Ginkgo and Warfarin Interaction in a Large Veterans Administration Population

Abstract

Ginkgo biloba is a widely used herbal product that could potentially have a severe interaction with warfarin, which is the most frequently prescribed anticoagulant agent in North America. Literature, however, provides conflicting evidence on the presence and severity of the interaction. In this study, we developed text processing methods to extract the ginkgo usage and combined it with prescription data on warfarin from a very large clinical data respository. Our statistical analysis suggests that taking concurrently with warfarin, gingko does significantly increase patients’ risk of a bleeding adverse event (hazard ratio = 1.38, 95%CI: 1.20 to 1.58, p<.001). This study also is the first attempt of using a large medical record databaseto confirm a suspected herb-drug interaction.

Introduction

The National Center for Complementary and Alternative Medicine (NCCAM) is an agency in the U.S. Department of Health and Human Services dedicated to defining the usefulness and safety of complementary therapies through rigorous scientific research. Complementary and alternative medicine (CAM) is a term used to describe two different treatments. NCCAM defines “complementary therapies” as “non-mainstream” approaches used in combination with allopathic medicine, whereas “alternative therapies” are those that replace conventional medicine with non-mainstream approaches (1). Although CAM was previously divided into different categories, NCCAM states there are two basic subgroups – natural products, and mind and body practices. Natural products consist of botanicals (including herbs), vitamins and minerals, and probiotics. These natural products are commonly found in dietary supplements.

According to the National Health and Nutrition Examination Study (NHANES 2003–2006), approximately half of all Americans use supplements and spend $15 billion annually (2,3). Ginkgo biloba is one of the most purchased dietary supplements used in the United States, and is used to treat a variety of conditions such as memory deficits or dementia, intermittent claudication, tinnitus, and many other health concerns. One of the main concerns with ginkgo use is increased bleeding risk. Ginkgo may decrease platelet aggregation and many case reports have suggested increased bleeding risk, as verified by a systematic review (4). The increased bleeding risk posed by ginkgo may therefore be of great concern in persons taking anticoagulants such as warfarin.

Warfarin is one of the most frequently prescribed anticoagulant agents in North America. Warfarin works by blocking the effects of vitamin K, inhibiting the synthesis of clotting factors and preventing thromboembolic events. Despite its widespread use, warfarin therapy is associated with increased risk of hemorrhage and interacts with various medications, dietary supplements, and some foods. Patients taking warfarin are monitored closely for abnormal or increased bleeding and receive frequent blood testing for the prothrombin time international normalized ratio (INR) to ensure the warfarin dose is adequate yet safe. A high, out-of-range INR is often associated with warfarin drug-drug or drug-herb interactions and indicates increased risk of bleeding.

The interaction between ginkgo and warfarin has not been adequately studied in patients taking both products in combination. Since ginkgo combined with warfarin is a potentially severe interaction that may result in bleeding, we chose it as the first test case in a research study that evaluated the possibility of this occurrence in patients identified as being on both products. In this study, we mined a large national clinical data repository called VINCI to investigate the ginkgo-warfarin interaction. With over 20 million unique patients and extensive medical records, we were able to identify thousands of patients who were using ginkgo and warfarin concurrently.

Methods

Data Source

We used existing electronic medical record data from the Veterans Administration (VA) Informatics and Computing Infrastructure (VINCI) database. Available to VA researchers, this database includes over 20 million unique patient electronic medical records from all VA hospitals and clinics in the United States, which are compiled using uniform coding of data elements. VINCI also includes a suite of research tools to facilitate analysis, such as natural language processing.

To identify concurrent usage of ginkgo and warfarin, we queried both structured and free text data in the VINCI database. We queried the clinical documents table for information containing the terms “ginkgo” and variants “ginkgo” and “ginko.” Of the matching documents approximately 50% used the terms ginkgo, 25% gingko, and 25% ginko. We also queried the filled prescriptions table for the term “warfarin” and its alternate brand names (Jantoven, Coumadin, Marevan, Lawarin, Waran, and Warfant). There were no occurrences of the alternate brand names, since they are not on the VA formulary, so future queries only used warfarin.

Natural Language Processing (NLP)

An NLP module was developed to further process the notes retrieved by the ginkgo query. We prioritized the NLP of ginkgo cases because little structured data are available for herbal supplement usage (Table 1).

Table 1.

