67 - Creation and Validation of Algorithms to Identify Patients with Moderate-to-Severe Osteoarthritis of the Hip and/or Knee and Inadequate/Intolerable Response to Multiple Pain Medications

¹Evidera, Waltham, MA, USA. ²Eli Lilly and Company, Indianapolis, IN, USA. ³Pfizer Inc., New York, NY, USA. ⁴Algosunesis, LLC, Longboat Key, FL, USA

Purpose
Osteoarthritis (OA) is a degenerative joint disease involving the cartilage and surrounding tissues that impacts approximately 32.5 million US adults.¹ While its etiology is multifactorial, risk increases with advancing age—2.1%, 19.8%, and 39.7% of persons aged 15-49 years, 50-69 years, and ≥70 years, respectively, are reported to have OA.² Two of the most commonly affected sites are the hips and the knees (it also can occur in hands, facet joints, and feet).³ OA is typically associated with pain, stiffness, and swelling, which collectively result in increasing levels of disability over time. Levels of disability can in fact preclude employment, with OA patients almost twice as likely as those without this condition to be unable to gain employment.⁴ In addition to its deleterious physical and psychologic impacts, OA is associated with substantial societal and economic burden. One relatively recent study estimated annual mean total costs of OA to be approximately $16,145⁵ (expressed in 2019 US dollars), of which direct and indirect costs represent 71% and 29%, respectively.
While the deleterious impacts of OA are substantial, they are experienced disproportionately among those with moderate-to-severe disease and/or with relatively high levels of pain.^3,6 Quantification of the burden among this subgroup would increase knowledge about OA that could in turn help inform development of new analgesic therapeutics. The most cost-efficient and timely means to undertake such examinations would be a retrospective cohort study based on analyses of large electronic healthcare databases. However, these databases typically lack the clinical information required to identify disease severity/progression, pain levels, and related measures. Conversely, while patients’ medical records typically contain the clinical information necessary to ascertain disease severity and pain levels, they lack cost data and may not offer “complete” capture of utilization of healthcare services, as they are site/physician-specific. Thus, we undertook a limited pilot study to determine whether case-selection algorithms can be developed for use in electronic healthcare databases to identify with reasonably good accuracy patients with moderate-to-severe OA of the hip and/or knee and inadequate/intolerable response to ≥2 pain-related medications, using data sources that contained both electronic healthcare data and medical records.
Methods
We used data from the Reliant Medical Group (“Reliant”). Reliant is a large, private, multispecialty group practice (>250 physicians in >20 locations) in central Massachusetts, with >1 million patient visits annually. For subgroups in the system, administrative data and curated electronic medical record data are available, as is the ability to conduct chart review. Institutional Review Board approval was obtained before information from patients’ charts was abstracted and analyses of their electronic health data was undertaken.
This pilot study comprised 14 adults (i.e., aged≥18 years), including 10 (71.4%) with at least one claim consistent with OA of hip or knee (defined as either≥1 medical claims resulting in diagnosis codes for OA of hip or knee, or≥1 medical claims for total/partial/revision of hip or knee arthroplasty); 2 (14.3%) with claims consistent with joint or back pain but without diagnosis codes for OA of hip or knee; and 2 (14.3%) without any claims that resulted in diagnosis codes for OA (any site), evidence of administration of hyaluronic (HA) acid or intra-articular (IA) steroids, or evidence of hip or knee arthroplasty. All criteria were assessed between January 1, 2014 and December 31, 2019 (“study period”).
We used literature review and expert opinion to develop case-selection algorithms, limited to measures expected in electronic healthcare databases (specifically, claims), to identify patients with moderate-to-severe disease and inadequate/intolerable response to ≥2 pain-related medications. Patient allocation via algorithm was compared against information obtained from medical charts, which were considered the gold standard for patients’ “true status”. We assessed algorithm performance based on sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy.
Results
Mean age was 69.1 years and 78.6% were women. Most (78.6% of the study sample) were Caucasian; 14.3% and 7.1% were Hispanic and unknown race/ethnicity, respectively. Median (interquartile range) body mass index was 32.6 (26.5 - 40.0) kg/m².
While several different pain and disease severity scales were assessed in patients’ charts, among those selected for the pilot test, the former was limited to numeric or verbal rating scales. Similarly, OA-specific severity assessment scales (e.g., Western Ontario and McMaster Universities Arthritis Index [WOMAC]) were not commonly recorded in patients’ medical charts. Patients averaged 6.1 pain assessments during the study period, albeit of varying frequency (range: weekly to once every two years). Functional disability was recorded in medical notes for 6 patients (42.9% of the study sample); information on degree of disability was often missing.
Chart review determined that 11 patients (78.6% of the study sample) had confirmatory evidence of OA of hip and/or knee. Among patients with “true” OA of the hip/knee, 7 (63.6%) had confirmatory evidence of moderate-to-severe disease. Among those with “true” moderate-to-severe disease, 5 (71.4%; 50.0% of patients with claims consistent with OA of the hip/knee; and 35.7% of the study sample) had confirmatory evidence of inadequate/intolerable response to≥2 pain-related medications.
PPV of our a priori algorithms for identifying patients with moderate-to-severe OA of the hip and/knee and inadequate/intolerable response to ≥2 pain-related medications ranged from 40.0% (≥2 medical encounters [inpatient or outpatient] with diagnoses of OA of the hip/knee within a 6-month period, ≥2 different opioids within a 90-day period) (2 of the 5 patients identified with this algorithm were “true positives”) to 100.0% (≥1 encounters with diagnoses of OA of the hip/knee, ≥1 mobility aids received within 30 days of an OA diagnosis [any site], ≥2 prescriptions for topical NSAIDs within 90 days of an OA diagnosis, and receipt of knee/hip arthroplasty [partial, complete, or revision]) (the single patient identified with this algorithm was a true positive). Estimates of sensitivity, specificity, and NPV for the algorithm with relatively low PPV were 40.0%, 66.7%, and 66.7%, respectively; corresponding estimates for the algorithm with relatively high PPV were 20.0%, 100.0%, and 69.2%, respectively. Overall accuracy ranged from 57.1% for the algorithm with relatively low PPV to 71.4% for the algorithm with relatively high PPV. Total true cases identified by any single algorithm ranged from 0 to 3.
Conclusions
While maximally efficient in terms of time and budget, electronic healthcare databases lack the necessary clinical detail to readily identify severity and pain associated with OA of the hip and/or knee. Nearly all patients in our pilot study were readily and accurately identified as having OA of the hip and/or knee using the a priori algorithms developed based on literature review and expert opinion; however, these algorithms were much less accurate in their ability to further differentiate by disease severity and/or pain levels. Moreover, no one algorithm identified more than 60% of total true cases. Classification of disease severity and pain levels was somewhat problematic as information required to inform these decisions was not consistently captured in patients’ medical records. Collectively, our findings suggest that potentially coupling detailed information from patients’ electronic records with their healthcare claims and/or using techniques of machine learning are needed to enable use of large, geographically representative healthcare data sources in the identification of this important subgroup of patients with OA of the hip and/or knee and subsequent quantification of burden of illness and levels of unmet need associated with current standard of care.