The Association Between Arthralgia and Vedolizumab Using Natural Language Processing

Abstract

Background

The gut-selective nature of vedolizumab has raised questions regarding increased joint pain or arthralgia with its use in inflammatory bowel disease (IBD) patients. As arthralgias are seldom coded and thus difficult to study, few studies have examined the comparative risk of arthralgia between vedolizumab and tumor necrosis factor inhibitor (TNFi). Our objectives were to evaluate the application of natural language processing (NLP) to identify arthralgia in the clinical notes and to compare the risk of arthralgia between vedolizumab and TNFi in IBD.

Methods

We performed a retrospective study using a validated electronic medical record (EMR)–based IBD cohort from 2 large tertiary care centers. The index date was the first date of vedolizumab or TNFi prescription. Baseline covariates were assessed 1 year before the index date; patients were followed 1 year after the index date. The primary outcome was arthralgia, defined using NLP. Using inverse probability of treatment weight to balance the cohorts, we then constructed Cox regression models to calculate the hazard ratio (HR) for arthralgia in the vedolizumab and TNFi groups.

Results

We studied 367 IBD patients on vedolizumab and 1218 IBD patients on TNFi. Patients on vedolizumab were older (mean age, 41.2 vs 34.9 years) and had more prevalent use of immunomodulators (52.3% vs 31.9%) than TNFi users. Our data did not observe a significantly increased risk of arthralgia in the vedolizumab group compared with TNFi (HR, 1.20; 95% confidence interval, 0.97–1.49).

Conclusions

In this large observational study, we did not find a significantly increased risk of arthralgia associated with vedolizumab use compared with TNFi.

INTRODUCTION

Joint pain, or arthralgia, is a well-recognized manifestation of inflammatory bowel disease (IBD), with prevalence ranging from 16% to 33%.^1–5 As the focus of IBD treatment is on life-threatening bowel manifestations, studies of arthralgia in IBD are limited. However, arthralgia can cause significant morbidity. Existing treatments for IBD include tumor necrosis factor inhibitors (TNFi’s) that systemically target inflammation and are also effective in treating inflammatory arthritis.⁶ Newer treatments for IBD aim to be gut selective in their immunosuppressive activities, with vedolizumab being the first on the market.^7–9 Since widespread use of the drug, there has been concern about whether such gut-selective treatments are be less effective in controlling joint symptoms. In some case reports, vedolizumab was linked to increased arthralgia; however, another study reported an improvement in joint pain.^{10, 11} The objective of this study was to assess for an association between arthralgia and vedolizumab use at a population level.

Studying arthralgia in IBD is challenging because it is not commonly coded by gastroenterologists. However, joint pain is typically mentioned in narrative clinical notes. These types of narrative data can be accessed using natural language processing (NLP). Briefly, NLP is a computer-based technology that, when applied to narrative electronic medical record (EMR) notes, can detect and extract clinical information to generate structured data. For example, NLP can process notes to identify which notes state that a patient is having joint pain. This information can then be converted to a structured variable, joint pain yes/no, to be included in an analysis.

The objective of this study was to test an approach for efficiently studying a potential adverse outcome, arthralgia, and its potential association with a relatively new treatment. Additionally, we compare the accuracy of billing codes for arthralgia against a standardized NLP definition.

METHODS

Data Source

We performed the study using data from the electronic medical records of 2 tertiary care hospitals, Brigham and Women’s Hospital and Massachusetts General Hospital. We identified all subjects who received an infusion for vedolizumab as of February 6, 2016. As a comparison group, we studied subjects who received a new prescription for TNFi in a validated EMR-based IBD cohort consisting of subjects with Crohn’s disease (CD) and ulcerative colitis (UC). The algorithm used to classify subjects with IBD from the EMR had a positive predictive value (PPV) of 97%.¹²

Study Population

A retrospective cohort study was conducted by including new vedolizumab and new TNFi users. The first date patients received vedolizumab or TNFi was defined as the index date. The 1-year period before the index date was the covariate assessment period. Each patient was followed from the index date until the study outcome, arthralgia, occurred or 1 year after the start of follow-up.

Exposures

In the primary analysis, we adopted an intent-to-treat scenario and categorized the exposure for each patient into vedolizumab or TNFi based on their drugs prescribed on the index date. Patients were assumed to receive their drugs throughout the 1-year follow-up period.

