Trajectories of oral bisphosphonate use after hip fractures: a population-based cohort study

Our population-based cohort study used linked data from the Victorian Admitted Episodes Dataset (VAED), Pharmaceutical Benefits Scheme (PBS), and National Death Index (NDI). VAED are routinely collected administrative data for all public and private hospital episodes in Victoria, Australia. The dataset contains comprehensive information on demographics, diagnoses, admission sources, and discharge destinations. Records for readmission within 1 day of discharge were combined with original admission as a single continuous hospitalization to account for transfers between hospitals. Hip fracture ascertainment using Australian administrative hospital data has a sensitivity (i.e., proportion of hip fractures captured in administrative data) around 95% and positive predictive value (i.e., proportion of recorded hip fractures were a new hip fracture) above 70% [19]. We restricted our investigation to first hip fracture to minimize false positives due to misclassification of subsequent episodes of care as new hip fractures. The PBS dataset contains information on pharmacy claims made by all Australian residents for medications subsidized by the Australian Government. This dataset contains records of medications dispensed at all community pharmacies, outpatient clinics and at hospital discharge. The PBS dataset does not contain information of medications dispensed during a hospitalization. NDI contains information on all deaths registered across Australia. Data from VAED and PBS were available from 1 July 2006 to 30 June 2018, whereas data from NDI were available from 1 July 2012 to 30 June 2018. Details on the datasets have been published previously [4].

All patients aged 50 years or above and discharged from any public or private hospital in Victoria between 1 July 2012 and 30 June 2017 following hip fractures were included. Patients with a principal diagnosis of hip fracture (International Statistical Classification of Diseases and Related Health Problems, Tenth Revision, Australian Modification [ICD-10-AM] S72.0-S72.2) [19‐22] were included. Patients with a hip fracture or cancer diagnosis (ICD-10-AM C00-C99) within 5 years prior to the index hip fracture were excluded to ensure we investigated incident hip fractures and their bisphosphonate use were not affected by other indications (e.g., bone metastases or hypercalcemia). To avoid including patients who discontinued due to intolerance after their first dose, our study included patients with at least 2 dispensings of any oral bisphosphonates (World Health Organization Anatomical Therapeutic Chemical Classification [ATC] M05BA04, M05BA07, M05BB02-M05BB07) within 12 months after discharge. Patients dispensed other osteoporosis medications (i.e., denosumab [ATC M05BX04], zoledronic acid [ATC M05BA08, M05BB08], and raloxifene [ATC G03XC01]) within the follow-up period were excluded. Patients dispensed denosumab within 6 months before discharge were also excluded to avoid including patients not initiated on bisphosphonates at discharge for this reason. Patients who died within 12 months after discharge were excluded [23, 24]. Patients who were hospitalized for more than 1 month in the 12 months post-discharge were also excluded to avoid missing data from lack of inpatient dispensing data.

Dispensings of oral bisphosphonates were ascertained for each 30-day period after discharge for 12 months. Each oral bisphosphonate dispensing in Australia represents 30 days’ supply. If there were two or more bisphosphonate dispensings in a 30-day period, the extra prescriptions were carried forward to the subsequent months for which there was no dispensing. Trajectories of oral bisphosphonate use were identified by group-based trajectory modeling with SAS PROC TRAJ [25]. The number of trajectories was first determined using the highest order possible with the procedure, which allows the maximum number of inflection points thus flexibility in fitting each trajectory [15]. Bayesian information criterion (BIC) across all models was compared and models with less negative BIC were preferred (Supplemental Table 1) [15]. A model with 4 trajectories was chosen as any larger number of trajectories did not provide stratification of clinical significance. Different orders for each trajectory were then explored (Supplemental Table 2). Average posterior probability of group membership (AvePP) and odds of correct classification (OCC) for each trajectory were calculated to evaluate adequacy of the models [15]. The final best-fitting model was chosen based on the following criteria: (a) AvePP > 0.7, (b) OCC > 5, (c) the highest order parameter of each trajectory was statistically significant, (d) the group proportions estimated by the model was similar to the actual proportions of individuals assigned to each group based on their maximum posterior probability, (e) the BIC, (f) parsimony principle, and (g) clinical judgement (Supplemental Table 2) [15].

