Chronic pain in rheumatic diseases (RDs), a symptom often neglected, is prevalent and an important contributor to poor health and reduced quality of life. Various studies report that up to a third of patients with inflammatory RDs have persistent pain, often attributed to pain sensitization, even when the underlying disease is controlled.1-3 Beyond personal suffering, chronic pain has considerable societal implications that include increased healthcare utilization and economic disadvantages. It therefore follows that chronic pain deserves to be highlighted as an important comorbidity of RDs, deserving attention, management, and further study. A reliable estimate of the prevalence of chronic pain is necessary to better understand the contributors, life impact, and treatment of RD-associated pain. Electronic health records (EHRs) are ideally poised to provide information on large numbers of unselected patients in a real-world setting with an opportunity to contribute to a better understanding of chronic pain.

A critical issue is whether the information currently available in usual EHRs can sufficiently categorize patients experiencing subjective symptoms. Information collected in EHRs is mostly in real time and entered in a structured format using controlled vocabulary, including demographics, diagnostic codes (eg, International Classification of Diseases [ICD] codes), procedure codes, laboratory test results, medications, and unstructured clinical free-text data.4,5 Developed to track and manage patients, the near universal adoption of EHRs has provided researchers with the opportunity to access real-life health data that is free of the inherent biases when patients are specifically selected for study.4 EHRs have evolved to broaden the scope of an electronic medical record with inclusion of data beyond the clinical encounter, such as information on prescriptions filled, health data from other clinicians, and patient-provided data. At the most basic level, diagnostic codes provide estimates of disease prevalence, but a more detailed study of data in the EHR can provide disease-associated information that can be used to improve patient care, answer research questions, and influence health policy. When EHRs are used for secondary purposes such as research, reliability in identification of cohorts is essential. Case definitions or phenotypes can be constructed using algorithms that can include various variables in the patient record. This expansion from simple binary information such as diagnosis can provide more in-depth information of disease characteristics, but identifying data parameters that should be included in algorithms to accurately phenotype medical illness is critical. This is especially relevant when addressing softer outcomes that are not specifically entered in a structured format. Algorithms rely on the data that has been entered and require validation before they can be fully accepted.6

As of January 2022, the ICD revised their diagnostic codes to include the indication for primary chronic pain.7,8 This was a key first: the 11th revision of the ICD (ICD-11) also provided other codes for the most common and relevant groups of pain conditions. Under the ICD-11, pain is not only recognized as a health condition on its own but is also recognized as a symptom that can be secondary to other underlying health issues. The revision makes clear that chronic pain can lead to disability and distress, which, unbelievably, was not recognized officially until 2 years ago. The invisible and stigmatizing condition of chronic pain has long been ignored. Although this code now exists, it is unclear how often it is being used and/or recorded in EHRs. Pain has no singular home and is often buried as a symptom in conditions such as RDs. Conditions characterized predominantly by pain, such as chronic migraines or fibromyalgia, should also use these chronic pain codes; however, there exists the inherent risk of failure to capture the full scope of chronic pain.

Over a decade ago, prior to the advent of the diagnostic codes above, Tian et al developed an algorithm that combined pain scores, diagnostic codes, and prescription opioids to identify chronic pain in a primary care community setting with a prevalence of 19%.9 By combining other variables that commonly associate with chronic pain into a multimodal algorithm, researchers have proposed a more reliable assessment of chronic pain prevalence. Accuracy in identifying disease-associated phenotypes is critical, especially in the absence of a specific biomarker, but use of surrogate measures could have a risk of inaccuracy by either over- or underidentification.

In this issue of The Journal of Rheumatology, Falasinnu and colleagues describe the use of EHRs to diagnose chronic pain in patients with RDs, with particular attention to accuracy in assessment.10 The EHRs of 3042 patients with RDs (ankylosing spondylitis, psoriatic arthritis, Sjögren syndrome, systemic lupus erythematosus, and systemic sclerosis) visiting any Stanford outpatient rheumatology clinic during 2019 were examined. Guided by the previous study of Tian et al,9 the researchers expanded and refined the choice of variables to include 4 criteria in various algorithms to estimate variations in the prevalence of chronic pain. The 4 variables chosen were pain scores (0-10), pain diagnostic codes according to ICD-10, Clinical Modification (ICD-10-CM; eg, abdominal pain, chest pain, cervicalgia, fibromyalgia), pain medication prescriptions including opioids, nonsteroidal antiinflammatory drugs, anticonvulsants, antidepressants, muscle relaxants, and acetaminophen, and pain interventions including, among others, nerve blocks, joint injections, acupuncture, and physical therapy. Using these expanded criteria, the overall prevalence of chronic pain was assessed to be 37%.

The methodology of this exercise began with assessment of prevalence of chronic pain using each of the unimodal criteria alone, and thereafter using various combinations. In the first iteration, using each of the 4 unimodal criteria alone, the prevalence of chronic pain ranged from 12% to 23%, with female individuals more commonly identified with chronic pain for all assessments. Prevalence rates were as follows: pain scores identified 18% of patients (20% female, 12% male); ICD-10-CM codes identified 23% female and 15% male patients; pain prescriptions 16% female and 19% male patients; and pain interventions 23% female and 13% male patients. However, the correlation coefficients between unimodal pain phenotyping were of low to medium strength. As a next step, the various variables were combined as bimodal (combination of 2), trimodal (combination of 3), and quadrimodal (combination of 4) phenotyping algorithms. Not surprisingly, when algorithms were used with 11 possible combinations, the prevalence estimates for chronic pain increased and ranged from 28% to 47%. According to all algorithms studied, female patients were more likely to be categorized with chronic pain. Sjögren syndrome and systemic lupus erythematosus were the conditions most frequently classified with chronic pain. Validation of the various algorithms was not reported.

