Risk Assessment for Anti-Infectives-Related Acute Kidney Injury Using the Japanese Adverse Drug Event Report Database

Background: Acute kidney injury (AKI) is associated with signicant increases in short- and long-term morbidity and mortality. Drug-induced AKI is a major concern in the present healthcare system. Our spontaneous reporting system (SRS) analysis assessed links between AKIs, along with patients’ age, as healthcare-associated risks and administered anti-infectives. We also generated anti-infectives-related AKI-onset proles. Method: We calculated adjusted reporting odds ratios (RORs) for reports of anti-infectives-related AKIs (per Medical Dictionary for Regulatory Activities) in the Japanese Adverse Drug Event Report database and evaluated associations between anti-infectives and age by association rule mining. We evaluated time-to-onset data and hazard types using the Weibull parameter. Results: Among 534,688 reports (submission period: April 2004–June 2018), there were 21,727 AKI events. Anti-infective treatments including glycopeptide antibacterials, uoroquinolones, third-generation cephalosporins, triazole derivatives, and carbapenems were associated with 596, 494, 341, 315, and 313 AKI incidences, respectively. Adjusted RORs of anti-infectives-related AKIs increased among older patients and were higher in anti-infective combination therapies [anti-infectives, ≥ 2; ROR, 2.75 (2.56–2.95)] than in monotherapies [ROR, 1.52 In association rule mining, the number of anti-infectives and age were associated with anti-infectives-related AKI lift values (as consequent). Moreover, 48.1% of AKIs occurred within 5 days (median, 5.0 days) of anti-infective therapy initiation. Conclusion: Thus, adjusted RORs derived from our new SRS analysis indicate potential AKI risks linked to age and number of administered anti-infectives.

derive an index for detecting drug-associated adverse events (AEs) [7,8]. We previously analyzed an SRS database and found that the combination of medications might increase the risk of AEs according to the index derived from the RORs [10][11][12]. In this study, we evaluated the anti-infectives-related AKI pro les using ROR.
Older patients often suffer from multiple diseases and receive several drugs, which is referred to as polypharmacy [13][14][15]. Polypharmacy is a well-known risk factor for AEs. Altered liver and kidney functions are considered a cause for changes in the pharmacokinetics in elderly patients [16]. Polypharmacy and the incidence of AEs increase with advancing age. By evaluating the adjusted RORs using a multivariate logistic regression analysis technique, Abe et al. showed that polypharmacy and age might be more closely linked to an increased risk of kidney disorder than liver disorder [10]. Association rule mining in large databases is a new analytical approach for evaluating association rules between variables [11,17]. Hatahira et al. applied this algorithm to assess the association rules among fall-related AEs, the number of administered drugs, and age in the JADER database [11]. Our study was focused on the association rules among anti-infectives-related AKI, the number of administered anti-infectives, and age. To our knowledge, time-to-onset pro les of anti-infectives-related AKI derived from SRS databases have been rarely reported. We evaluated the time-to-onset data of AKI relative to the anti-infective therapy initiation.