Primary Outcome of Arthralgia

Arthralgias were defined using the ICD9 code “pain in joint” (719.4-719.4x). The NLP definition of arthralgia was the concept unique identifier (CUI) for the main concept “arthralgia ” (C0003862), as defined by the Unified Medical Language System (UMLS; release 2016AA),¹³ which includes commonly used terms that describe arthralgia such as “joint pain.” From the primary Metathesaurus Relationship File (MRREL), we identified additional concepts related to arthralgia. These related concepts, defined as “child” (CHD) concepts of arthralgia, included 36 concepts with descriptions of more than 20 sites or regions of joint paint such as “pain in ankle” and “hand joints pain” and other descriptions of joint pain with different characteristics such as “diffuse arthralgia” and “arthralgia started suddenly”. A dictionary was generated using UMLS for the total 37 concepts mentioned above (Supplementary Table 1) containing different expressions of each concept. For example, “hip arthralgia” and “painful hip” are different expressions of the same concept, “hip pain” (with a CUI of C0019559). Notes were processed using a published NLP software NILE to detect and record any positive mention of the 37 concepts.¹⁴ Negated concepts, for example, “no joint pain,” were not included in the analysis. NLP data for arthralgia were generated for every patient note.

We determined the performance characteristics of the ICD9 vs NLP definition for arthralgia compared with gold standard labels assigned through medical record review of 100 patients records randomly selected from the IBD cohort.¹² A broad definition of joint pain attributed to any cause, inflammatory or noninflammatory, was used as a positive case of arthralgia.

Covariates

We used the following covariates for confounder adjustments: age, sex, Deyo comorbidity index,¹⁵ TNFi, immunomodulatory agents and aminosalicylates, steroid use, baseline arthralgia (defined as ≥1 NLP arthralgia mentions), and follow-up time for IBD.

Statistical Methods

To compare the accuracy of ICD-defined arthralgia compared with NLP, we performed a medical record review of a random 100 subjects for evidence of arthralgia. We then constructed 2 × 2 tables comparing ≥1 ICD9 codes for arthralgia, compared with arthralgia defined as chart review (gold standard). The same was performed for ≥1 mentions for the concept of arthralgia extracted using NLP compared with the gold standard. Performance characteristics were reported as sensitivity, specificity, negative predictive value (NPV), and PPV.

We used the chi-square test and the Student t test to examine the differences in baseline characteristics between the vedolizumab and TNFi groups. To better control confounding and preserve statistical power, we calculated the stabilized inverse probabilities of treatment weights (IPTWs) for all groups and weighted them in the baseline table and Cox proportional hazard regression models, using TNFi as the reference. A propensity score was calculated for each patient using a multivariable logistic regression model conditioned on all covariates included in Table 1. A stabilized IPTW was then developed by multiplying the IPTW in the vedolizumab and TNFi groups by the marginal prevalence of the treatment actually received.¹⁶ The mean of the stabilized IPTW was checked to examine outliers and whether the cohort was weighted appropriately.

Table 1:

Baseline Characteristics

	Original Cohort			IPTW Weighted Cohorta
	Vedolizumab (n = 367)	TNF (n = 1218)	P	Vedolizumab (n = 350)	TNF (n = 1225)	P
Mean age (SD), y	41.2 (14.8)	34.9 (16.2)	<0.01	39.0 (13.9)	36.6 (16.8)	<0.01
Male sex	42.5	46.7	0.16	45.9	45.4	0.96
Mean Deyo CCI (SD)	0.3 (0.6)	0.2 (0.6)	0.02	0.3 (0.6)	0.3 (0.6)	0.35
Medications
5-aminosalicylates	25.9	26.1	0.93	28.7	26.2	0.35
Immunomodulators	52.3	31.9	<0.01	41.0	37.1	0.18
Steroid	81.5	40.2	<0.01	51.8	50.0	0.56
IBD type			<0.01
Ulcerative colitis	29.7	29.0
Crohn’s disease	69.7	70.5
Mixed	0.6	0.5
Baseline arthralgia,b %	46.1	28.5	<0.01	35.2	32.9	0.41
Follow-up time, mo	85.3 (69.8)	70.3 (57.4)	<0.01	72.4 (60.6)	73.4 (59.2)	0.77

Immunomodulators: methotrexate, 6-mercaptopurine, balsalazide, mesalamine, sulfasalazine, azathioprine.

^aIPTW conditioned on all variables in Table 1 (mean IPTW [SD], 0.99 [0.52]).

^b≥1 mentions of NLP arthralgia.

In addition to the IPTW-weighted Cox model, we conducted secondary analyses with an additional adjustment for age in the Cox model and a subgroup analysis excluding subjects with prevalent arthralgia during the baseline period. Additionally, we used the NLP mentions of arthralgia, or the burden of arthralgia, as an outcome to approximate the potential severity of joint symptoms. To do so, we counted the number of visits with positive mentions of arthralgia during the baseline period and 1 year after the index date. The number of visits with positive mentions of arthralgia was divided by the total number of visits during each period at the patient level to calculate a “% arthralgia burden” for each subject. A signed-rank test was performed for each subject comparing % arthralgia burden in the 1-year period before and after the index date.

