPRD and could not be evaluated as potential risk factors in this study. Finally, despite relatively large sample size, there were relatively few events, and several 95% CIs are wide, raising the risk of imprecision.

Mycophenolate appears to be commonly discontinued for abnormal blood-test results within the first year of prescription; however, this becomes less frequent after the first year. Discontinuations for severely abnormal blood-test results were even less common. These data may be used when counselling patients of their risks and benefits when considering mycophenolate.

For anti-TNF-alpha, among a cohort of moderate to severe PsO patients starting anti-TNF-alpha drugs for the first time we found that it was uncommon for the treatment to be stopped due to abnormal blood-test results, and that this became even less common after the first year of treatment. This is consistent with reports that the median interval between starting anti-TNF-alpha drugs and DILI and neutropenia ranged between 14 and 18 weeks and 12 and 17 weeks, respectively, although late-onset liver enzyme elevations and neutropenia were also reported in these studies.215,218,364–366 Increasing age and pre-existing liver disease were independent risk factors for stopping anti-TNF-alpha due to abnormal monitoring blood-test results consistent with previous reports.261 Consistent with previous reports of higher risk of DILI with infliximab, we found significantly higher risk of stopping infliximab treatment with abnormal monitoring blood-test results.365,367

In this study, among those who stopped treatment for abnormal blood-test results, most did so due to abnormal LFT results, and, of these, the vast majority had mildly elevated liver enzymes. Only a few participants stopped their treatment due to low WBC count and low platelet count. No participant had either leucopenia or thrombocytopenia that was severe enough to increase the risk of infection and bleeding respectively in this study. Previous studies have reported occasional patients with severe neutropenia and infection,215 although severe thrombocytopenia increasing the risk of bleeding was not reported previously to the best of our knowledge. Two patients stopped treatment with abnormal renal function. This is not unexpected as anti-TNF-alpha drugs are occasionally associated with granulomatous nephrotoxicity.368,369

Strengths of this study included the use of a nationwide cohort of PsO patients from the UK and Republic of Ireland, increasing the generalisability of the study findings. Furthermore, the reasons for stopping treatment were systematically documented in BADBIR. Free-text reasons for stopping treatment and MedDRA coding were individually reviewed when there was diagnostic uncertainty. This reduced the chances of missing outcome data. Another strength is the large sample size – to the best of our knowledge this is the largest study of its kind to date. Nevertheless, there are several limitations that need to be considered when interpreting the study results. Some patients who stopped treatment due to inefficacy and mildly abnormal monitoring blood tests were considered to have stopped treatment for abnormal blood-test results. This may have increased the event rate. Lack of information on blood-test results is another limitation. Information on disease severity and concurrent medications was not included in the data analysis. There were very few participants treated with infliximab and certolizumab which increased the imprecision around estimates of incidence of side effects with these drugs.

In conclusion, this study showed that anti-TNF-alpha drugs are uncommonly stopped due to abnormal monitoring blood-test results, especially after 1 year of treatment.