Methods

Publicly available FAERS data since January 1 2004 to September 30, 2021 was downloaded from FDA website as raw data. Hypoglycemic medications were mapped to the anatomic therapeutic classification (ATC) as A10 class. “Osteomyelitis” was defined as all of the adverse events (AEs) containing the key word “osteomyelitis”, which were determined by the Standardized MedDRA Query (SMQ, version 23.0) terminology [19]. Dataset of “Diabetes” was composed by all reports in FAERS with indication containing the key word “diabetes”.
Criteria of exclusion (Figure 1 ) were applied: each potential case was subjected to data cleaning procedure to make sure the removal of reports which were officially deleted by FDA authority, duplicated, missing caseid and date, or with inaccurate data for gender and age, and then was filtered with the targeted drug as the primary suspected (PS) drug. All the reports containing A10 medications other than the targeted one were removed, to minimize the possibility of interfering effects.
Data mining procedure using reporting odd ratio (ROR) method [21, 22] was introduced to investigate the disproportionality in reporting ratio caused by interested drug-AE pairs compared with random drug-AE pair, which were then tandem with Bayesian confidence propagation neural network (BCPNN) method introduced by Bate A et al. in 1998 [23], deducing linkage between the target drug and event by a prior possibility. The association between “diabetes” and AEs was also investigated. Drug-AE pairs which could generated stronger signals than the same AEs paired and diabetes were screened out and demonstrated with a heatmap. Osteomyelitis was picked as the major precursor to lower extremity amputation.
All the drugs and drug groups mentioned above were subjected to descriptive analysis for demographics, including gender, age category, annual report counts, occupation of the reporter, role of the targeted drug, and outcomes. Since hypoglycemic medications may sometimes be used by non-diabetic individuals or purpose [24, 25], meanwhile a considerable proportion of reports presents no specific indications or missing information, all interested drugs or drug groups were performed in duplicates with or without filtering diabetes as indication (Figure 1 ). Reports referring to competing interfering such as drugs known for causing osteomyelitis, including zoledronic acid, alendronate sodium were cast out, as well as reports listing osteology conditions as indications and adverse events (AE), since osteomyelitis may occurs preferentially in patients with diabetic ulcers, lower extremity amputation, metatarsal excision [17]. Because that osteomyelitis might occur preferentially in patients with known infection [18], we excluded reports containing competing indications and AEs which are typically reported preferentially among users of SGLT2i, in order to minimize the bias due to dilution or competition [26, 27], such as diabetic foot [28], infection [29], especially for genital infection, genitourinary tract infections (GUTI), urinary tract infection, diabetic ketoacidosis (DKA), and Fournier’s gangrene (FG), as well as reports that listing all the anti-biotics or becaplermin [30]. Furthermore, the using of insulin and its analogs (A10A) is typically considered a proxy of disease severity or advanced disease stage [26, 27], we category reports referring to A10A as a control group. In addition, the gender bias of the osteomyelitis reports was also put to the investigation. Reports referring to testosterone and estrogen were extract and underwent the same cleansing procedure described above.
The developing trend of ROR on quarterly basis was investigated. We designed a procedure to mimic the accumulation of FAERS data in real world by adding up every quarter of data into the dataset. Series of quarterly ROR (q-ROR) value was generated for interested drug (drug group)-osteomyelitis pairs. Chi-square tests (Chi2 tests) were induced to compare the changing tendencies of q-ROR of given pairs, as well as tendencies prior to and since SGLT2i were approved by FDA, to eliminate the interfering effect cause by comorbidities or concomitants.
Data process were conducted with R Studio 4.1.2 (RStudio), using logistic regression model. For ROR, a signal was determined as count of drug-AE pair (a) larger than 3, plus the value of the ROR higher than 1, as well as lower limit of the 95% confidence interval (95%CI) exceeding 1. For BCPNN, a signal was defined as the value of lower limit of information component (IC025) exceeding 0, which to be specific, IC025 value between 0 and 1.5 was defined as a weak signal, while IC025 between 1.5 and 3 was considered as a medium signal, and IC025> 3 was considered as a strong signal.