Warfarin, known as a vitamin K antagonist, inhibits coagulation factors II, VII, IX, and X and non-specifically inhibits the intrinsic and extrinsic pathways of the blood coagulation process.¹ It is difficult to predict the drug concentration of warfarin due to different metabolic processes or serum protein binding ability in individual patients, requiring frequent blood tests to adjust drug concentration.¹ Direct oral anticoagulants (DOACs) directly inhibit thrombin or coagulation factor Xa and are convenient due to a fixed dose without monitoring.^2,3 Previous meta-analysis and population-based cohort studies reported that DOAC users had higher risks of gastrointestinal bleeding (GIB) compared with warfarin users.^2,3 This issue was even more relevant as DOAC has no specific antidotes for anticoagulation reversal.

However, recent real-world publications substantially downsized the risk of GIB associated with DOAC.^4-6 A recent systematic review and meta-analysis of 37 real-world publications showed no significant difference in the major GIB between DOAC and conventional treatment, such as warfarin or anti-platelet agents (adjusted hazard ratio [HR], 1.02; 95% confidence interval [CI], 0.94 to 1.10).⁶ A recent review article also reported that DOAC does not raise the risk of major GIB compared to warfarin.⁷ Due to the known risk of GIB in DOAC users, they are often prescribed together with acid suppressants, such as proton pump inhibitors, in a real-world practice. When DOAC was used with proton pump inhibitors, DOAC showed a lower risk of upper GIB and mortality than warfarin.⁸ A systematic review and meta-analysis also reported that acid suppressants significantly reduced the risk of GIB in DOAC users, with an overall relative risk of 0.70 (95% CI, 0.61 to 0.82).⁹ Therefore, the risk of GIB associated with DOAC use in real-world practice may be lower than those initially reported.

This study aimed to quantify the comparative risk of GIB in new DOAC and warfarin users in real-world practice based on a common data model (CDM).

1. Concept of CDM

Most hospitals in South Korea use electronic health record (EHR) systems; however, a large number of medical codes for diagnosis, medications, and procedures are incompatible with international coding systems. Therefore, CDM concept has been introduced for research using the same analysis algorithm among numerous EHRs, as CDM has standardized database with the same frame through distributed research networks.^10-13 In CDM research, analysis tool works in the local environment within a single hospital to ensure the security of the individual identification information, which means that an analysis program code is sent to each institution and only the statistical results analyzed within the institution are collected, and the patient data do not leave the institution.^10-12 In CDM database, the meaning of each content is represented with standard concepts, and content-related concepts are stored with their concept-ids as foreign keys to the concept table. The CDM framework has three important modules: (1) mapping local vocabularies to standard concepts; (2) extraction, transformation and loading of EHR data into CDM data; and (3) developing a general analysis interface.^10,11 When the data of patients are loaded into the CDM, the original “patient_ids” are deleted, and the random “person_ids” are generated for CDM data to maintain personal security, privacy and confidentiality. Since source variable names are generally expressed in nonstandard terms, these terms were standardized into standard concepts with mapping, which is crucial for facilitating patient data standardization.^10,11 The quality of CDM data is excellent as CDM is basically based on the same structure and standardized definition. The quality issues of CDM data was already described and verified in many previous studies.^12,14,15

2. Data source and ethics

This study was based on EHR data of each hospital converted to CDM from eight hospitals, including Kyung Hee University Hospital at Gangdong (n=887,370), Ajou University Medical Center (n=2,765,795), Daegu Catholic University Medical Center (n=958,429), Kangdong Sacred Heart Hospital (n=1,194,685), Kyung Hee University Medical Center (n=1,168,640), Kangwon National University Hospital (n=567,439), Pusan National University Hospital (n=1,753,001), and Soonchunhyang University Hospital (n=1,098,041). The number of participants, study period, DOAC users, and warfarin users before and after propensity score matching (PSM) are presented in Table 1. We used Athena (https://athena.ohdsi.org/) as a CDM vocabulary browser to find the matching standard concept in CDM database for concept mapping. To minimize mapping errors and data loss, two authors (J.M.C. and M.K.) carefully checked the concept mappings. Our institutional review board approved this study (IRB number: KHNMC 2023–09-008). Informed consent was waived for this study as it was based on the CDM database, which has no data security issues.