Comparison of # of dual use patients (ginkgo + warfarin) identified using structured data alone versus structured data with text notes.

	# Patients (Structured Data)	# Patients (Free Text Notes + Structured Data)	# Overlapping Patients
Ginkgo + warfarin	9	9,862	8

For NLP development, we randomly selected 100 patients with notes containing any mention of ginkgo or one of its spelling variants (n=441) to create an annotated data set. Two reviewers developed a guideline to establish true positive cases and conducted chart review. The inter-reviewer agreement was calculated (Cohen’s kappa = 0.82).

Based on the manual review, we first crafted a set of processing rules to classify highly prevalent document templates (n=41). These processing rules identified positive occurrences of ginkgo in patient supplement lists recorded within the documents as well as negative occurrences of ginkgo in standard documents instructing the patient not to take ginkgo prior to an upcoming surgery. Then using the annotated documents that do not contain templates, we trained a support vector machine (SVM) model to classify the remaining notes not covered by the template rules. The SVM developed was conducted using the Waikato Environment for Knowledge Analysis (WIKA) sequential minimal optimization (SMO) algorithm with the default parameters and bag-of-word features. The final NLP module first applies the template rules and then applies the SVM model.

To test the NLP module, we further annotated another 200 randomly selected notes retrieved by the ginkgo query and calculated the sensitivity and specificity of the NLP module. On the 200 randomly selected ginkgo notes, the NLP model reached a sensitivity of 97%, specificity of 87%, and F measure of 93%. Applying this NLP model to all ginkgo related notes (n=836,506), 600,107 documents and 132,061 patients were identified as positive.

The documentation of ginkgo usage often does not specify the start date or duration. The warfarin exposure was calculated using the VINCI pharmacy fill record. Co-administration was established when patients were exposed to both ginkgo and warfarin. Combining the NLP results with medical fill records, we found 54,139 combined use events in 9,862 distinct patients (Table 2).

Table 2.

ICD-9-CM Codes for Major Bleeding Events

Diagnosis	ICD-9 Code
Gastrointestinal Bleeding	530.82, 531.2, 531.4, 531.6, 532.2, 532.4, 532.6, 533.2, 533.4, 533.6, 534.2, 534.4, 534.6, 535.x1, 537.83, 562.02, 562.03, 562.12, 562.13, 569.3, 578.x
Intracranial Bleeding	430.x, 431.x, 432.0, 432.1, 432.2, 432.9, 851–854

Key: ICD-9-CM – International Classification of Diseases, Ninth Revision, Clinical Modification

Sample Size

The study sample consisted of all patients in the VINCI database that had at least one warfarin order during the study period (years 2008 to 2014), which provided a sample size of n=807,399 patients. Of these, n=11,003 also used ginkgo at least once concurrently with warfarin, and so composed the warfarin + ginkgo group (gingko group). The remaining n=796,396 formed the warfarin only group (non-gingko group).

Bleeding Events

Bleeding is the most frequent complication of warfarin therapy (5). A large body of evidence evaluating the safety of warfarin therapy is available. In clinical trials, bleeding events are classified as fatal, major, life-threatening, clinically significant, overt, or minor. ‘Major bleeding’ is the most common safety outcome cited in clinical trials but the definition varies across trials. According to the International Society on Thrombosis and Hemostasis, the definition of major bleeding should be based on objective criteria and only include events which are life-threatening, utilize major health-care resources, or result in death (5). A list of ICD-9 codes (Table 2) for bleeding events based on the above criteria have been identified and used in a number of clinical trials and analyses evaluating warfarin safety (5–8). We used this list to identify bleeding events in our patient population (Table 3).

Table 3.

Bleeding category of bleeding events in patients on warfarin, and warfarin plus ginkgo combination, after limiting follow-up to one year and dropping bleeding events on first day of followup (consistent with the Figure 2 graph)

bleeding category (see Appendix 1)	n (% of 122,827 bleeding events)
9	83,802 (68.2)
8	26,066 (21.2)
16	5,793 (4.7)
12	3,759 (3.1)
11	2,682 (2.2)
7	2,396 (2.0)
14	1,764 (1.4)
10	808 (0.7)
6	541(0.4)
15	372 (0.3)
13	364 (0.3)
1	252 (0.2)
2–5, 17	0

Statistical Analysis of Ginkgo-Warfarin Interaction

The comparison of bleeding events between the warfarin only and the warfarin plus Ginkgo groups was made using Cox regression. The cumulative bleeding risk is displayed graphically with Kaplan-Meier plots. Multivariable Cox regression models were also fitted, controlling in a fixed covariate fashion for several binary comorbidities: age 75 or older, history (Hx) of heart failure, Hx of high blood pressure, Hx of vascular diease, Hx of stroke, Hx of diabetes, Hx of hypertension, Hx of renal disease, Hx of liver disease, and Hx of alcohol use. These comorbidities were identified using ICD codes.

