Risk of Upper GI Bleeding From Different Drug Combinations
Risk of Upper GI Bleeding From Different Drug Combinations
Data were obtained from a network of 7 electronic health record (EHR) databases from 3 countries. The EU-ADR Project (Exploring and Understanding Adverse Drug Reactions by integrative mining of clinical records and biomedical knowledge) has successfully established a platform that integrates data from various repositories of European EHRs for evaluation of drug safety.
We analyzed data from 3 primary care databases (Integrated Primary Care Information [IPCI, The Netherlands]; Health Search/CSD Longitudinal Patient Database [HSD, Italy]; and Pedianet [Italy]) and 4 administrative/claims databases (Aarhus University Hospital Database [Aarhus, Denmark], PHARMO Institute [PHARMO, The Netherlands], and the regional databases of Lombardy [UNIMIB, Italy] and Tuscany [ARS, Italy]). The characteristics and study periods of the databases are shown in Table 1. All of these databases have been extensively used in epidemiological studies. Subjects can enter and may also leave the database at any time for several reasons (eg, death, moving out of the region, leave of practice). The primary care databases capture all prescriptions from general practitioners and some from secondary care (eg, repeat prescriptions). The study protocol was approved by the review board for all databases.
The study population included all people registered in the database network with at least 1 year of valid and continuous data. A self-controlled case series (SCCS) analysis was performed on all identified cases of UGIB. The SCCS is a case-only study (ie, control subjects are not included) in which the relative incidence of UGIB is estimated for exposed and nonexposed time in each case. Each case serves as its own control. The SCCS method assumes that all cases in the analysis should (1) have exposed and unexposed person-time, (2) experience an UGIB, and (3) contribute follow-up time before and after the UGIB. The primary advantage of the SCCS is that it automatically adjusts for confounding factors that are fixed within subjects (ie, genetic factors, sex, chronic disease, or other comorbidity).
From the study population, we identified all subjects who experienced an UGIB during follow-up by using pertinent disease codes from the different coding systems in each database. UGIB was assessed by using hospital discharge codes (in claims databases) or general practitioner diagnosis/recordings (in primary care databases). We included all codes indicating gastroduodenal ulcers and hemorrhages, melena, and hematemesis. Codes for variceal bleeding specifically were not included. We only included codes corresponding to an acute UGIB, because for the SCCS the outcome should be an acute event with a clear disease onset. Supplementary Table 1 shows the corresponding codes for each coding system. A free-text search of clinical narratives was performed in IPCI and HSD. A validation study was conducted in 4 of the databases used in the current study and showed a high concordance for International Classification of Diseases (ICD)-9 (positive predictive value [PPV] of 78% and 72%) and ICD-10 codes (PPV of 77%) that was not seen with the International Classification for Primary Care coding system (PPV of 21% for codes and free text only).
We focused on concomitant use of nsNSAIDs, COX-2 inhibitors, and low-dose aspirin with other drugs reported to be associated with an increase or decrease in risk of UGIB. The drug groups of interest were as follows: (1) nsNSAIDs, (2) COX-2 inhibitors, (3) low-dose aspirin, (4) high-dose aspirin, (5) corticosteroids, (6) SSRIs (citalopram, fluoxetine, and paroxetine were assessed individually), (7) GPAs, (8) aldosterone antagonists, (9) calcium channel blockers, (10) anticoagulants, (11) antiplatelets, and (12) nitrates. Drugs of interest were categorized according to the World Health Organization's Anatomical Therapeutic Chemical (ATC) classification system.Supplementary Table 2 shows the corresponding ATC codes. We created mutually exclusive exposure categories: no use of any drug of interest (reference group), use of only one drug of interest, or concurrent use of nsNSAIDs, COX-2 inhibitors, or low-dose aspirin with one other drug of interest (Supplementary Figure 1). All other combinations of drugs of interest and combinations of >2 drugs were combined in a separate category. Fixed drug combinations were included in the corresponding drug combination group. Duration of exposure was calculated by dividing the total number of prescribed/dispensed pills by the number of pills per day or defined daily dosages. We assumed that all dispensed drugs were consumed. All exposed and unexposed person-time was therefore included in the analysis. Drug dose and frequency were not taken into account because such information is not consistently recorded in all databases.
(Enlarge Image)
Supplementary Figure 1.