Table 1 . Summary of Data Source.

Database	Total number of participants	Study period	Before propensity score matching			After propensity score matching
			Warfarin users	DOAC users		Warfarin users	DOAC users
Kyung Hee University Hospital at Gangdong	887,370	2006/06–2023/03	980	2,309		358	358
Ajou University Medical Center	2,765,795	1994/01–2023/04	2,802	4,947		899	899
Daegu Catholic University Medical Center	958,429	2005/01–2023/05	1,812	3,561		596	596
Kangdong Sacred Heart Hospital	1,194,685	1986/11–2022/09	531	1,971		229	229
Kyung Hee University Medical Center	1,168,640	2008/01–2022/02	889	3,055		311	311
Kangwon National University Hospital	567,439	2003/01–2022/01	799	2,177		297	297
Pusan National University Hospital	1,753,001	2011/02–2019/08	1,653	2,857		520	520
Soonchunhyang University Hospital	1,098,041	2003/05– 2021/05	352	2,032		137	137
Total			9,818	22,909		3,347	3,347

DOAC, direct oral anticoagulant..

3. Study design and cohort definition

We designed a retrospective, observational, and comparative cohort study of new DOAC and new warfarin users aged >18 years (Fig. 1). New DOAC users (target cohort) were defined as adults prescribed DOACs (dabigatran, rivaroxaban, edoxaban, or apixaban) for >90 consecutive days. New warfarin users (comparative cohort) were defined as users newly prescribed warfarin for >90 consecutive days. Any patients with the following criteria were excluded from the cohort: (1) history of GIB before cohort entry; (2) history of previous study drug use; and (3) age <18 years. We also excluded patients with chronic liver disease (concept_id 4212540), chronic kidney disease (concept_id 46271022), renal dialysis (concept_id 4146536), and kidney transplantation (concept_id 4322471). All DOACs and warfarin on the South Korean market were included in this analysis. Patients were considered eligible when they were continuously exposed to study medications from the index date for their prescription period until development of GIB, follow-up loss from the hospital, or termination of prescription. The target cohort excluded any warfarin users after DOAC exposure, and the comparative cohort also excluded DOAC users after warfarin exposure. Both cohorts were censored or terminated when GIB was identified. In this study, the “time-at-risk” was defined from 1 day after the index date to the end of the observation, during which the patients were followed up and the longest observation period for any patient was 455 days. The end of observation of a patient included the following criteria: no follow-up prescription of the medications that was first prescribed at the index date 180 days after the last prescription, any prescription of opposite medications that was initially prescribed at the index date, death of the patient, or end of data availability. To validate our study, the time-at-risk end was also evaluated with 455 days after cohort start and with 30 days after cohort end.

Figure 1. The study flowchart of the included patient-based retrospective cohort data from eight hospitals. Patients included in both cohorts with a history of gastrointestinal bleeding who did not have at least 1 day at risk and were not matched on propensity score were excluded. Finally, 3,347 propensity-matched pairs between the direct oral anticoagulants and warfarin users were included.

GIB (outcome cohort) was defined as any patients diagnosed with gastrointestinal hemorrhage (concept_id 192671, 4100660), upper GIB (concept_id 4291649, 193250, 4318535) or lower GIB (concept_id 4338544, 4318536, 4318829), and/or treated with endoscopic control of bleeding (concept_id 2109184, 2108900, 44784306, 2109100), identified by SNOMED-CT codes for matched discharge diagnoses. To compare the risk of GIB, we matched the target and comparative cohorts using 1:1 PSM, with age group, sex, and comorbidity score as fixed independent variables. We attempted to match each patient in both cohorts with a similar propensity score based on nearest-neighbor matching without replacement. The assessment of imbalance between baseline characteristics after matching was measured with standardized mean difference, and <10% were considered acceptable.¹³

