The central problem was the way non-jaw osteonecrosis appeared in the reporting vocabulary. FAERS has no preferred term for avascular necrosis, so extra-oral cases are often recorded under the site-unspecified term “osteonecrosis.” That term is also used for osteonecrosis of the jaw, a complication closely associated with bisphosphonates and denosumab. The researchers therefore constructed cases at the report level, beginning with explicit non-jaw terms and adding site-unspecified reports only when they lacked evidence of jaw involvement, such as an oral or dental co-event or an indication referring to an oral malignancy or dental procedure. This adjudicated approach identified 24,087 cases, compared with 1,521 under a conservative definition limited to explicit non-jaw terms and 33,964 under a more permissive definition. A further 73 reports were removed because the suspected drug appeared to have been prescribed for osteonecrosis itself.

An initial scan of 432 suspect-drug ingredients produced 47 signals under four complementary disproportionality criteria: reporting odds ratio, proportional reporting ratio with chi-square testing, the multi-item gamma-Poisson shrinker, and the Bayesian confidence propagation neural network. A new descriptive measure, the jaw-propensity index, calculated the share of each drug’s osteonecrosis-family reports explicitly coded to the jaw. Six drugs with established jaw-centered osteonecrosis clustered at the high end: denosumab had an index of 0.94, risedronate 0.78, zoledronic acid 0.71, ibandronate 0.65, pamidronate 0.55 and alendronate 0.502. By contrast, tenofovir disoproxil had an index of 0.001, and the glucocorticoid signals ranged from 0 to 0.28. The final glucocorticoid-and-asparaginase core did not change classification when the threshold was varied from 0.30 to 0.70.

Jaw exclusion removed 41 of the 47 signals, but the study also found that litigation-related reporting could independently magnify associations. Restricting the analysis to health-professional reports sharply reduced the tenofovir disoproxil reporting odds ratio from 157.9, with a 95% confidence interval of 153.4 to 162.6, to 6.33, with an interval of 5.38 to 7.45. Of 7,027 US tenofovir disoproxil osteonecrosis cases, 5,937, or 84.5%, carried a lawyer reporter code. Emtricitabine and efavirenz showed similar patterns. Lawyer-coded reports represented 25.8% of all adjudicated cases but only 2.6% of the deduplicated FAERS corpus. The researchers treated the lawyer code as a lower bound on solicitation, because recruited reports may also enter through consumers or manufacturers with no occupation recorded.

Independent comparisons supported the view that the extreme US antiretroviral estimates reflected reporting structure rather than a comparable signal in every jurisdiction. Canada Vigilance produced a tenofovir disoproxil reporting odds ratio of 6.24, with a wide 95% interval of 2.00 to 19.4, based on only three cases. JADER produced 10.5, with an interval of 6.60 to 16.8, from 18 cases. Both were far below the all-reporter FAERS estimate and closer to its health-professional estimate. Within reports carrying an HIV indication, the Japanese estimate was 2.75, with an interval of 1.26 to 6.00, whereas the corresponding US estimate remained 27.5, with an interval of 25.2 to 30.1, because 94% of US HIV-indication cases named tenofovir disoproxil or a co-formulated partner. These findings do not exclude a modest tenofovir–osteonecrosis association, but they do not support the magnitude seen in the unfiltered US data.

The external databases more consistently reproduced the glucocorticoid pattern. The pooled glucocorticoid reporting odds ratio was 10.3, with a 95% interval of 8.28 to 12.8, in Canada and 54.4, with an interval of 48.3 to 61.3, in Japan. In JADER, individual estimates included 39.3 for prednisolone, based on 729 cases, 28.7 for methylprednisolone, based on 207 cases, and 13.0 for dexamethasone, based on 184 cases. Asparaginase also replicated strongly in Japan, with a reporting odds ratio of 75.7, an interval of 61.4 to 93.3 and 103 cases. Rankings of drugs by jaw-propensity index were highly concordant across databases, with Pearson correlations of 0.985 for Canada and 0.960 for Japan, although this comparison was supported largely by separation into jaw-dominated and non-jaw clusters rather than by identical values for every drug.