Classification of (A) prescriptions and (B) drug categories.
To estimate the relative incidence of UGIB, incidence rate ratios (IRRs) with 95% confidence intervals (CIs) were obtained using conditional Poisson regression by comparing the incidence rate of UGIB during periods of drug exposure with the incidence rate during all other observed time periods. Age-adjusted IRRs were calculated within each database and by pooling all data together (IRRp). To account for heterogeneity between the databases, pooling of data was also performed by a random effects meta-analytic model on the database-specific risk estimates resulting in an overall IRR.
To estimate the magnitude of drug interaction (excess risk), the following measures were calculated: the relative excess risk due to interaction (RERI), the proportion attributable to interaction (AP), and the synergy index (S). Interaction on an additive scale meant that the observed effect of the drug combination was larger than the sum of the effects of the drugs separately but less than multiplicative. If the IRR of the combination was more than the sum of the 2 drugs separately, interaction (at least on an additive scale) was present. Corresponding 95% CIs were also calculated for the RERI using the Hosmer–Lemeshow delta method. The estimated measure of the RERI, AP, or S itself does not provide any information on risk and cannot be interpreted in isolation. However, based on the relative risk, it can be concluded that an excess risk is present when the RERI is larger than 0 and the CIs around it do not cross 0. Additionally, it may be concluded that there is more excess risk with a RERI of 1 than with a RERI of 2 (see Supplementary Table 3 for more details).
Population attributable risk (PAR) was calculated to estimate the proportion of UGIB in the general population that is attributable to concomitant use of drugs using the following formula: PAR = (p * [IRR − 1])/(p * [IRR − 1]+ 1). For this calculation, drug utilization data from the participating databases (data not shown) were used to derive the prevalence of exposure (p) to which the IRR pertained.
Because increasing age confers additional risk of UGIB, analyses by stratifying on age (with a cutoff of 60 and 70 years) and sex were conducted to investigate effect modification by age or sex. To explore the possibility of confounding by contraindication, we performed a sensitivity analysis by truncating the drug exposure at the time of the event. A pooled analysis excluding the IPCI database was performed due to the low PPV in IPCI.
Patients and Methods
Data Sources
Data were obtained from a network of 7 electronic health record (EHR) databases from 3 countries. The EU-ADR Project (Exploring and Understanding Adverse Drug Reactions by integrative mining of clinical records and biomedical knowledge) has successfully established a platform that integrates data from various repositories of European EHRs for evaluation of drug safety.
We analyzed data from 3 primary care databases (Integrated Primary Care Information [IPCI, The Netherlands]; Health Search/CSD Longitudinal Patient Database [HSD, Italy]; and Pedianet [Italy]) and 4 administrative/claims databases (Aarhus University Hospital Database [Aarhus, Denmark], PHARMO Institute [PHARMO, The Netherlands], and the regional databases of Lombardy [UNIMIB, Italy] and Tuscany [ARS, Italy]). The characteristics and study periods of the databases are shown in Table 1. All of these databases have been extensively used in epidemiological studies. Subjects can enter and may also leave the database at any time for several reasons (eg, death, moving out of the region, leave of practice). The primary care databases capture all prescriptions from general practitioners and some from secondary care (eg, repeat prescriptions). The study protocol was approved by the review board for all databases.
Study Design
The study population included all people registered in the database network with at least 1 year of valid and continuous data. A self-controlled case series (SCCS) analysis was performed on all identified cases of UGIB. The SCCS is a case-only study (ie, control subjects are not included) in which the relative incidence of UGIB is estimated for exposed and nonexposed time in each case. Each case serves as its own control. The SCCS method assumes that all cases in the analysis should (1) have exposed and unexposed person-time, (2) experience an UGIB, and (3) contribute follow-up time before and after the UGIB. The primary advantage of the SCCS is that it automatically adjusts for confounding factors that are fixed within subjects (ie, genetic factors, sex, chronic disease, or other comorbidity).
