Term Coverage of Dietary Supplements Ingredients in Product Labels

Abstract

As the clinical application and consumption of dietary supplements has grown, their side effects and possible interactions with prescribed medications has become a serious issue. Information extraction of dietary supplement related information is a critical need to support dietary supplement research. However, there currently is not an existing terminology for dietary supplements, placing a barrier for informatics research in this field. The terms related to dietary supplement ingredients should be collected and normalized before a terminology can be established to facilitate convenient search on safety information and control possible adverse effects of dietary supplements. In this study, the Dietary Supplement Label Database (DSLD) was chosen as the data source from which the ingredient information was extracted and normalized. The distribution based on the product type and the ingredient type of the dietary supplements were analyzed. The ingredient terms were then mapped to the existing terminologies, including UMLS, RxNorm and NDF-RT by using MetaMap and RxMix. The large gap between existing terminologies and ingredients were found: only 14.67%, 19.65%, and 12.88% of ingredient terms were covered by UMLS, RxNorm and NDF-RT, respectively.

Introduction

According to the National Health and Nutrition Examination Survey (NHANES) 2003-2006, 53% of U.S. adults took at least one dietary supplement, most of which were multivitamin and multi-mineral supplements¹. Recently, the National Health Statistics Reports indicated that about 40% of Americans use some form of complementary and alternative medicine (CAM) and that non-vitamin/non-mineral dietary supplements were still the most commonly used complementary and alternative health approach². The prevalence of supplement use was estimated to be 69% in 2011 when occasional and seasonal use was taken into account³.

Dietary supplements are typically used to complement conventional medicine with the goal of achieving better healthcare outcomes; however, dietary supplements were found to result in 23,000 emergency room visits yearly in the U.S. according to the study conducted by Food and Drug Administration (FDA) and the Centers for Disease Control and Prevention (CDC)⁴. Moreover, one out of four people taking a prescription medicine also uses an herbal supplement, increasing the possibility of drug-supplement interactions (DSIs). For example, warfarin can interact with supplements such as Panax ginseng and Gingko biloba, which can lead to severe adverse effects such as spontaneous postoperative bleeding⁵. In particular, supplements are increasingly used by patients diagnosed with cancer to help strengthen their immune system and ease the side effects of treatments. Unfortunately, our ability to readily identify adverse effects from herb and dietary supplements and their reactions with conventional Western medications are currently limited and the reports on such interactions occur infrequently in clinical practice⁶. Therefore, it is necessary to gather information on the ingredients in these supplements to facilitate medication safety efforts.

New drugs are typically tested for their efficacy and toxicity before market approval. However, U.S. Food and Drug Administration (FDA) regulates dietary supplements differently from conventional food and drugs under a separate regulation called Dietary Supplement Health and Education Act of 1994 (DSHEA). DSHEA requires appropriate labeling of dietary supplements⁷. These labels contain rich information such as suggested use, ingredients, product indication, target population, and other necessary safety precautions. In particular, the dietary supplement labeling guideline requires all dietary ingredients to be listed on the label, and the synonyms for the dietary ingredients can be used⁷. Such information provides a great resource to analyze the ingredients of dietary supplements from the perspective of terminology.

To support effective mining of product labels information for DSIs and supplements adverse effects studies, it is vital to understand the representation of supplements and ingredients in dietary supplement labels as well as the coverage of standard biomedical terminologies for supplements and their ingredients. One possible reason for the lack of extensive studies on DSIs is an absence of a standard and accepted terminology for dietary supplements. In a recent study comparing terms with different resources, we found that none of five major online databases covered all supplement terms⁸. Another prior study also evaluated the supplement term coverage in both medication lists and clinical notes in the electronic health record (EHR)⁹. To the best of our knowledge, the investigation of supplement term representation and coverage on dietary supplement labels is limited and deserves further investigation. In this study, we sought to evaluate terminology coverage of data elements in supplement product labels by cross-validating online supplement databases with EHR patient data. We also sought to investigate the adequacy of standard terminologies for representing supplements, especially their product ingredients, based on existing product labels.