Results

At least one bleeding event was noted in 143,360 of the n=796,396 non-ginkgo, warfarin only, patients (18.0%) and in 2,484 of the n=11,003 warfarin patients who at some point while on warfarin were also noted to be using ginkgo (22.6%). It was discovered, however, that the first bleeding event was most frequently noted on the first day that warfarin was noted in the EMR (24.4% of non-ginkgo patient with bleeding event, and 16.4% of ginkgo patients with bleeding event). The bleeding events after day 1 were relatively uniform across the follow-up period. (Table 4)

Table 4.

Timing of first bleeding event and first mention of ginkgo in EMR

Including first 30 days of warfarin use
	Warfarin Only (non-ginkgo group) [n = 796,396]	Warfarin + Ginkgo (ginkgo group) [n = 11,003]
Had a bleeding event, n (% of group size)	143,360 (18.0)	2,484 (22.6)
Three most frequent days-on-Warfarin when first bleeding event occurred, n (% of bleeding events)	day 1: 34,951 (24.4) day 2: 301 (0.2) day 8: 219 (0.2)	day 1: 408 (16.4) day 4: 7 (0.3) day 21: 7 (0.3)
Three most frequent days-on-Warfarin when ginkgo use first noted, n (% of group size)		day 1: 3,033 (27.6) day 2: 41 (0.4) day 20: 35 (0.3)
After eliminating first 30 days of warfarin use
	Warfarin Only (non-ginkgo group) [n = 716,671]	Warfarin + Ginkgo (ginkgo group) [n = 9,601]
Had a bleeding event, n (% of group size)	122,964 (17.2)	392 (4.1)
Three most frequent days-on-warfarin when first bleeding event occurred	day 36: 418 (0.3) day 31: 408 (0.3) day 35: 394 (0.3)	day 33: 3 (0.8) day 34: 7 (0.8) day 42: 6 (0.8)
Three most frequent days-on-warfarin when ginkgo use first noted, n (% of group size)		day 36: 47 (0.5) day 31: 45 (0.5) day 51: 41 (0.4)

It is not likely the bleeding would occur so frequently on the first day of warfarin use, since warfarin does not have that rapid of an anticoagulant effect. Apparently, many patients that were already on warfarin went to a Veterans Administration (VA) hospital to seek care for a bleeding event. These could have been patients who already had an EMR encounter at the hospital, or a patient who sought care there for the first time. Either way, the bleeding event would give the providers cause to inquire about warfarin use and to note its use in the EMR. So on that day, many of these patients had warfarin use noted in their EMR for the first time. To reduce this measurement bias, the first 30 days of warfarin use (first 30 days of follow-up) were next eliminated from the dataset, which reduced the sample size somewhat. This also insured that warfarin dose was stable at a level acceptable to the provider and warfarin induced anticoagulation had reached a therapeutic level. This effectively eliminated the uncharacteristic spike in bleeding events at the first day of follow-up, which was now day 31 on warfarin. (Table 4)

Next, we created a new day 1 of follow-up. For the non-ginkgo group, this was day 31, recoded to day 1. For the ginkgo group, it was the first day that ginkgo was noted in the EMR after the first 30 days of warfarin use. For some ginkgo patients, this was warfarin day 31, and for others, it was somewhat uniform across days after that. The first day was recoded to day 1 in this group, as well. The follow-up, then, represented the days on warfarin only for the non-gingko group, and days on warfarin + ginkgo combination for the ginkgo group. Follow-up ended with the first bleeding event, or the end of therapy of these agents in the EMR, whichever came first. This is consistent with a time to first event survival analysis (Cox regression).