Methods
Ethical approval was not sought for this study because the study was an observational study without any research subjects. The JADER dataset can be downloaded from the website of the Pharmaceuticals and Medical Devices Agency (PMDA) (www.pmda.go.jp). The JADER database is publicly available. All data from the JADER database were fully anonymized by the PMDA before we used them. This study used a dataset containing information recorded between April 2004 and June 2018. The JADER database consists of four tables: 1) DEMO (patients' demographic information); 2) DRUG (drug information); 3) REAC (AE information); and 4) HIST (primary disease information). We built a relational database system that integrates the four data tables using the FileMaker Pro 14 software (FileMaker, Santa Clara, CA, USA). In the DRUG table, the causality of each drug was assigned a code depending on its association with adverse drug reactions, such as "suspected drug," "concomitant drug," or "interacting drug." All drugs in the "suspected drug," "concomitant drug," and "interacting drug" association classes were used for the analyses.
To identify an AE signal, we calculated the crude RORs by comparing one of the index groups with the reference group. Each ROR was calculated from a two-by-two contingency table; it is the ratio of the odds of reporting AEs versus all other events associated with the drug of interest compared with the reporting odds for all other drugs present in the database. RORs were expressed as point estimates with 95% con dence intervals (CIs). A ROR estimate of less than 1 indicated the absence of a potential exposureevent relationship, but if it was more than 1, it indicated a potential exposure-event safety signal. The signal of a drug-event combination was positively identi ed when the lower limit of the 95% CI of the ROR exceeded 1. The positive identi cation of a signal required two or more cases [7,18].
Using RORs allowed adjustments by multivariate logistic regression analysis and offered the advantage of controlling for covariates [6,[19][20][21][22]. We calculated the adjusted RORs according to previous reports [21,22]. To calculate adjusted RORs, only reports containing reporting year, sex, age, and the number of administered anti-infectives were used. The following multivariate logistic model was used for analysis: where Y is the reporting year, S is the sex, A is the age-strati ed group (≤ 19 years, 20-29 years, 30-39 years, 40-49 years, 50-59 years, 60-69 years, 70-79 years, 80-89 years, and ≥ 90 years), and N is the number of administered anti-infectives. The adjusted RORs were calculated using two reference groups, the female, 20-29-year-old group and the zero anti-infective group. A Wald test can be used to evaluate the effect of adding a speci c term. Because the difference in the − 2 log-likelihoods follows a chi-square distribution with one degree of freedom, adding an interaction term was statistically signi cant (p < 0.05).
To comparatively evaluate the effect of variables, we selected explanatory variables using a stepwise method [23] at a signi cance level of 0.05 (forward and backward, Table 2). The contributions of selected variables in the nal model were evaluated using the likelihood ratio test. A p ≤ 0.05 indicated statistical signi cance of the difference in the − 2 log-likelihoods following a chi-square distribution with 1 degree of freedom. Chi-Square is the likelihood-ratio chi-square test of the hypothesis that all regression parameters are zero.
Association rule mining is focused on nding frequent co-occurring associations among a collection of items. Given a set of transactions T (each transaction is a set of items), an association rule can be expressed as X [the antecedent of the rule (left-hand-side, lhs)] → Y [the consequent of the rule (righthand-side, rhs)], where X and Y are mutually exclusive sets of items [24]. The association rule was evaluated by measures of support, con dence, and lift. The support of the rule is de ned as the percentage of transactions in T that contain both X and Y [24,25]. The support was calculated as follows:

Support = P(X∩Y) = {X∩Y}/{D}
where D is the total number of transactions in the database.
The con dence of the association rule is the ratio of the support of the itemset X∩Y to the support of the itemset X, which roughly corresponds to the conditional probability P(Y|X) [23]. Because con dence is an indicator of the accuracy of related rules, an association rule with high con dence is critical. The formula for calculating con dence is as follows: The lift value is the ratio of the con dence of the rule and the expected con dence of the rule. It is de ned as follows: Lift evaluates the independence of X and Y, suggesting that the greater the lift value, the stronger the relationship. If X and Y are independent, the lift is 1. If X and Y are positively or negatively correlated, the lift is > 1 or < 1, respectively.
The Chi-squared value to evaluate the association rules is de ned using the values of con dence, support, and lift according to of the single rule [26, 27]: The association rule mining was performed using the apriori function of the arules library in the arules package of the R software (version 3.5.1) [28]. In the rst step, the apriori algorithm searched for itemsets that had more than minimum support as predetermined by the user [29,30]. In the second step, the rules were generated by selecting the itemsets that were based on a threshold of con dence from those found in the rst step. Because all possible rules were enumerated from a large database, the rst step was a narrow path. Therefore, to extract association rules e ciently, the thresholds of the minimum support and con dence were de ned depending on factors such as the size of data and the number of items. In this study, we de ned the minimum support and con dence thresholds as 0.0001 and 0.01, respectively. Furthermore, the maximum size of mined frequent itemsets (maxlen; maximum length of itemset/rule: a parameter in the arules package) was restricted to 3.
To assess the time-to-onset pro le, the median time from the rst prescription of each report to the onset of AKI was used in conjunction with the interquartile range and Weibull shape parameter (WSP). We selected an analysis period of 90 days after therapy initiation. We used the WSP test for the statistical analysis of time-to-onset data to describe a non-constant incidence rate of AEs. The WSP represents the failure rate distribution against time and was used to evaluate hazard functions for detecting AEs. The scale parameter α of the Weibull distribution determines the scale of the distribution function. A larger scale value (α) stretches the distribution whereas a smaller scale value shrinks the data distribution. The shape parameter β of the Weibull distribution determines the shape of the distribution function. The WSP β value indicates the hazard without a reference population; when β is equal to 1, the hazard is estimated to be constant over time. If β > 1, the hazard is considered to increase over time [31,32]. The information obtained from the WSP could be of complementary value for the pharmacovigilance analysis using ROR.
All data analyses were performed using JMP 12.0 (SAS Institute, Cary, NC, USA).