Methods

We extracted supplement product information and their ingredients from the Dietary Supplement Label Databases (DSLD), and then normalized and formed a comprehensive list of ingredients followed by mapping them to standard terminology, including UMLS, RxNorm, and NDF-RT. The term coverage in these terminologies was investigated. In our assessment, we compared the term overlap among three resources, including supplement labels, online resources, and the EHR medication list. We first introduce DSLD, MetaMap, and RxMix which are used in this study, followed by the details of methods.

DSLD

The DSLD is created by the National Institutes of Health (NIH) Office of Dietary Supplement (ODS) and U.S. National Library of Medicine (NLM). The database collects full label contents from dietary supplement products including both currently available products as well as products which are no longer on the market in the U.S., as well as those consumed by National Health and Nutrition Examination Survey (NHANES) participants. Each product in DSLD provides four types of label information: product information (including serving information, product type, supplement form, target groups), dietary supplement facts (e.g., usage, ingredients), label statement (e.g., FDA statement, precautions, and suggested use), and contact information.

Unified Medical Language system (UMLS) and MetaMap

UMLS is a repository that integrates over 100 medical vocabularies and provides a unified platform which can be used to develop or enhance applications. Metathesaurus, as one of the three main components in the UMLS, contains over 2 million terms and codes from many different vocabularies such as Systematized Nomenclature of Medicine - Clinical Terms (SNOMED CT), Current Procedural Terminology (CPT), Logical Observation Identifiers Names and Codes (LOINC), etc. Each concept in the Metathesaurus has a Concept Unique Identifier (CUI), which can be used to map to various dictionaries and compare data from different sources. For this paper, UMLS was used to normalize data from different sources (i.e., supplement ingredients).

To map the biomedical texts to the UMLS Metathesaurus, the NLM developed and maintains a natural language processing tool, called MetaMap, which lexically and syntactically analyzes texts and provides a list of mapping concept candidates. We used MetaMap to find the matched UMLS concepts in Metathesaurus. The versions of Metamap and UMLS Metathesaurus used in this study are 2016 and 2015AB, respectively.

RxNorm, NDF-RT, and RxMix

RxNorm provides a standardized nomenclature of clinical drugs by integrating 12 drug vocabularies¹⁰. RxNorm contains names not only for prescription drugs but many over-the-counter drugs on the U.S. market. RxNorm provides the ability to normalize the drug names along with drug dosage, generic name, chemical components, and dosage forms to support the communications between different systems. The version of RxNorm used in this article is “04- Jan-2016”.

National Drug File - Reference Terminology (NDF-RT) is developed by the U.S. Department of Veterans Affairs, Veteran Health Administration. It is used to represent drug characteristics such as ingredients, chemical structure, and additional information about their molecular interactions and kinetics¹¹. The version of NDF -RT used in this article is “2016.01.04”.

RxMix is a web application integrating various functions from the RxNorm and NDF-RT application program interfaces (APIs) to allow users to create their own workflow to conduct a certain task, such as finding drug class from a drug name¹². In this paper, RxMix was used to map terms to RxNorm and NDF-RT concepts.

Extracting and analyzing product information from DSLD

Step 1: Downloading the data from DSLD. For each supplement in DSLD, the data containing the abovementioned label information are presented in four separate comma-separated values (CSV) files that are available for download. DSLD consists of 45,455 up-to-date records, downloaded through batch processing by Python according to the unique DSLD identifier of each supplement.

Step 2: Extracting the product information of the supplements. The information pertaining to the product name and the product type can be found in the Product Information. DSLD used the LanguaL™ system for the classification of the supplements since LanguaL™ is an “automated method for describing, capturing and retrieving data about food”¹³. In total, 12 categories (e.g., single vitamin and single mineral supplement) were applied to label more than 45,000 dietary supplements. It is noteworthy that no generic names for these supplements can be found in the database, and the manufacturers have come up with their unique nomenclature for these mixtures. Therefore, the most useful information in terms of terminology in the product information label is the LanguaL™ supplement type. LanguaL™ assigns the letter “A” to all dietary supplements and a four-digit code following the letter for further classification. The supplement types for each supplement were extracted and the distribution of the supplements in each type was then analyzed.

Normalizing and analyzing supplement ingredient list

Since no generic names could be found for the supplements in DSLD, more emphasis were put on scrutinizing the ingredients of each supplement. Here, we describe how to extract, normalize and analyze ingredients for each supplement.

