Uncertainty of current algorithm for bisphosphonate-related osteonecrosis of the jaw in population-based studies: a systematic review

Introduction

Bisphosphonates (BPs) are pyrophosphate analogues that bind to the hydroxyapatite crystals in hard tissue and suppress osteoclast-mediated bone resorption by inducing the apoptosis of osteoclasts.1 Since the first approval of nitrogen-containing BPs for the prevention and treatment of osteoporosis in 1995, numerous clinical trials have demonstrated their efficacy in the treatment of various bone diseases including osteoporosis, multiple myeloma, and metastatic bone cancers.2, 3

However, growing concern has been raised regarding the potential complications after prolonged use of BPs. In 2011, the US Food and Drug Administration (FDA) officially issued a cautious statement on the causal association between long-term use of BPs and bisphosphonate-related osteonecrosis of the jaw (BRONJ), atypical femoral fractures, gastrointestinal cancer, etc.4 One recent study from the FDA Adverse Event Reporting System even reported 171-fold increased risks of BRONJ after administration of zoledronate.5 However, because many aspects and mechanisms underlying BRONJ remain unknown, numerous clinical and epidemiologic studies have been conducted.6-9

The recent increase in accessibility to large population-based health-care databases has enabled researchers to carry out large-scale epidemiologic studies on BRONJ.^(10–14) Such studies have successfully evaluated representative samples from large populations of patients within their prospective cohorts, reporting an incidence ranging from 0.01% to 4.3%.12

However, epidemiologic studies reporting BRONJ incidence levels on the lower end of this range generally possess several methodological limitations. The small number of BRONJ cases in the associated cohorts implies relatively low statistical power, whereas selection bias in nested case-control designs is inevitable because of limited medical information of controls in the database.11, 15 Furthermore, insurance claim data were primarily designed for reimbursement, not for medical research. A more appropriate measure may be the National Health and Nutrition Examination Survey, which contains detailed, individual medical information.

However, the most significant factor that renders these previous large-scale epidemiological studies of BRONJ less reliable is uncertainty with regard to the identification algorithm used to diagnose patients with BRONJ according to International Classification of Diseases (ICD) diagnostic codes. Although the first reports of BRONJ appeared in 2003,16 a specific ICD code—M87.1, Osteonecrosis due to drugs—was first introduced in 2010. Until that time, various codes for reimbursement had been applied, mostly according to convenience. Even for ICD code K10.2 (inflammatory conditions of the jaw), which has been regarded as the most frequently used diagnostic code,7, 10, 17 various clinical aspects of jaw diseases may be involved, such as chronic osteomyelitis, osteoradionecrosis, osteitis, premalignant lesions, nonspecific bone diseases, and unconfirmed bone tumors. Moreover, current diagnostic criteria for BRONJ according to the American Society for Bone and Mineral Research (ASBMR)18 and the American Association of Oral and Maxillofacial Surgeons (AAOMS)19 are characterized only by clinical manifestation, thus rendering BRONJ identification according to ICD code even more uncertain.

In spite of the aforementioned limitations, it is indisputable that claim-based studies using large samples present the most promise for studying disease of such rarity as BRONJ. To assess the relevance of previous epidemiologic studies and their methodology, the present study involved two main components. First, large population-based observational studies on BRONJ were systematically reviewed. Then, the authors evaluated the validity of the claims-based algorithm for BRONJ identification via estimation of positive predictive values.

Materials and Methods

Results

Systematic literature review

Using our search criteria, a total of 2016 articles were identified through Medline via the PubMed search. A total of 655 studies written in languages other than English or published before 2003 were excluded, while 998 studies were further excluded after review of titles and abstracts. After the full search, eight studies were identified from the reference lists of populated results, and 355 studies were excluded because of various reasons, as detailed in Fig. 1. Finally, 17 studies with relevant epidemiologic data were included in our analysis (Fig. 1).

The studies varied with regard to source population, study design, sample size, definition of exposure and outcomes, and control of confounding variables (Table 1). Of the included studies, 12 studies primarily used the population-based claim databases,7, 11-15, 17, 21-25 while five studies used limited clinical data.5, 8, 10, 26, 27 Some studies5, 7, 8, 10-12, 26 used data after the first report of BRONJ in 2003;16 however, most used data regardless of time. The sample included eight cohort studies and nine case-control studies.