RESULTS

Baseline Characteristics of Included Patients

We studied 367 vedolizumab and 1218 TNF inhibitor new users identified from the IBD cohort. Vedolizumab users were older (41.2 ± 14.8 vs 34.9 ± 16.2 years), had a higher mean Deyo comorbidity index, and had more prevalent use of immunomodulators (52.3% vs 31.9%) and steroids (81.5% vs 40.2%) than the TNFi group (Table 1).

ICD vs NLP Definition of Arthralgia

Among the 100 randomly selected patients for medical record review of arthralgia, the prevalence was 43.3%. Compared with chart-reviewed labels for arthralgia, the ICD9 code for arthralgia had a PPV of 79%, and NLP had a PPV of 90%. The performance characteristics were better for NLP than ICD9, including sensitivity for arthralgia of 52% for ICD9 and 83% using NLP (Table 2). Thus, for the primary analyses to test the association between vedolizumab and arthralgia, we used the number of NLP mentions of arthralgia as the outcome.

Table 2:

Performance Characteristics of Identifying Arthralgia in Patients Using ICD9 Codes Compared With NLP

Characteristics	ICD9 for Arthralgia		NLP for Arthralgia
		95% CI		95% CI
Positive predictive value	0.79	0.59–0.92	0.9	0.76–0.97
Negative predictive value	0.71	0.59–0.81	0.88	0.77–0.95
Sensitivity	0.52	0.36–0.68	0.83	0.69–0.93
Specificity	0.89	0.78–0.96	0.93	0.82–0.98

Abbreviation: CI, confidence interval.

Comparative Risk of Arthralgia Between Vedolizumab and TNFi

Vedolizumab users had a higher prevalence of arthralgia during the baseline period (46.1% vs 28.5%). After weighing by the IPTW, all covariates were well-balanced between 2 groups except for the mean age (Table 1). The crude arthralgia rate during the 1-year follow-up period was 54.2 per 100 person-years in the vedolizumab group and 40.4 per 100 person-years in the TNFi group (P = 0.002) (Table 3).

Table 3:

Crude Arthralgia Rates in Entyvio and TNF Cohorts

	Prevalent Cohort		Incident Cohort
	Vedolizumab (n = 367)	TNFi (n = 1218)	Vedolizumab (n = 198)	TNFi (n = 871)
Incident arthralgia during follow-up	139	385	44	187
Follow-up, person-year	256.4	953.9	167.5	751.1
Arthralgia rate/100 person-year	54.2*	40.4*	26.3	24.9

Prevalent cohort includes all subjects initiating vedolizumab or TNFi; incident cohort is the same as prevalent cohort with the exception of subjects with arthralgia prior to vedolizumab or TNFi initiation.

*P = 0.002.

The rate of incident arthralgia was 26.3 per 1000 person-years in the vedolizumab group and 24.9 per 1000 person-years in the TNFi group (P = 0.68). The mean time of first arthralgia after the index date (SD) for the TNFi group was 0.31 (0.30) years, and for vedolizumab it was 0.20 (0.23) years. We observed that the majority of arthralgia occurred within 6 months of initiating treatment in both groups (Supplementary Figure 1A, B). In the IPTW-weighted analysis, the hazard ratio for arthralgia in the vedolizumab group compared with TNFi was 1.20 (95% confidence interval, 0.97–1.49).

In the secondary analyses, we observed similar results when further adjusting by age or studying only subjects with incident arthralgia after initiating vedolizumab or TNFi (Table 4). Additionally, for both treatments, the burden of arthralgia was generally lower after initiating treatment (vedolizumab: P = 0.048; TNFi: P = 0.009).

Table 4:

Results of Cox Regression Models Assessing Risk of Arthralgia

	Unadjusted HR	IPTW Weighted Modela, HR	IPTW + Age Model, HR	Incident Cohort
TNFi	Ref	Ref	Ref	Ref
Vedolizumab	1.63 (1.34–1.98)	1.20 (0.97–1.49)	1.18 (0.95–1.46)	1.13 (0.80–1.60)

Abbrevaition: HR, hazard ratio.