Case Definition
From the study population, we identified all subjects who experienced an UGIB during follow-up by using pertinent disease codes from the different coding systems in each database. UGIB was assessed by using hospital discharge codes (in claims databases) or general practitioner diagnosis/recordings (in primary care databases). We included all codes indicating gastroduodenal ulcers and hemorrhages, melena, and hematemesis. Codes for variceal bleeding specifically were not included. We only included codes corresponding to an acute UGIB, because for the SCCS the outcome should be an acute event with a clear disease onset. Supplementary Table 1 shows the corresponding codes for each coding system. A free-text search of clinical narratives was performed in IPCI and HSD. A validation study was conducted in 4 of the databases used in the current study and showed a high concordance for International Classification of Diseases (ICD)-9 (positive predictive value [PPV] of 78% and 72%) and ICD-10 codes (PPV of 77%) that was not seen with the International Classification for Primary Care coding system (PPV of 21% for codes and free text only).
Definition of Exposure
We focused on concomitant use of nsNSAIDs, COX-2 inhibitors, and low-dose aspirin with other drugs reported to be associated with an increase or decrease in risk of UGIB. The drug groups of interest were as follows: (1) nsNSAIDs, (2) COX-2 inhibitors, (3) low-dose aspirin, (4) high-dose aspirin, (5) corticosteroids, (6) SSRIs (citalopram, fluoxetine, and paroxetine were assessed individually), (7) GPAs, (8) aldosterone antagonists, (9) calcium channel blockers, (10) anticoagulants, (11) antiplatelets, and (12) nitrates. Drugs of interest were categorized according to the World Health Organization's Anatomical Therapeutic Chemical (ATC) classification system.Supplementary Table 2 shows the corresponding ATC codes. We created mutually exclusive exposure categories: no use of any drug of interest (reference group), use of only one drug of interest, or concurrent use of nsNSAIDs, COX-2 inhibitors, or low-dose aspirin with one other drug of interest (Supplementary Figure 1). All other combinations of drugs of interest and combinations of >2 drugs were combined in a separate category. Fixed drug combinations were included in the corresponding drug combination group. Duration of exposure was calculated by dividing the total number of prescribed/dispensed pills by the number of pills per day or defined daily dosages. We assumed that all dispensed drugs were consumed. All exposed and unexposed person-time was therefore included in the analysis. Drug dose and frequency were not taken into account because such information is not consistently recorded in all databases.
(Enlarge Image)
Supplementary Figure 1.
Classification of (A) prescriptions and (B) drug categories.
Main Statistical Analyses
To estimate the relative incidence of UGIB, incidence rate ratios (IRRs) with 95% confidence intervals (CIs) were obtained using conditional Poisson regression by comparing the incidence rate of UGIB during periods of drug exposure with the incidence rate during all other observed time periods. Age-adjusted IRRs were calculated within each database and by pooling all data together (IRRp). To account for heterogeneity between the databases, pooling of data was also performed by a random effects meta-analytic model on the database-specific risk estimates resulting in an overall IRR.
To estimate the magnitude of drug interaction (excess risk), the following measures were calculated: the relative excess risk due to interaction (RERI), the proportion attributable to interaction (AP), and the synergy index (S). Interaction on an additive scale meant that the observed effect of the drug combination was larger than the sum of the effects of the drugs separately but less than multiplicative. If the IRR of the combination was more than the sum of the 2 drugs separately, interaction (at least on an additive scale) was present. Corresponding 95% CIs were also calculated for the RERI using the Hosmer–Lemeshow delta method. The estimated measure of the RERI, AP, or S itself does not provide any information on risk and cannot be interpreted in isolation. However, based on the relative risk, it can be concluded that an excess risk is present when the RERI is larger than 0 and the CIs around it do not cross 0. Additionally, it may be concluded that there is more excess risk with a RERI of 1 than with a RERI of 2 (see Supplementary Table 3 for more details).
Population attributable risk (PAR) was calculated to estimate the proportion of UGIB in the general population that is attributable to concomitant use of drugs using the following formula: PAR = (p * [IRR − 1])/(p * [IRR − 1]+ 1). For this calculation, drug utilization data from the participating databases (data not shown) were used to derive the prevalence of exposure (p) to which the IRR pertained.
Sensitivity Analyses
Because increasing age confers additional risk of UGIB, analyses by stratifying on age (with a cutoff of 60 and 70 years) and sex were conducted to investigate effect modification by age or sex. To explore the possibility of confounding by contraindication, we performed a sensitivity analysis by truncating the drug exposure at the time of the event. A pooled analysis excluding the IPCI database was performed due to the low PPV in IPCI.
Source...