Table 1. Summary of Epidemiologic Studies on BRONJ and Quality Assessment

Author, Year	Population data acquisition	ONJ patient identification	Criteria for BRONJ confirmation	Study design	No. of investigated population / no. of cases	Exposure (bisphosphonate)	Quality assessmenta
Zavras 200624	Insurance medical claims data, US 2001–2004	CPT codes	—	Cohort	255,757 / 224	Oral or IV BP / ever exposure	5
Wilkinson 200721	SEER claims data, US 1995–2003	ICD-9 codes or CPT codes	—	Case-control	16,073 / 95	IV BP / cumulative dose estimation	7
Mavrokokki 20078	Survey to OMS, AU and NZ 2004–2005	Surveys to OMS members	Exposed bone over 6 weeks	Cohort	— / 158	Oral or IV BP / ever exposure, type	5
Pazianas 200825	Insurance research database, US 2002–2005	CPT codes	—	Case-control	3505 / 697	Oral BP / duration	6
Cartsos 200813	NHIRD, US 2000–2006	ICD-9 codes or CPT codes	—	Cohort	714,217 /378	Oral or IV BP / ever exposure	5
Lo 201026	Kaiser Permanente of Northern California, US 2006	Survey, telephone interviews, clinical exam	AAOMS criteria	Cohort	8572 / 9	Oral BP / type, duration	8
Barasch 201127	DPBRN US, 2000–2008	Dental records, surveys, and telephone interviews	Necrotic jaw regardless of 8 weeks’ exposure or radiation	Case-control	— / 191	Oral or IV BP / type, duration, frequency, dose	6
Fellows 201122	Two large HMOs participating DPBRN US, 1995–2006	ICD-9 codes or CPT codes and chart review	Exposed necrotic lesion in maxilla or mandible	Case-Control	572,583 / 23	Oral or IV BP / type, duration, dose	7
Baillargeon 201123	5% national sample of Medicare beneficiaries, US 2000–2007	ICD-9 codes and CPT codes	—	Case-Control	9161 / 24	IV BP / ever exposure	5
Tennis 201214	HealthCore Integrated Research Database, US 2001–2006	ICD-9 codes or CPT codes and chart review	Nonspecific (“based on clinical judgment”)	Cohort	77,786 / 38	Oral or IV BP / type, duration	7
Vestergaard 201217	National Administrative Data, Denmark 1996–2006	Modified ICD-10 codes	—	Case-control	103,562 / 55	Oral BP / cumulative dosage	7
Solomon 201312	Two large insurance-based healthcare claims, US 2005–2010	ICD-9 codes and clinical exam	Nonspecific	Cohort	105,043 (BP user) / 8	Oral BP / ever exposure	6
Hallmer 201410	Four OMS database, Denmark 2003–2010	ICD-10 codes and chart review	AAOMS definition	Cohort	1.2 million / 55	Oral or IV BP / type	6
Lin 20147	NHIRD, Taiwan 2003–2007	ICD-9 codes	—	Case-control	18,030 / 25	Oral alendronate / type, duration, cumulative dosage	8
Huang 201515	NHIRD, Taiwan 2000–2010	ICD-9 codes	— (Not-specified)	Case-control	19,399 / 121	Oral BP / duration of exposure, dosage	7
Kwon 201511	NHIRD, South Korea 2004–2010	ICD-10 codes	—	Case-control	109,787 / 212	Oral BP / type, duration	6
Zhang 20165	FDA Adverse Event Reporting System (FAERS) 2010–2014	FAERS reports; “Osteonecrosis of the Jaw”	—	Cohort	— / 17,117	Oral and IV BP, RANKL inhibitor, antiangiogenic agents, m-TOR inhibitor	5

• CPT = current procedural terminology; BP = bisphosphonates; ICD = International Classification of Diseases; SEER = Surveillance, Epidemiology, and End Results; DPBRN = Dental Practice-Based Research Networks; HMOs = health maintenance organizations; NHIRD = National Health Insurance Research Database.
• ^a According to guidelines of Preferred Reporting Items for Systematic Review and Meta-Analyses20 and modified from Solomon et al.12 Refer to Supplemental Table S1 for details.