4. Covariates

A total of 237 covariates were used for extensive PSM between the new DOAC and new warfarin users, including age, sex, index year, Charlson comorbidity index, comorbidities, and drugs prescribed 365 days before the index date, with regularized logistic regression models (Table 2).^10-13 We reported covariates over 5% of the total patients before PSM. In this analysis, the pharmacological variables were as follows: agents acting on the renin-angiotensin system, antibacterials for systemic use, antidepressants, anti-inflammatory and antirheumatic products, antithrombotic agents, beta-blockers, calcium channel blockers, diuretics, drugs for acid-related disorders, drugs for obstructive airway diseases, drugs used in diabetes, lipid-modifying agents, opioids, or psycholeptics. General medical history included chronic obstructive lung disease, gastroesophageal reflux disease, diabetes mellitus, hypertensive disorder, hyperlipidemia, pneumonia, or neoplasm. Cardiovascular disease history included atrial fibrillation, coronary arteriosclerosis, cerebrovascular disease, heart failure, heart disease, ischemic heart disease, pulmonary embolism, or venous thrombosis. Charlson comorbidity score and CHADS² were used to assess the overall comorbidity burden.

Table 2 . Distribution of Baseline Characteristics in the Overall Population from the Five Hospitals between Warfarin and DOAC Users before and after Propensity Score Matching.

Characteristic	Before propensity score matching, No. (%)				After propensity score matching, No. (%)
	Warfarin (n=9,818)	DOAC (n=22,909)	SMD		Warfarin (n=3,347)	DOAC (n=3,347)	SMD
Age group, yr
<40	496 (5.1)	552 (2.4)	0.140		176 (5.3)	132 (3.9)	0.063
40–49	936 (9.5)	891 (3.9)	0.227		259 (7.7)	210 (6.3)	0.057
50–59	1,819 (18.5)	2,372 (10.4)	0.234		575 (17.2)	535 (16.0)	0.032
60–69	2,435 (24.8)	5,367 (23.4)	0.032		849 (25.4)	804 (24.0)	0.031
70–79	2,971 (30.3)	8,169 (35.7)	0.115		1,008 (30.1)	1,116 (33.4)	0.069
≥80	1,161 (11.8)	5,558 (24.3)	0.328		480 (14.3)	550 (16.4)	0.058
Female sex	4,026 (41.0)	10,833 (47.3)	0.127		1,363 (40.7)	1,433 (42.8)	0.042
General medical history
Hypertensive disorder	3,673 (37.4)	9,019 (39.4)	0.040		1,248 (37.3)	1,230 (36.8)	0.011
Hyperlipidemia	1,585 (16.1)	4,873 (21.3)	0.132		572 (17.1)	574 (17.2)	0.002
Diabetes mellitus	1,530 (15.6)	3,446 (15.0)	0.015		486 (14.5)	504 (15.1)	0.015
Gastroesophageal reflux disease	482 (4.9)	1,934 (8.4)	0.142		182 (5.4)	185 (5.5)	0.004
Pneumonia	472 (4.8)	1,470 (6.4)	0.070		171 (5.1)	170 (5.1)	0.001
Chronic obstructive lung disease	569 (5.8)	1,069 (4.7)	0.051		158 (4.7)	137 (4.1)	0.031
Neoplasm history	1,100 (11.2)	4,490 (19.6)	0.234		485 (14.5)	527 (15.8)	0.035
Cardiovascular disease history
Heart disease	6,537 (66.6)	15,165 (66.2)	0.008		2,086 (62.3)	2,009 (60.0)	0.047
Atrial fibrillation	3,875 (39.5)	8,811 (38.5)	0.021		1,204 (36.0)	1,236 (36.9)	0.020
Ischemic heart disease	1,391 (14.2)	3,048 (13.3)	0.025		432 (12.9)	477 (14.3)	0.039
Venous thrombosis	722 (7.4)	3,068 (13.4)	0.199		360 (10.8)	407 (12.2)	0.044
Pulmonary embolism	678 (6.9)	3,024 (13.2)	0.211		307 (9.2)	353 (10.5)	0.046
Cerebrovascular disease	900 (9.2)	2,149 (9.4)	0.007		307 (9.2)	298 (8.9)	0.009
Coronary arteriosclerosis	427 (4.4)	785 (3.4)	0.048		122 (3.6)	144 (4.3)	0.034
Medication use
Antithrombotic agents	7,560 (77.0)	14,271 (62.3)	0.324		2,272 (67.9)	2,400 (71.7)	0.083
Drugs for acid-related disorders	6,295 (64.1)	14,605 (63.8)	0.007		1,971 (58.9)	2,081 (62.2)	0.067
Anti-inflammatory and antirheumatic products	5,133 (52.3)	10,766 (47.0)	0.106		1,607 (48.0)	1,667 (49.8)	0.036
Beta blocking agents	4,717 (48.1)	10,830 (47.3)	0.015		1,523 (45.5)	1,491 (44.6)	0.019
Diuretics	4,840 (49.3)	10,168 (44.4)	0.099		1,430 (42.7)	1,526 (45.6)	0.058
Lipid modifying agents	3,886 (39.6)	9,913 (43.3)	0.075		1,352 (40.4)	1,329 (39.7)	0.014
Agents acting on the RAS	4,161 (42.4)	9,052 (39.5)	0.058		1,325 (39.6)	1,317 (39.4)	0.005
Opioids	4,211 (42.9)	8,983 (39.2)	0.075		1,271 (38.0)	1,318 (39.4)	0.029
Antibacterials for systemic use	4,165 (42.4)	9,331 (40.7)	0.034		1,238 (37.0)	1,321 (39.5)	0.051
Calcium channel blockers	3,938 (40.1)	8,863 (38.7)	0.029		1,193 (35.7)	1,193 (35.7)	<0.001
Psycholeptics	3,516 (35.8)	7,380 (32.2)	0.076		1,003 (30.0)	1,049 (31.4)	0.030
Antidepressants	2,357 (24.0)	5,599 (24.4)	0.010		724 (21.6)	780 (23.3)	0.040
Drugs for obstructive airway diseases	2,424 (24.7)	5,791 (25.3)	0.014		735 (22.0)	743 (22.2)	0.006
Drugs used in diabetes	2,319 (23.6)	4,605 (20.1)	0.085		598 (17.9)	671 (20.1)	0.056