To investigate if a bleeding event created an information bias, where the provider would more carefully inquire about ginkgo use if a bleeding event occurred, we created a table in a case-control fashion (Table 5). If bleeding created an obvious information bias, giving the provider cause to inquire about ginkgo use, ginkgo would be expected to be noted in the EMR more often than in patients who never have a bleeding event. The opposite occurred. Ginkgo was noted one-fifth as often in patients with at least one bleeding event compared to patients without a bleeding event (Table 5). This reveals that a ginkgo information bias is only subtle, if it does exist.

Table 5.

Probability (%) of noting Ginkgo in the EMR, conditional upon a bleeding event was also noted in the EMR.

		Ginkgo noted at least once in the EMR
At least one bleeding event noted in the EMR	No	9209/602916 (1.5%)
	Yes	392/123356 (0.3%)

The follow-up was limited to one year, under the assumption that if ginkgo interacts with warfarin, its effect should be seen by then. It was suspected that ginkgo might first appear in the EMR because a bleeding event had occurred, giving the provider reason to more carefully inquire about herbal use with the patient. Table 6 shows the patients with follow-up periods of days 1, 2, or 3, to determine if bleeds are occurring at the time of beginning of follow-up. This was true about one-third of the time in ginkgo group. Unfortunately, there is no way to extract from the EMR when ginkgo exposure actually first occurred, just when its use was first recorded.

Table 6.

Days of follow-up after eliminating first 30 days of warfarin use and restricting follow-up to one year

	Warfarin Only (non-ginkgo group) [n = 717,831]		Warfarin + Ginkgo (ginkgo group) [n = 9,601]
	No bleed [n = 677,855]	Bleed [n = 38,816]	No bleed [n = 9,270]	Bleed [n = 331]
Follow-up period in days for first three days, n (%)
1	880 (0.1)	408 (1.0)	2,721 (29.3)	121 (36.6)
2	714 (0.1)	331 (0.9)	70 (0.8)	5 (1.5)
3	694 (0.1)	296 (0.8)	34 (0.4)	2 (0.6)

Ignoring time-at-risk, a bleeding event was observed in in 331 (3.4%) patients in the warfarin plus ginkgo group and in 38,816 (5.4%) patients in the warfarin only group. After accounting for time-at-risk in a univariable Cox regression model, ginkgo was associated with a higher risk of bleeding (hazard ratio = 2.08, 95%CI: 1.87 to 2.32, p<.001). The association is shown as a Kaplan-Meier curve in Figure 1. The association was unchanged after adjusting for co-morbidities (hazard ratio = 2.08, 95%CI: 1.87 to 2.32, p<.001).

Figure 1.

Kaplan-Meier graph of risk of bleeding for one year of follow-up, which hazard ratio from a univariable Cox regression model.

To assess if the association was just a matter of the high incidence of bleeding noted on the first day of ginkgo follow-up, the analysis was repeated after dropping subjects with one day of follow-up. This left a sample size of n=6,759 with in the warfarin plus ginkgo group, with 210 bleeding events. Ignoring time-at-risk, a bleeding event was observed in 210 (3.1%) patients in the warfarin plus ginkgo group and in 38,408 (5.4%) patients in the warfarin only group. After accounting for time-at-risk in a univariable Cox regression model, ginkgo was associated with a higher risk of bleeding (hazard ratio = 1.38, 95%CI: 1.20 to 1.58, p<.001). The association was unchanged after adjusting for co-morbidities (hazard ratio = 1.38, 95%CI: 1.20 to 1.58, p<.001).

Discussion

Although the potential bleeding risk of Ginkgo biloba has been much discussed in the literature, initial concerns were based on case reports. These reports described a temporal association between ginkgo use and the bleeding events (4). Some evaluations of randomized controlled trials have not found a higher bleeding risk (10). A systematic review and meta analysis of 18 randomized controlled trials looked at the impact of ginkgo on hemostasis parameters associated with bleeding risk and found a significant reduction in blood viscosity. However, there were no effects on other factors, such as ADP-induced platelet aggregation, fibrinogen concentration, activated partial thromboplastin time, and prothrombin time. Thus the authors concluded there was no higher bleeding risk. Another evaluation looked at claims data in the Taiwan National Health Insurance Research Database (11). The authors concluded there was no significant correlation to the risk of hemorrhage but did note that caution should be exercised in elderly persons and those with known bleeding risk.