Surrogates used for the identification of ONJ were mostly ICD codes, current procedural terminology (CPT) codes, or both. Among ICD codes, inflammatory conditions of the jaw (K10.2) were frequently used as ONJ surrogates, while some studies used the newly proposed code—M87.1, osteonecrosis due to drugs11—or various related diagnostic codes.7, 10, 12, 22 After primary identification, only four studies performed manual confirmation of BRONJ by medical chart review or clinical examination,10, 12, 14, 22 and two studies applied AAOMS or ASBMR criteria for confirmation of BRONJ.10, 26

Settings for BP exposure were also varied. Seven studies included cases of oral BP exposure only,7, 11, 12, 15, 17, 25, 26 and two studies included cases of intravenous BP exposure only,21, 23 while seven studies included both routes.8, 10, 13, 14, 22, 24, 27 One study investigated various types of medications that could lead to development of the disease.5 Most studies examined limited information regarding exposure such as type of BP or whether simply exposed, although seven studies additionally examined duration of exposure. Detailed individual estimates of cumulative exposure according to type, dose, duration, and frequency were investigated in six studies;7, 8, 17, 21, 22, 27 however, only two studies investigated cumulative doses for risk estimates.7, 21 With regard to the control of confounding factors, six studies examined detailed patient characteristics, including dental and medical comorbidities,7, 11, 15, 17, 21, 25 whereas others presented limited dental or basic demographic information^{(8,14,22–24,27)} or no adjustment.5, 10, 12, 13, 26

Thirteen studies provided risk estimates for BRONJ (Table 2). Regardless of study design, type of BP exposure, and outcomes, most studies reported that use of BPs was associated with significantly increased risk of BRONJ.^{(5,11,13–15,17,21–24,27)} On the other hand, two studies reported no significant association,7, 25 while protective effects of oral BPs for BRONJ were shown in one study using claims data for patients with osteoporosis.13 In one study using adverse event reports from the US FDA, significantly increased risks for BRONJ were associated with denosumab, antiangiogenic drugs (sunitinib, bevacizumab, sorafenib), and m-TOR inhibitor (temsirolimus, everolimus) use, with odds ratios ranging from 1.4 to 13.8.

Table 2. Main Results of Risk Estimates for BRONJ in Overall Studies

Author, Year	Target population	No. of cases	No. of exposed non-cases, or controls if not available	Risk estimates	Exposures	Outcomes (surrogate for BRONJ)
Zavras 200624	Cancer	20	5830	OR 4.2 (2.67–6.72)	IV BP	CPT code
Wilkinson 200721	Cancer	95	14,349	HR 4.94 (3.33–7.34)	IV BP	Either ICD or CPT codes
Pazianas 200825	Osteoporosis	697	2808	OR 0.91 (0.70–1.19)	Oral BP	CPT code
Cartsos 200813	Osteoporosis	150	176,739	OR 0.65 (0.54–0.79)	Oral BP	ICD code
	Cancer	39	8168	OR 4.47 (3.19–6.27)	IV BP	ICD Code
Barasch 201127	Non-cancer	30	81	OR 7.2 (2.1–24.7)	BP	BRONJa
Fellows 201122	—	23	575,583	OR 13.4 (4.0–45.0)	“Any risk factor” including oral BP and medical comorbidities	BRONJa
Baillargeon 201123	—	24	2272	HR 1.65 (0.71–3.80)	IV BP	Either ICD or CPT codes
Tennis 201214	Cancer	26	46,516	RR 8.8 (2.02–38.36)	IV BP	“Probable” + “Possible” BRONJa
Vestergaard 201217	Osteoporosis	33	103,529	HR 3.15 (1.44–6.87)	Oral BP (alendronate)	ICD code
Lin 20147	Osteoporosis	25	18,030	HR 0.87 (0.47–1.58)	Oral BP (alendronate)	ICD code
Huang 201515	Osteoporosis	121	19,278	HR 2.05 (1.58–2.65)	Oral BP	ICD code
Kwon 201511	Osteoporosis	212	109,787	OR 3.86 (2.41–6.19)	Oral BP (“continuous user”)	ICD code
Zhang 20165	Cancer	11,490	— (Not specified)	OR 125.2 (115.7–135.8)	Zoledronate	FAERS report
	Cancer	1184	— (Not specified)	OR 4.9 (4.4–5.4)	Denosumab

• BP = bisphosphonates; CPT = current procedural terminology; ICD = International Classification of Diseases.
• ^a Confirmed by clinical examination or medical chart review.