DOAC, direct oral anticoagulant; SMD, standardized mean difference; RAS, renin-angiotensin system..

5. Statistical analysis

We used the health big-data platform based on CDM supported by the Korean National Project.¹² Categorical variables were presented as numbers (percentage) and normally distributed continuous variables as mean (standard deviation). The tools for CDM analysis embedded in the ATLAS platform (version 2.12.0) were used in the initial analysis. R packages, version 4.3.1 (R Foundation for Statistical Computing, Vienna, Austria), were used to support the Cox model analysis and Kaplan-Meier estimation. R program supports the comparative analysis functions in the CDM, including creating the analysis dataset, constructing the baseline demographics study, and building the Cox regression analysis. Cox proportional hazard models were used to compare GIB in the matched cohorts. The HR and 95% CIs for GIB were calculated. To test the statistical significance of the differences between the observed cohorts, the Kaplan-Meier curves using log-rank tests were depicted for the percentage of event-free patients. Only the first event was included in all time-to-event analyses. Statistical significance was defined as two-sided p-values <0.05. We used 0.2 of the pooled standard deviation of the logit of the propensity score as the caliper width for PSM. Statistical heterogeneity was assessed using chi-square and I² statistics. A fixed-effects model was used when the heterogeneity (p<0.05, I² >50%) is absent; otherwise, a random-effect model was used.

A total of 9,818 new DOAC users and 22,909 new warfarin users from eight hospitals met the eligible criteria before PSM. Patients included in both cohorts who had a history of GIB, were not at least 1 day at risk, and did not match the propensity score were excluded (Fig. 1). Finally, 3,347 propensity-matched pairs of new DOAC and new warfarin users were included. Standardized mean differences were lower than 0.1 after PSM regarding age group, sex, medical history, cardiovascular disease history, and medication use.