We investigated factors that have been associated with adherence and persistence to oral bisphosphonates in previous research, including selected comorbidities related to cognitive, mental, and gastrointestinal health [11]. We also investigated factors that were associated with 1-year mortality following hip fracture in our previous research [4]. These factors included age, sex, discharge to residential aged care facilities (RACF), and frailty [4]. Frailty was assessed using the validated Hospital Frailty Risk Score (HFRS), a weighted score of 109 diagnoses recorded within 2 years prior to discharge [26]. Patients were categorized as having low frailty risk (HFRS < 5), intermediate frailty risk (HFRS 5–15), and high frailty risk (HFRS > 15) [26]. Comorbidities were identified from ICD-10-AM diagnosis codes recorded within 5 years prior to discharge. Previous falls and previous osteoporosis were identified from ICD-10-AM diagnoses recorded within 5 years prior to admission. A 5-year lookback period was used to improve the detection of comorbidities across multiple hospitalizations for each patient. This was because not all comorbidities may have been recorded at the index hospitalization. Pre-admission bisphosphonate use (any oral bisphosphonate dispensing within 6 months prior to admission) and pre-admission polypharmacy (5 or more different medications dispensed within 60 days prior to admission) were also assessed. Factors associated with different trajectories compared to the persistent adherence trajectory were assessed using multivariate multinomial logistic regression. The model was adjusted for age, sex, discharge to RACF, HFRS, comorbidities, previous falls, previous osteoporosis diagnosis, pre-admission bisphosphonate use, and pre-admission polypharmacy. Statistical significance was defined as p-value < 0.05. All analyses were performed using SAS (version 9.4) and R (version 4.0.0). As management of osteoporosis is sex-specific, sex-stratified subgroup analysis was performed to examine any differences across sex in bisphosphonate use after hip fractures. Sensitivity analysis was conducted to explore the possible effect of subsequent fractures on the analysis by excluding patients with second hip fractures from the multivariate multinomial logistic regression.

The study was approved by Australian Institute of Health and Welfare Ethics Committee (EO2018-4–468) and Monash University Human Research Ethics Committee (14,339).

The four trajectories identified by group-based trajectory modeling were labelled as “persistent adherence,” “delayed dispensing,” “early discontinuation,” and “late discontinuation.” The proportions of patients that were categorized into the different trajectories were 66% (n = 1191), 17% (n = 305), 9% (n = 161), and 9% (n = 154) respectively. Patients with “persistent adherence” had > 80% probability of monthly bisphosphonate dispensing throughout the 12 months. The probability of dispensing bisphosphonates for patients with “delayed dispensing” was low initially and increased to 50% at 6 months after discharge. The probability of bisphosphonate dispensing for patients with “early discontinuation” and “late discontinuation” decreased to 50% at around 3 and 8 months respectively (Fig. 1 and Supplemental Table 3).

Patients aged 85 years or older and patients with intermediate frailty risk (HFRS 5–15) were more likely to have persistent adherence, as evidenced by their lower risk of delayed dispensing (\(\ge\) 85 years old versus 50–64 years old: relative risk [RR] 0.38, 95% confidence interval [CI] 0.22–0.64; HFRS 5–15 VS HFRS < 5: RR 0.66, 95% CI 0.47–0.93). In addition, patients who were on bisphosphonates pre-admission had a lower risk of delayed dispensing (RR 0.28, 95% CI 0.21–0.39). Patients with dementia were at higher risk of late discontinuation when compared with persistent adherence (RR 1.92, 95% CI 1.09–3.38), while patients with anxiety disorder had a higher risk of early discontinuation (RR 1.92, 95% CI 1.08–3.43) (Table 2).