Discussion

Since the first report of BRONJ in 2003, there have been various clinical and experimental attempts to understand the nature of the disease. Currently, however, there is still insufficient evidence regarding whether a true causal association exists between BP use and ONJ. Moreover, the specific risks and pathophysiologic mechanisms have yet to be clearly defined.19, 28 Given the extremely low incidence and lack of clinical data regarding risks of BRONJ, recent epidemiological studies using large databases have appeared advantageous in investigating the risk of this rare adverse drug event by providing longitudinal data and access to large reference populations.11, 26 In the present study, the authors conducted a systematic review of population-based observational studies on BRONJ. The included studies varied in study quality and presented inconsistent findings. Moreover, a high level of heterogeneity with regard to methodology (ie, different study setting, sample size, operational definition of outcomes, exposure, adjustment for confounding factors, etc.) made further qualitative synthesis not possible.

First, because of the different methods of acquisition for population data and variations in setting/study design, incidence rates and prevalence estimates varied dramatically up to hundreds of times.12, 22, 26 Risk estimates were also inconsistent, and some studies even reported protective effects of oral BPs for ONJ.13 Definition of exposed non-cases or controls was also unclear, thus making qualitative synthesis difficult.5, 11, 22, 27 Although BRONJ has been reported since the early 2000s, some studies performed sampling regardless of calendar year, including patients diagnosed before 2003.13, 15, 17, 21, 22, 24, 25, 27

Second, despite the fact that BRONJ is considered an adverse event induced primarily by BP administration, few studies provided detailed BP information regarding specific dose, duration of exposure, adherence to the drugs, and subsequent dose-response relationships. Most studies examined limited information such as BP type or whether exposure had occurred,5, 8, 10, 12, 13, 23, 24 whereas some studies additionally examined duration of exposure.11, 14, 25, 26 Considering the pharmacokinetics of BPs, which exhibit low bioavailability because of poor absorption and fast plasma clearance, their long dosing interval and extremely long skeletal half-life, applying cumulative dosing criteria rather than duration of BP exposure would be more appropriate for the determination of causal relationships.1, 19, 29 In addition, risk estimations under inappropriate settings of exposure (eg, without classifying the route of BP administration) were observed in some studies.5, 15 Because BPs present significantly different efficacy according to potency and bioavailability of each type,1 well-designed analysis of outcomes for each setting are required to establish accurate dose-response relationships.

Third, with regard to the control of confounding factors, some studies examined detailed patient characteristics containing dental and medical comorbidities,7, 11, 15, 17, 21, 25 whereas others presented limited basic demographic information or no adjustment.5, 10, 12, 13, 26 Given that many dental and medical risk factors for BRONJ have been elucidated,19 extensive consideration for possible confounding factors would be beneficial for future analyses.

AAOMS defines BRONJ according to the presence of exposed and necrotic bone in the maxillofacial region that does not heal within 8 weeks among patients with history of BP administration.30 However, this definition has continuously been criticized for its ambiguity and focus on clinical aspects alone.18, 31 Generally, this cannot easily be confirmed without close examination by oral and maxillofacial specialists; therefore, the AAOMS definition might not represent a feasible operational definition for large population-based epidemiology. Much lower prevalence and/or incidence was observed in studies that utilized manual confirmation of BRONJ diagnoses compared with other studies in which no such confirmation had occurred.8, 12 Moreover, even in the studies utilizing the AAOMS definition or similar criteria, the window period for disease induction was not taken into consideration except in a single study,14 thus reducing the accuracy of outcomes.

A considerable number of included studies used ONJ surrogates to determine outcomes rather than manual confirmation. Currently, however, there are few reports validating whether ONJ surrogates specifically reflect cases of BRONJ. If there is a significant difference between ONJ surrogates such as ICD/CPT codes and the actual disease, the relevance of previous ONJ epidemiologic studies using large population-based data may be jeopardized. Therefore, the authors evaluated the validity of the claims-based algorithm used for the identification of ONJ via estimation of positive predictive values.