1. Baseline characteristics of DOAC and warfarin users

Table 2 shows the baseline characteristics of the study population. After PSM, the proportion of new warfarin and DOAC users consistently increased up to 70–79 years; however, it decreased for patients ≥80 years. After PSM between DOAC and warfarin users, the most common general medical history finding was hypertensive disorders, and followed by hyperlipidemia and diabetes mellitus. The most common history of cardiovascular disease was heart disease, followed by atrial fibrillation and ischemic heart disease. In our study, any medications that may modify the risk of GIB, such as antithrombotic agents, anti-inflammatory drugs, antirheumatic drugs, and drugs for acid-related disorders, were well matched between new DOAC and new warfarin users (Table 2). The most common medications used before cohort entry were antithrombotic agents, followed by drugs for acid-related disorders (58.9% for warfarin users and 62.2% for DOAC users).

2. Risk of GIB between DOAC and warfarin users

We conducted Cox proportional hazard analyses to compare the risk of GIB between new DOAC and warfarin users after PSM (Fig. 2). The risk of GIB was not different between both cohorts (HR, 0.95; 95% CI, 0.65 to 1.40; p=0.808). When the time-at-risk end was reevaluated with 455 days after cohort start, the risk of GIB was not different between both cohorts (HR, 0.83; 95% CI, 0.60 to 1.15; p=0.256) (Supplementary Fig. 1A). When the time-at-risk end was modified as 30 days after cohort end, the risk of GIB was not different between both cohorts, either (HR, 0.99; 95% CI, 0.78 to 1.26; p=0.947) (Supplementary Fig. 1B).

Figure 2. Kaplan-Meier plots for the gastrointestinal bleeding between direct oral anticoagulant (DOAC) and warfarin users. The difference between new DOAC and new warfarin users was not statistically significant (hazard ratio, 0.79; 95% confidence interval, 0.50–1.25; p=0.319).

3. Risk of GIB between DOAC and warfarin users in older patients

Fig. 3 shows the Kaplan-Meier plots for GIB between DOAC and warfarin users in older patients. New DOAC users were not associated with GIB than new warfarin users in older subjects >65 years (HR, 1.00; 95% CI, 0.69 to 1.52; p=0.997) (Fig. 3A) and in older subjects >75 years (HR, 1.21; 95% CI, 0.68 to 2.10; p=0.509) (Fig. 3B).

Figure 3. Kaplan-Meier plots for the gastrointestinal bleeding between direct oral anticoagulant (DOAC) and warfarin users in older patients (A: older adults >65 years and B: older adults >75 years). There was no statistical significance for the difference between new DOAC and new warfarin users in older patients >65 years (hazard ratio [HR], 1.00; 95% confidence interval [CI], 0.69–1.52; p=0.997) as well as in older patients >75 years (HR, 1.21; 95% CI, 0.68–2.10; p=0.509).

4. Risk of GIB between DOAC and warfarin users according to sex

Fig. 4 shows the Kaplan-Meier plots for GIB between new DOAC and new warfarin users according to sex. Compared with new warfarin users, new DOAC users had similar risks of GIB in the male cohort (HR, 0.84; 95% CI, 0.49 to 1.44, p=0.524) (Fig. 4A) and female cohort (HR, 0.79; 95% CI, 0.42 to 1.49; p=0.461) (Fig. 4B).

Figure 4. Kaplan-Meier plots for the gastrointestinal bleeding between direct oral anticoagulant (DOAC) and warfarin users according to sex (A: male cohort and B: female cohort). The difference between new DOAC and new warfarin users in male (hazard ratio [HR], 0.84; 95% confidence interval [CI], 0.49–1.44; p=0.524) and female (HR, 0.79; 95% CI, 0.42–1.49; p=0.461) patients was not statistically significant.

5. Risk of GIB according to DOAC subtypes in DOAC users

Fig. 5 shows the Kaplan-Meier plots for GIB in new DOAC users according to the DOAC subtypes. The risk of GIB was lower in “edoxaban or apixaban” users than in “rivaroxaban or dabigatran” users by 26% (HR, 0.74; 95% CI, 0.69 to 1.00; p=0.049).