When stratified by sex, females aged 85 years or older were more likely to have persistent adherence, as demonstrated by their lower risk of delayed dispensing (RR 0.39, 95% CI 0.21–0.72). However, there was no difference for males aged 85 years or older (RR 0.35, 95% CI 0.12–1.03). However, the estimates remained similar. Similarly, females but not males with intermediate frailty risk were more likely to display persistent adherence, as evidenced by their lower risk of delayed dispensing (females: RR 0.68, 95% CI 0.46–1.00; males: RR 0.57, 95% CI 0.26–1.24) with intermediate frailty risk. For patients who were on bisphosphonates pre-admission, decreased risks remained evident for delayed dispensing in both sexes (males: RR 0.17, 95% CI 0.06–0.49; females: RR 0.30, 95% CI 0.21–0.42). Higher risk of early discontinuation for patients with history of anxiety disorder remained significant for males (RR 9.81, 95% CI 2.24–42.9) but was not evident in females (RR 1.37, 95% CI 0.69–2.70). Additionally, females hospitalized for falls previously had higher risk of early discontinuation (RR 1.80, 95% CI 1.16–2.78). Higher risk of late discontinuation in patients with dementia became insignificant for both males (RR 1.82 95% CI 0.49–6.78) and females (RR 1.89 95% CI 1.00–3.58) (Tables 3 and 4).

Sex, type of residence, history of depression, gastroesophageal reflux, previous osteoporosis diagnosis, and pre-admission polypharmacy were not associated with different trajectories in both main and sex-stratified analyses (Tables 2, 3, and 4). Sensitivity analysis excluding patients with second hip fractures from the analyses produced similar results for both overall and sex-stratified analyses. The study is reported in accordance with STROBE statement (Supplemental Table 4) [27].

Strengths of our study include the use of linked administrative health data for all patients with hip fracture in Victoria [42]. Our group-based trajectory modeling was based on 30-day dispensing rather than prescription data or 60-day or 90-day dispensing. Group-based trajectory modeling was an agnostic approach to identify trajectories which avoided potential bias arising from defining adherence using arbitrary cut-offs or dichotomization of patients [16, 43, 44]. Particularly, our model identified patients with delayed dispensing and late discontinuation. These dispensing patterns would not have been evident had we calculated the possession ratio (MPR), another adherence measure with common cut-offs of low (MPR < 0.5), moderate (MPR 0.5 to < 0.8), and high (MPR 0.80 to 1.00) adherence. Our study allowed identification of risks factors for bisphosphonate use patterns that would have been obscured otherwise. Another strength of our study is our stringent inclusion and exclusion criteria. By including only those with at least 2 dispensings of bisphosphonate and excluding patients who used other anti-osteoporosis medications during follow-up, our results are not biased by patients who discontinued due to intolerance or contraindications or those who switched to other anti-osteoporosis medications.

While this study included the entire population in Victoria, Australia, one limitation of our study is that it may not be generalizable to other populations. Patterns of bisphosphonate use may be different in other populations due to differences in guidelines around medication selection, initiation, availability of dental examination, and medication reimbursement criteria. Other limitations include we only included patients dispensed two or more oral bisphosphonate prescriptions and did not investigate the characteristics of patients not initiated on bisphosphonates. While we adjusted for pre-admission bisphosphonate use in our analysis, we did not account for the duration of previous use. Our study also did not investigate the use of parenteral antiresorptives (e.g., denosumab, zoledronic acid) because of their different dosing and dispensing frequencies. We did not exclude patients dispensed zoledronic acid within 1 year prior to discharge. However, we do not anticipate this would result in significant overestimation of delayed dispensing because oral bisphosphonates and denosumab are the most common antiresorptives in Australia [6]. Misclassification bias, common in administrative data, is also present in our study. However, data quality was maximized with regular Australian Government data integrity audits [45] and our frequent monthly data also minimized the effect of single dispensing on trajectory determination. It was a strength of our study that we used a 5-year lookback period to maximize detection of comorbidities. However, we acknowledge that this may have led to the inclusion of comorbid conditions that were no longer acutely symptomatic. We were also not able to investigate why patients exhibit their respective trajectories. However, this may be partly related to patients’ beliefs regarding the benefits and harms of medications, which have been demonstrated to impact adherence [46].