Results
The JADER database contains 534,688 reports submitted between April 2004 and June 2018, and we identi ed 21,727 AKI events. According to the Anatomical Therapeutic Chemical (ATC) Classi cation System (www.whocc.no/atc_ddd_index/), 145 anti-infectives were selected and categorized into 36 ATCdrug classes (Table 1).
In the top ve anti-infective therapies, glycopeptide antibacterials (ATC code: J01XA), uoroquinolones (ATC code: J01MA), third-generation cephalosporins (ATC code: J01DD), triazole derivatives (ATC code: J02AC), and carbapenems (ATC code: J01DH), we identi ed 596, 494, 341, 315, and 313 reported AKIassociated AEs, respectively ( Table 2). The lower limit of the 95% CI (con dence interval) of ROR was > 1 for the following drug groups: combinations of penicillins, incl. beta-lactamase inhibitors (ATC code: The adjusted RORs and 95% CIs are summarized in Table 3 The association rule mining technique was applied to AKI using demographic data, such as the age group, sex, and the number of anti-infectives administered (Table 4). We e ciently extracted the association rules by setting the optimized support thresholds and con dence thresholds to 0.0001 and 0.01, respectively, and limiting maxlen (parameter in the arules package) to 3. The result of the mining algorithm was a set of 36 rules: it is presented as the heat map of the lift and support derived from the strati ed age group, sex, and the number of administered anti-infectives (Table 4, Fig. 1). For AKI AEs caused by anti-infectives, combination therapies tended to have higher lift values than monotherapies; the lift values were 2.36 for combination therapies (≥ 2 anti-infectives) and 1.38 for anti-infective monotherapies ( Table 4, id. [6,15], Fig. 1). The lift values increased relative to the interaction between age and the number of administered anti-infectives (Table 4, id. [1][2][3][4], Fig. 1).
Combinations containing the complete information on the treatment start date and AE onset date were extracted for the time-to-onset analysis. We evaluated 14 anti-infectives categories for which the number of cases was more than 100 and the lower limit of the 95% CI exceeded 1 according to Table 2 (Table 5, Fig. 2). Figure 2 shows a histogram of the number of AKI onsets in relation to the number of days after anti-infective treatment initiation (from day 0 to day 90). The median period (interquartile range) until AKI onset caused by anti-infectives was 5.0 (2.0-11.0) days for orally (per os, po) administered antiinfectives and 5.0 (2.0-9.0) days for administration by intravenous (iv) injection. The upper limits of the 95% CI of the β value were less than 1 for po administered anti-infectives.