Thus the existing literature does not confirm the ginkgo-anticoagulant interaction. The effects and true risks of this interaction are difficult to estimate, based on the limited quantity and quality of existing reports. Leveraging a large database, with documentation of herb use in a high number of patients with cardiovascular disease, as is the case with this project, a positive signal between warfarin and ginkgo was found. This is new information that may help to elucidate the relationship between gingko and wafarin usage and bleeding events.

One limitation we noted is that the documentation rate of herb usage in medical records is possibly low. As shown in the Table 7, the NLP extracted ginkgo use in this study is fairly consistent with those reported in the literature.

Table 7.

Prevalence of Ginkgo usage derived from the NLP of VHA notes and the prevalence reported by literature.

Population	Time Period	Sample Size	Prevalence
National sample of adults (12)	Within 12 months	23,393	1.8%
Nursing home patients (13)	Within 12 months	68,403	0.4%
Veterans with cancer (14)	Current use	200	0%
Veteran outpatients (15)	Current use	458	10%
Veterans with cancer (16)	Current use	200	1%
Veterans patients on warfarin	Concurrent with warfarin	726,272	1.3%

Another limitation of study is the lack of detailed information in the EMR. The length of time of ginkgo is obviously longer than what is noted in the EMR. This could have impacted the statistical result. However, the separation of the two curves in Figure 2 occurring after day 30 of follow-up helps support the increased risk observed with the warfarin-ginkgo combination. The dosage and frequency of herbal product use is not regulated and rarely recorded. Information bias is well-known to exist in medical records, with missing information and more complete information being in sicker patients (17).

Figure 2.

Kaplan-Meier graph of risk of bleeding for one year of follow-up, after dropping patients with only one day of follow-up, which hazard ratio from a univariable Cox regression model.

A final limitation is the possibility of confounding by other medications. That is, some patients might have received other medications that could have increased their risk of bleeding, such as aspirin or other antiplatelet agents. No attempt was made in this study to collect those data in order to control for such medications. However, to use confounding by other medications as an explanation for the direction of the observed study outcome, with higher bleeding risk for the warfarin plus ginkgo group, these other medications would had have to have been prescribed more often in the warfarin plus ginkgo group, and it is not likely providers would have chosen to do that.

Despite the limitations, our analysis provides information that is previously unavailable to researchers and clinicians. Given the wide use of ginkgo in the US population, the bleeding risks associated with gingko-wafarin usage is worth noting. In the future, we plan to investigate other suspected herb-drug interactions.

Appendix 1. Bleeding Categories

Category	ICD9 Codes	Description
1	530.82	Esophageal hemorrhage
	531.2	Acute gastric ulcer with perforation
2	531.4	Chronic or unspecified gastric ulcer with hemorrhage
	531.6	Chronic or unspecified gastric ulcer with hemorrhage and perforation
	532.2	Acute duodenal ulcer with hemorrhage and perforation
3	532.4	Chronic or unspecified duodenal ulcer with hemorrhage
	532.6	Chronic or unspecified duodenal ulcer with hemorrhage and perforation
	533.2	Acute peptic ulcer of unspecified site with hemorrhage and perforation
4	533.4	Chronic or unspecified peptic ulcer of unspecified site with hemorrhage
	533.6	Chronic or unspecified peptic ulcer of unspecified site with hemorrhage and perforation
	534.2	Acute gastrojejunal ulcer with hemorrhage and perforation
5	534.4	Chronic or unspecified gastrojejunal ulcer with hemorrhage
	534.6	Chronic or unspecified gastrojejunal ulcer with hemorrhage and perforation
6	537.83	Angiodysplasia of stomach and duodenum with hemorrhage
	562.02	Diverticulosis of small intestine with hemorrhage
7	562.03	Diverticulitis of small intestine with hemorrhage
	562.12	Diverticulosis of colon with hemorrhage
	562.13	Diverticulitis of colon with hemorrhage
8	569.3	Hemorrhage of rectum and anus
9	578*	Gastrointestinal hemorrhage
10	430*	Subarachnoid hemorrhage
11	431*	Intracerebral hemorrhage
	432.0	Nontraumatic extradural hemorrhage
12	432.1	Subdural hemorrhage
	432.2	<non-existent>
	432.9	Unspecified intracranial hemorrhage
13	851*	Cerebral laceration and contusion
14	852*	Subarachnoid subdural and extradural hemorrhage following injury
15	853*	Other and unspecified intracranial hemorrhage following injury
16	854*	Intracranial injury of other and unspecified nature