Restricting each non-glucocorticoid signal to the disease for which the drug was prescribed separated additional drug-like signals from indicators of an underlying condition. Mycophenolate mofetil initially had a reporting odds ratio of 2.58, with an interval of 2.21 to 3.01, and remained elevated when only health professionals were considered and when glucocorticoid-unexposed reports were examined. Yet within transplant or immunosuppression indications, its ratio fell to 0.89, with an interval of 0.68 to 1.16, leaving no detected excess beyond the disease baseline. Voxelotor similarly fell from 5.06 overall to 0.97, with an interval of 0.71 to 1.32, within sickle-cell disease. Among 20 non-glucocorticoid drugs evaluated within their home indications, 14 had no residual excess, five met the study’s residual criterion, and one was not evaluable. The five residuals were asparaginase, pegaspargase, testosterone, arsenic trioxide and imiglucerase.

Asparaginase had the strongest case because it combined statistical, mechanistic and cross-jurisdiction evidence. Its within-leukemia reporting odds ratio was 7.14, with a 95% interval of 5.10 to 10.0, while pegaspargase had a lower but directionally consistent ratio of 2.00, with an interval of 1.26 to 3.17. The source article describes mechanisms through which asparaginase can promote hypercoagulability and hypertriglyceridemia, both established osteonecrosis risk factors, and notes experimental evidence that it can potentiate glucocorticoid-induced osteonecrosis. By comparison, the residuals for testosterone in hypogonadism, arsenic trioxide in acute promyelocytic leukemia and imiglucerase in Gaucher disease were modest and could plausibly reflect treatment of patients with more severe disease, greater exposure to co-therapies or unmeasured differences in severity and duration. The spontaneous reports did not contain enough information to distinguish these explanations from a drug contribution.

Timing provided a further separation between the major signal clusters. Weibull models gave glucocorticoids a subacute onset profile, with shape parameters from 0.44 to 0.74 and fitted median times to onset of 71 to 265 days. Several antiresorptives had increasing-hazard profiles, consistent with delayed jaw osteonecrosis, although the antiresorptive class as a whole was heterogeneous. The litigation-heavy antiretrovirals had especially delayed patterns: tenofovir disoproxil had a fitted median of 1,292 days, emtricitabine 1,244 days and efavirenz 1,438 days. A separate machine-learning check, which did not receive the disproportionality results, also ranked consumer reporting, lawyer reporting and glucocorticoid co-exposure as the leading predictors of case status. Its area under the receiver-operating-characteristic curve was 0.765. This exercise used the same reports, so it was an internal consistency check rather than independent validation.

The study’s conclusions are constrained by the nature of spontaneous reporting. Missing concomitant medications can hide glucocorticoid exposure; adverse events are under-reported; notoriety can increase reporting; and large clusters of reports can depress the apparent ratios of less frequently reported drugs. Indeed, all 34 negative-control drugs had ratios below 1, with a median of 0.16. Removing antiresorptive and litigated-antiretroviral reports raised that median to 0.38, and removing glucocorticoid reports as well raised it to 0.44, while most remaining drug ratios increased by roughly 2.35-fold. The analysis also used external case definitions that only approximated the primary adjudication, and the secondary analyses were exploratory rather than multiplicity-adjusted. The researchers concluded that, after accounting for jaw misclassification, litigation-related reporting and confounding by indication, the reproducible spectrum is glucocorticoid-centered, with asparaginase as the best-supported residual. Three additional drugs remain unresolved, and cohort studies would be required to determine whether the residual associations reflect causation.

Subject of Research: Pharmacovigilance analysis of drug reporting signals for non-jaw osteonecrosis

Article Title: Glucocorticoid-centered drug spectrum of non-jaw osteonecrosis, a FAERS pharmacovigilance study with independent validation in Canada Vigilance and JADER

Article References: Qi, L., Qiu, Y., Wang, X., Zhang, T., & Jiang, H. (2026). Glucocorticoid-centered drug spectrum of non-jaw osteonecrosis, a FAERS pharmacovigilance study with independent validation in Canada Vigilance and JADER. Frontiers in Pharmacology, 17, Article 1977191. https://doi.org/10.3389/fphar.2026.1977191

Image Credits: AI Generated

DOI: 10.3389/fphar.2026.1977191

Keywords: non-jaw osteonecrosis, FAERS, pharmacovigilance, glucocorticoids, asparaginase, reporting bias, jaw osteonecrosis, adverse events, disproportionality analysis, Canada Vigilance, JADER, Glucocorticoid-centered