Figure 5. Kaplan-Meier plots for the gastrointestinal bleeding between rivaroxaban or dabigatran and edoxaban or apixaban users. The risk of gastrointestinal bleeding was 26% lower in edoxaban or apixaban users than in rivaroxaban or dabigatran users (hazard ratio, 0.74; 95% confidence interval, 0.69–1.00; p=0.049).

To the best of our knowledge, this is the first distributed network analysis using CDM data to investigate the risk of GIB in new DOAC and warfarin users. In this real-world database, the risk of GIB in new DOAC users was comparable to that in new warfarin users (HR, 0.95; 95% CI, 0.65 to 1.40; p=0.808), consistent with previous studies.^6,7 Our real-world data supports the recent evidence that the risk of GIB associated with DOAC may be lower than those initially reported.^2,3 In addition, the risk of GIB was not different between new DOAC and new warfarin users in older and both sexes cohorts. As demonstrated in a systematic review and meta-analysis, which showed a reduced risk of GIB associated with DOAC by combined use of acid suppressants,⁹ the combined use of acid suppressants may reduce the risk of GIB in DOAC users in real-world practice. Our data showed that 58.9% of warfarin users and 62.2% of DOAC users were prescribed “drugs for acid-related disorders” before cohort entry. Among anticoagulant users, GIB events was not lower but rather similar in the group prescribed drugs for acid-related disorders compared to the group not receiving such prescriptions (Supplementary Fig. 2). Since the patients using drugs for acid-related disorders is typically considered as a high-risk group of GIB, it is inferred that their GIB events might have decreased by the use of these medications to the level of the low-risk group of GIB, who had not used drugs for acid-related disorders. We also found that the risk of GIB was lower in edoxaban or apixaban users than in rivaroxaban or dabigatran users by 26% (HR, 0.74; 95% CI, 0.69 to 1.00; p=0.049).

GIB is the most common bleeding complication among anticoagulant users and is associated with considerable morbidity and mortality (5% to 15%).^16,17 Early reports on the risk of GIB from randomized controlled trials (RCTs), systematic reviews, and observational studies showed a 25% to 30% higher risk of GIB in DOAC users compared to warfarin users.^2,3,18 However, recent publications reported a substantially decreased risk of GIB in DOAC users.⁴ A meta-analysis of data from 43 RCTs and 41 real-world reports showed no significant difference in the risk of major GIB among DOAC users (1.19%) compared to conventional treatment (0.92%) for various indications.⁶ Our CDM-based analysis also supports the evidence that the GIB risk does not differ significantly between new DOAC and new warfarin users in real-world practice.

Recent studies have shown a higher risk of GIB associated with dabigatran and rivaroxaban, but a lower risk of GIB associated with apixaban and edoxaban,^2,6,19-23 which are consistent with the findings of our study. Previous observational studies showed that dabigatran and rivaroxaban have a trend toward age-related risk of GIB compared to warfarin.^3,19,20 In a population-based study, dabigatran was associated with a higher risk of GIB than warfarin in older patients ≥75 years (HR, 1.30; 95% CI, 1.14 to 1.50).¹⁹ In Medicare data, dabigatran use also had higher GIB rates when compared to warfarin in older patients ≥75 years (HR, 1.28; 95% CI, 1.14 to 1.44).²⁰ Recent meta-analyses of observational and clinical trial data comparing dabigatran and rivaroxaban to warfarin supported older age as dabigatran/rivaroxaban-related GIB risk.^21,22 These studies suggested that warfarin may be safer in older patients ≥75 years. However, in a recent review of five RCTs, apixaban and edoxaban significantly reduced major GIB events in older patients with atrial fibrillation.^23,24 A meta-analysis of RCTs conducted in 75-year-old patients showed that apixaban significantly reduced major GIB compared to warfarin.²⁵ Recent real-world data also confirmed a better safety of apixaban than rivaroxaban among Medicare beneficiaries with atrial fibrillation older than 65 years.²⁶ Their mechanism has not yet explained this potential difference in GIB risk among subtypes of DOAC. In our study, the risk of GIB did not differ between new DOAC and new warfarin users in older patients, as DOAC in our study included all subtypes. Therefore, the risk of GIB in DOAC users should be interpreted cautiously according to the DOAC subtypes.