^aIPTW conditioned on all variables in Table 1 (mean IPTW [SD], 0.99¹⁸ [0.52]).¹⁹

DISCUSSION

In this study, we tested an approach to assess for a potential association between arthralgia and vedolizumab use in a large EMR-based cohort with clinical data extracted using NLP. This study highlights a potential adverse effect where coding was suboptimal in the EMR but is mentioned in the narrative progress notes. Our approach using narrative data to study arthralgia may explain why the prevalence of arthralgia in this cohort was higher than previous IBD studies, at 43.3%. Although we observed an overall higher prevalence of arthralgia in vedolizumab vs TNFi users, there was no difference in incident arthralgia symptoms between new initiators of vedolizumab compared with TNFi after adjusting for potential confounders.

We interpret these data to suggest the following. Although TNFi therapies have a direct effect on inflammation in the joints, extra-intestinal manifestations are thought to be driven by gut inflammation.¹⁷ Thus, the therapy most effective for the bowels will likewise have an impact on joint symptoms. In line with this fact, we observed no differences in the rates of arthralgia in vedolizumab vs TNFi initiators after adjusting for potential confounders using IPTW.

When comparing vedolizumab with TNFi in the unadjusted analysis, we observed higher rates of arthralgia in the vedolizumab compared with the TNFi group. The finding of a higher prevalence among vedolizumab users was not unexpected. Vedolizumab is commonly prescribed after TNFi due to insurance requirements demonstrating inadequate response first of a TNFi. Vedolizumab users in general had a longer disease duration and potentially had more severe disease, as evidenced by the need for vedolizumab. In turn, these subjects may be more likely to have arthralgia due to active inflammation in the gut and may have more opportunities to develop arthralgia due to the longer disease duration. However, after adjusting for potential confounders including follow-up time, no differences in arthralgia were observed. Additionally, when studying only subjects with incident arthralgia after initiation of vedolizumab or TNFi, we did not observe a difference in rates between the 2 groups.

In the present study, we also highlight that applying NLP to identify subjects with arthralgia improved the sensitivity for cases by 31% compared with using ICD9 codes alone. Additionally, the standardized UMLS definition of arthralgia used in this study yielded good performance characteristics against gold standard manual chart reviews, with a PPV of 90%. This provides some data for the use of NLP with standardized UMLS definitions to screen for adverse outcomes that are suboptimally coded in the EMR data.

A limitation of this study includes the fact that it was performed using EMR data from 2 tertiary care centers, which may not be generalizable to other institutions. However, 2 advantages to this approach are the large number of IBD subjects and the ability to perform an active comparator design. Additionally, the severity of joint pain could not be measured because it was not uniformly measured at each visit. Thus, the study focused on the presence or absence of arthralgia. Even with a formal evaluation, clinically it is often difficult to distinguish drug-induced arthralgia from IBD-related arthralgia, underscoring a potential reason for the paucity of studies in this area. To approximate the burden of arthralgia, we assessed the number of times arthralgia is mentioned before and after either vedolizumab or TNFi initiation, standardized by follow-up time. In this small secondary analysis, we observed fewer mentions of arthralgia after treatment. Thus, while the primary analysis showed no difference in the rate of arthralgia after initiation of vedolizumab or TNFi, the burden or frequency of mentions for arthralgia was reduced after initiation of either treatment. Due to the relatively small number of individuals in the vedolizumab cohort, we did not stratify the analysis by CD or UC status.

Patients can have many reasons for joint pain. Misclassification of joint pain from other causes such as osteoarthritis, if unbalanced between the 2 groups, could attenuate any signals for increased arthralgia in any treatment groups. The IPTW approach was applied to mitigate this issue. Lastly, this study followed patients for up to 1 year. This was done because vedolizumab is a relatively new treatment with limited follow-up data in the EMR. We standardized the follow-up time for all patients to 1 year to allow for similar opportunities to develop and have arthralgia reported in their medical record. Additionally, based on our data, the majority of arthralgia occurred within 6 months of initiation of either treatment. Future studies with longer follow-up periods can provide additional information on arthralgia, and the nature of the arthralgia in IBD patients treated with long-term vedolizumab compared with TNFi.

In summary, we performed a study using NLP to efficiently assess potential associations between a relatively new treatment and an adverse outcome, arthralgia, that is infrequently coded by gastroenterologists. In this study of a large IBD cohort followed at a large tertiary care center, we did not observe an increased rate of incident arthralgia after vedolizumab initiation compared with subjects initiating TNFi. More studies are needed to replicate these findings with longer-term follow-up.

SUPPLEMENTARY DATA

Supplementary data are available at Inflammatory Bowel Diseases online.

Conflicts of interest: None of the authors has any conflicts of interest or disclosures to declare.

Supported by: This project was supported by the National Institute of Health p30 AR072577; A.A. receives support from the Crohn’s and Colitis Foundation; K.P.L. receives support from the Harold and DuVal Bowen Fund.