Discussion
AKI is a complication in clinical care that can be linked to a variety of anti-infectives. Signals indicating an association with AKI were detected in many categories of anti-infectives (Table 2). Polymyxins (ATC code: J01XB) had the highest crude ROR among 36 ATC-drug classes of anti-infectives ( Table 2). The detailed mechanism of AKI by polymyxin remains unclear [33]. In our study, 106 out of 107 reports on polymyxin-related AKI indicated colistin (ATC code: J01XB01) administration, which was associated with an AKI incidence rate of approximately 10-55% [34]. ROR signals were also associated with other antiinfectives. Aminoglycosides cause tubular cell toxicity, and vancomycin is linked to acute interstitial nephritis [34]. The incidence rate of nephrotoxicity is reportedly up to 58% in patients treated with aminoglycosides, but most recent reviews suggest rates of 5-15% [4]. All AKI reports in antimycotics were from amphotericin B. Amphotericin B causes AKI when used as monotherapy or combination therapy [4] and raises blood urea nitrogen (BUN) and serum creatinine in 80% of patients receiving a complete course of amphotericin B therapy [35].
The crude ROR values indicated the occurrence of AKI in anti-infectives-treated patients in the age ranges of 70-79 years, 80-89 years, and ≥ 90 years and in patients receiving anti-infective monotherapy or combination treatment (≥ 2 anti-infectives) ( Table 3). However, the crude ROR is insu cient for assessing the relative strength of causality between drugs and AEs and only provides an approximation of the signal strength [7,19]. The crude RORs were used to make adjustments by multivariate logistic regression analyses, which mitigated the effects of covariates. The adjusted RORs of anti-infectivesrelated AKI increased with the age group (Table 3). Furthermore, the adjusted RORs tended to be higher in anti-infective combination therapy than in monotherapy; the adjusted ROR was 2.75 (2.56-2.95) for ≥ 2 anti-infectives and 1.52 (1.45-1.61) for one anti-infectives. These results suggested that the patient age and the number of administered anti-infectives are related to the occurrence of AKI. Accordingly, our study further demonstrated that the number of administered anti-infectives and age were both associated with the lift value of anti-infectives-related AKI (as the consequent) ( Table 4, Fig. 1). Aging is known to decrease renal drug elimination [36], which is associated with an increased risk of AKI by high drug exposure in the elderly. Rybak et al. reported AKI incidence rates of 5%, 11%, and 22% in patients treated with vancomycin monotherapy, aminoglycoside monotherapy, and combination therapy consisting of vancomycin and one aminoglycoside, respectively [37]. Thus, anti-infective combination therapy may increase the risk of AKI in older patients, which should be considered more carefully in clinical practice.
The time-to-onset analysis derived the daily numbers of onset events. We found that 48.1% of antiinfectives-related AKI occurred within 5 days of treatment initiation, and the median for anti-infectivesrelated AKI onset was 5.0 days post-initiation (Table 5, Fig. 2). We did not detect statistically signi cant differences of time-to-onset pro les among the different types of anti-infectives (14 ATC-drug classes) or the route of administration (iv versus po).
There are inherent limitations in using SRS data. For example, the length of the post-launch period of the drug, the noti cation of AEs, over-reporting, and under-reporting affect SRS analysis. There was no suitable comparison group, and data on patient characteristics were incomplete. Multivariate logistic regression analysis was used to adjust potential data bias. The results of this study were partially re ned using the multivariate logistic regression analysis technique. Therefore, covariates-corrected adjusted RORs are likely to have improved odds accuracy compared to that of regular RORs. It has been reported that angiotensin-converting enzyme inhibitors, angiotensin receptor blockers, nonsteroidal antiin ammatory drugs, calcineurine inhibitor (cyclosporine, tacrolimus), sulfonamides, acyclovir, rifampin, phenytoin, interferon, and proton pump inhibitors are involved in AKI [5]. In this study, we did not evaluate the effect of concomitant drugs other than anti-infectives. More reliable epidemiological studies will be needed to derive the causal constraints from this analysis.

Conclusions
The JADER database, which includes clinicians' reports of potential AE concerns related to drugs, is a useful tool for pharmacovigilance because it is based on real-world data derived from clinical practice. To our knowledge, this is the rst study to report the association between anti-infectives and AKI using SRS. We used adjusted RORs to identify the risk of anti-infectives-related AKI linked to older patients and the number of administered anti-infectives. Based on association rule mining technique, the number of administered anti-infectives and patient age were both associated with the lift value of anti-infectivesrelated AKI. The median period until anti-infectives-related AKI onset was 5 days after therapy initiation. We believe that our data will provide guidance for reducing the incidence of AEs in elderly patients receiving polypharmacy.
Abbreviations AKI: Acute kidney injury; SRS: Spontaneous reporting systems; JADER: Japanese Adverse Drug Event Report; PMDA: Pharmaceuticals and Medical Devices Agency; MedDRA: Medical Dictionary for Regulatory Activities; WSP: Weibull shape parameter Declarations Ethics approval and consent to participate Ethical approval was not sought for this study because the study was an observational study without any research subjects. All results were obtained from data openly available online from the PMDA website. All data from the JADER database were fully anonymized by the regulatory authority before we accessed them.

Consent for publication
Not applicable.

Competing interests
Ryogo Umetsu is an employee of Micron Inc. The rest of authors have no con ict of interest.

Funding
This research was partially supported by JSPS KAKENHI Grant Number, 17K08452.

Authors' contributions
All authors have contributed to this scienti c work and approved the nal version of the manuscript. SN, SH, and MN designed this study, performed the data analyses, and wrote the manuscript. RU, and YN involved in methodology and software. KS, RM, MT, KM, YY, MI, and RS assisted the data curation and validation. JL supervised the drafting of the manuscript. All authors took responsibility for the integrity of the data and accuracy of the data analysis.

Availability of data and materials
The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.