The current study10 can be seen as an update and a refinement of the previous study by Tian et al.9 This latter study was conducted in a multisite community health center in Connecticut and included data of 38,520 patients who had at least 1 medical visit between March 2011 and February 2012. Using an algorithm that included ICD-9 codes that identified chronic pain, pain scores ≥ 4 on a scale of 10, and receipt of at least 90 days of opioid medication, chronic pain was identified in 19% of patients. Validation of the chronic pain algorithm was by review of 381 randomly selected charts, with 74 identified with chronic pain, with a sensitivity of 84.4% and a specificity of 97.7%. Further analysis showed that those with chronic pain were more likely to be women, have increased primary care services and behavioral health visits, and that 43% had received at least a single prescription of opioids in the study year.

Plausible explanations for the increased prevalence of chronic pain reported in this current RD population10 include the following: pain accompanies most rheumatic complaints, inclusion of more recently used pain medications such as antiepileptics and antidepressants, and the inclusion of procedures that included joint injections and blocks. Future studies in the rheumatology population should include patients with rheumatoid arthritis, as well as information on contextual factors such as education, work and economic status, healthcare utilization, and information on prescribers. Importantly, the practicing rheumatologist should ask about the status of the disease, including duration of illness, severity, and specific treatments.

EHR data-driven phenotyping is increasingly used to add another depth of study of disease, including associations, risk factors, and longitudinal outcome, with information used to design clinical trials that address identified issues. Diagnostic codes alone are a poor estimate of disease associations, with a tendency to underreport. In a study using EHRs to identify nonalcoholic fatty liver disease, disease prevalence was underestimated at 40%, but by adding type 2 diabetes mellitus using ICD-10 codes, accuracy was improved to 95%.11 This algorithm was further refined and provided increased accuracy by the inclusion of natural language and triglyceride levels.12 Other examples include algorithms that have been used to identify conditions such as decompensating heart failure13 and the epidemiology of uterine fibroids.14 The increasing sophistication in using data that has not been defined a priori or that may be found in the unstructured record provides an opportunity to expand study beyond simply diagnosis, but with the opportunity to examine factors such as disease severity, associations, and probabilities.15 Reliable phenotyping is therefore the key to identifying disease-specific cases or characteristics.

The current study by Falasinnu et al10 provides a cautionary message about the inherent inaccuracies when subcategorizing patients within a specific disease state by use of EHRs. Researchers will need to clearly identify the ideal variables to be included in any algorithm that seeks to identify features of disease that are not entered as structured or hard data to mitigate the risks of inaccuracy. Clearly, the gold standard is the clinical assessment to specifically answer a question, but the ability to access a vast amount of data in large populations and across various jurisdictions is a positive step toward better understanding of chronic illnesses.

Well-designed EHRs are positioned to become the standard of healthcare research, with many advantages. Beyond facilitating patient care, EHRs have the potential to be a treasure trove of real-world clinical data for epidemiological study. However, there still remains a lack of standardization in terminology and various data elements that will hinder EHR-based research. Developers will need to balance a fine line to ensure that pertinent information is entered, but not at the expense of copious redundant information that will lead to worker fatigue and inaccuracy in recording. Complexity, inaccuracies, and missing data in the EHR are currently inherent challenges. Electronic phenotyping will improve as algorithms incorporate clinical variables, including the ability to extract meaningful information from clinical notes by use of natural language processing.16 From an ethical standpoint, it is also important to respect patients’ privacy rights and ensure that data are collected and handled in a manner that is compliant with applicable laws and regulations. When EHRs eventually function seamlessly, they have the potential to accumulate a vast amount of real-world healthcare data that will be valuable to understand health status from many aspects. The ability to connect across networks of healthcare providers will allow for longitudinal multicenter studies from diverse populations to have the ability to identify trends and risk factors that were not evident from smaller studies. Any step forward to facilitate the care of patients with RDs is welcomed. Over time, EHRs will become commonplace and easier to complete for healthcare providers, who will in turn be able to focus on relevant information that is both clinically applicable and meaningful epidemiologically. With the advent of the new diagnostic codes for chronic pain, newer algorithms will emerge in the coming years to replace those used by Tian and colleagues.9 Vast opportunities lie ahead for researchers studying pain through the use of EHRs, but conclusions will be valid only if data capture is rigorous and accurate.

HC is funded by a Merit Award from the Department of Anesthesiology and Pain Medicine at the University of Toronto; is president-elect of the Canadian Pain Society, and president of the Canadian Consortium for the Investigation of Cannabinoids. MAF was a core member of the Health Canada Science Advisory Committee on Health Products Containing Cannabis 2020-2022; and lead on the Canadian Rheumatology Association position statement on cannabis for the rheumatology patient.

See Chronic pain phenotyping algorithms, page 297