In the present study, a population-based hospital registry was used to identify potential cases of ONJ, and validation of all cases by medical chart review was possible. Our findings showed that the overall PPV of the algorithm currently used to identify BRONJ is very low, indicating low validity and possible overestimation of ONJ occurrence. Among the empirically employed ICD-10 codes, most exhibited PPV values less than 5%. Only ICD code K10.2 (inflammatory conditions of the jaw) was associated with a relatively high PPV of 26.18%, which increased to 74.47% after confinement to BP users. Newly proposed ICD codes specific to BRONJ (M87.1, osteonecrosis due to drugs) exhibited very low PPVs. Only a few confirmed ONJ cases were captured by code M87.1, whereas most were captured by code K10.2. A mixed algorithm that set ICD code K10.2 as the main diagnosis and M87.1 as a secondary diagnosis exhibited a high PPV with a limited number of samples. Other attempts to increase PPV value were not effective.

A subgroup analysis of specific cancer patients is noteworthy to be discussed. Given that breast cancer, prostate cancer, and multiple myeloma are the main indications of high-dose IV BP use, a high PPV was expected. However, the PPV of this group was just slightly higher than those obtained from other subgrouping attempts to increase the PPV, and the value was never sufficient to be used as a significant surrogate for ONJ. Patients who had potential ONJ codes but were not confirmed as ONJ presented with various potential ICD codes. When the PPV was estimated according to each specific ICD code, K10.2 showed the highest PPV (44.05%, 95% CI 33.22–55.30). Non-ONJ patients with K10.2 presented with various inflammatory conditions of the jaw, such as chronic osteomyelitis, osteoradionecrosis, osteitis, unconfirmed bone tumors, and nonspecific radiographic lesions that could not be classified as stage 0 BRONJ, and etc. This indicates ICD codes were not sufficient to be significant surrogates, even in patients taking high doses of IV BPs; therefore, other reliable and specific operational definitions for identification of BRONJ are necessary.

These results of the present study indicate the limitations of current ICD algorithm-based identification of BRONJ cases in large-population epidemiologic studies and are consistent with the results of a similar recent investigation involving a Scandinavian cohort.32, 33 ICD code K10.2 has been globally utilized for the diagnosis of BRONJ. However, this database itself was constructed for reimbursement purposes rather than research purposes, and code K10.2 includes various clinical aspects of jaw pathology.7, 10, 17 Although a more specific code, M87.1, was introduced in 2010, the frequency of use has been recorded as very low. This might be attributed to the absence of site-specification of M87.1 to the jaw. Therefore, it is possible that dentists prefer to use other jaw-specific ICD codes. Moreover, the WHO has recently revised the disease classification as of 2016, suggesting the application of code K10.2 (osteonecrosis [drug-induced] [radiation-induced] of jaw [acute] [chronic] [suppurative]) for BRONJ, further complicating the identification of cases within large population-based databases. Currently, in the absence of specific ICD codes for BRONJ, it is recommended that the ICD code K10.2 be used in combination with M87.1 to improve PPV. Considering that the current algorithm did not sufficiently identify actual disease, interdisciplinary efforts to establish uniform ICD codes specific to BRONJ are essential. Moreover, implementation in all clinics that diagnose and treat ONJ should be carried out to warrant relevant uses in future epidemiological researches.

A fundamental question might be raised regarding an association between BP and ONJ, which was addressed in previous epidemiologic studies. The majority of included studies showed a tendency to identify their outcome surrogates themselves with the actual disease of ONJ. Considering the extremely low PPV, it is difficult to draw the conclusion that bisphosphonate is causally associated with ONJ based on the results of previous studies. Also, without manual confirmation of outcomes, conclusions should be limited to outcome surrogates but not BRONJ. Moreover, given the current definition of BRONJ in which the outcome is conditional on an exposure, cause-effect relationships cannot be properly established without bias. Because of such limitations of previous epidemiologic approaches hindering the establishment of the causal association, therefore, great care should be taken in interpretation of those results. Additional research is necessary to establish cause-effect relationships, such as investigations on dose-response relationships or threshold effects among BP users.

Despite such difficulties in conducting ONJ epidemiologic studies and the low validity of current claims-based algorithms, it is hard to deny that the large population-based approach may be one of the most effective and practical strategies to investigate such an uncommon disease. There is an urgent need to develop more reliable and specific operational definitions for the identification of BRONJ cases in large population databases. Furthermore, detailed information about BP administration should be collected for establishing dose-response relationships, and well-designed confounding controls should be taken into consideration in future ONJ epidemiologic studies.

Disclosures

All authors state that they have no conflicts of interest.