Our study has several strengths. First, previous studies compared the risk of GIB in DOAC or warfarin users with specific indications, such as atrial fibrillation or venous thrombosis; however, the risk of GIB was compared without specific disease indications in this study. Therefore, our study shows the risk of GIB in anticoagulant users, regardless of their real-world indications. However, GIB risk in the real world should be carefully interpreted, as we excluded patients with a history of GIB and those with chronic liver disease, chronic kidney disease, renal dialysis, or kidney transplantation. Second, as in other CDM studies, our target and comparative cohorts were well matched after extensive PSM.13,²⁷ Our new warfarin users and new DOAC users were well matched for any medications that may influence the risk of GIB, such as antithrombotic agents and drugs for acid-related disorders. Therefore, we overcame the limitations of confounding bias, although this study was designed as a retrospective real-world analysis. Although most RCTs focusing on GIB have ruled out medications such as antithrombotic agents, anti-inflammatory drugs, antirheumatic drugs, and drugs for acid-related disorders that may affect GIB, they do not reflect real-world medical practice. Therefore, we adjusted for these drugs without excluding them. Third, CDM-based distributed network research is a powerful system that can analyze numerous data worldwide by sharing only algorithms with keeping personal security.^13,27 Finally, we used a new user design and only included subjects first exposed to either the target or comparative drugs.

Despite its strengths, this study has several limitations. First, the baseline characteristics of the eight hospitals may have differed because of the different prescription patterns of physician or patient needs. However, these effects may not have been significant because this study was an integrated analysis of data from eight hospitals, and CDM data was based on the standardized data with the same structure. Second, due to the study’s retrospective nature, the medications prescribed may not have matched the actual drug intake of the patients. However, DOAC and warfarin are critical drugs for preventing complications; therefore, most patients may take the prescribed medication in the real world. Third, we defined the control group as warfarin users to enhance comparability with DOAC cohorts; therefore, the control group may not reflect the general healthy population. However, healthy controls may not be appropriate for analyzing the risk of GIB because DOAC users usually have comorbidities and take multiple drugs. Fourth, the risk of GIB may be affected by the previous use or dose of anticoagulants. In our study, previous use of anticoagulants before cohort entry was excluded. In addition, DOAC is used with a fixed dose and warfarin is also used with 2 to 5 mg titrating bleeding tendency. Therefore, the risk of GIB may less affected by the previous use or dose of anticoagulants. Fifth, the coding for GIB is much more accurate than other disease coding and widely used without an operational definition in claim-based study. However, GIB events might be under- or over-estimated as many potential cases of GIB do not receive the disease coding or suspicious GIB might be included within the disease coding. Finally, large-scale propensity matching with 237 covariates can reduce the differences between the two cohorts and limit overfitting.²⁸

In conclusion, the risk of GIB in new DOAC users is comparable to that in new warfarin users in real-world practice. In DOAC users, however, the risk of GIB was lower in edoxaban or apixaban subgroups than rivaroxaban or dabigatran subgroups.

This research was supportedby common data model data sharing from Ajou University Medical Center, Ajou University School, Suwon, Republic of Korea.

G.H.K. is an editorial board member of the journal but was not involved in the peer reviewer selection, evaluation, or decision process of this article. No other potential conflicts of interest relevant to this article were reported.

Study concept and design: J.M.C. Data acquisition: J.M.C., H.H.J., W.W.S., S.Y.R., J.H.K., G.H.K., J.P. Data analysis and interpretation: J.M.C., M.K. Drafting of the manuscript: J.M.C. Critical revision of the manuscript for important intellectual content: J.M.C. Statistical analysis: J.M.C., M.S.K. Study supervision: J.M.C. Approval of final manuscript: all authors.

Supplementary materials can be accessed at https://doi.org/10.5009/gnl230406.