Step 1: Extracting ingredient names and ingredient types. The information pertaining to the ingredient names and ingredient types was further extracted from Dietary Supplement Facts. Therefore, it is straightforward to iterate through the items in each dietary supplement facts file and extract the relevant information. The number of ingredients in each supplement was also recorded. Then the information were categorized into 20 different types (e.g., vitamin and amino acids) which appeared in the data. Among these, two categories of ingredients (i.e., default and header) were excluded in this study as these types of ingredients are not specifically classified in the database. There are 60,005 items that are belong to this class that was named “default”. The “header” type of ingredients is used to organize the presentation structure of dietary supplement facts on the label, while the information itself does not contain any useful information about the ingredients.

Step 2: Normalizing the ingredient information. After extracting the ingredient raw data, it was discovered that further data cleaning is required before any terminology mapping could be done. We removed information including comment in parentheses, brackets and braces, HTML marker residue from webpage sources left in the CSV files during the transfer process, non-alphanumeric characters and various duplicates for a single ingredient type with regular expression filters. It was found that the different formulations of the same ingredients contributed to the majority of redundant information. For example, the ingredient acai has multiple duplicates in the record: acai extract, acai powder, acai freeze-dried powder, acai concentrate, acai juice, etc. The word “extract”, “powder”, “concentrate” does not contribute to efficient mapping. To eliminate all these duplicates, a keyword list with formulations (e.g., extract, powder, concentrate) was manually summarized and used to preprocess these ingredient list using additional mapping steps. The list of original and normalized ingredients and the Python source code are available at: https://github.com/Schneitzer/Term-Coverage-of-Dietary-Supplements-in-Product-Labels.

Step 3: Analyzing the ingredient information. After normalizing the ingredient information, a list of unique terms and their ingredient types were generated. The normalizing efficiency can also be deduced by comparing the number of items in each set before normalizing and after normalizing.

Mapping ingredient terms to existing terminologies

After normalizing the ingredient data, seven subsets were chosen without loss of generality, i.e., amino acids, animal part or source, botanical, chemical, enzyme, hormone, and vitamin, for mapping.

Step 1: Mapping the list of normalized ingredients to UMLS using MetaMap. The mapping result contains a matching score according to the accuracy of mapping could be measured. A score of 1000 represents a perfect match, while lower scores indicates a partial match. It is possible that a certain ingredient could be mapped to multiple keywords, which will obfuscate its clinical significance due to a partial match. Therefore, only mappings with 1000 matching score were further analyzed in this study.

Step 2: Filtering the list of ingredient information according to the semantic types. The UMLS semantic types of the exact matched concepts were examined as some of those concepts may not refer to dietary supplements. For example, the concept “beta carotene” has two exact match in UMLS Metathesaurus, but only one of them refers to “organic chemical, pharmacologic substance, and vitamin”; the other maps to “Beta carotene Measurement”, which is a laboratory procedure. In addition, although the ingredient information has been preprocessed and normalized, there are still items that may contain long phrases that may produce false positives. Therefore, only a certain subset of semantic types (e.g., “Enzyme, organic chemical”) that are closely related to dietary supplements were manually chosen to filter out the mapping noise.

Step 3: Mapping the list of normalized ingredients from each ingredient type to RxNorm and NDF-RT using RxMix¹². The normalized lists were directly uploaded to RxMix workflow for mapping. Two function, RxNorm:findRxcuiByString (with normalized string match parameter selected) and NDF-RT:findConceptsByName, were applied to find the unique concept identifiers in RxNorm and NDF-RT terminologies.

Evaluating ingredient term coverage of terminologies

After the ingredient information was mapped to UMLS, RxNorm and NDF-RT terminologies, the unique concepts were collected although many ingredient terms can be mapped to the same concept although we normalized the ingredients terms described above. The number and percentage of the unique mapped concepts in each terminology were calculated to indicate the unique terms of the ingredient information of DSLD that can be mapped to all the databases included in this research.

Results

Extraction of supplement list from DSLD

All 45,455 supplements from DSLD databases were extracted and classified by 12 LanguaL™ product types. Number and percentage of each product type were shown in Table 1.

Table 1:

The supplement representation in DSLD; number, percentages and LanguaL™ product type.

LanguaL™ Product Type	Number	Percentage
DIETARY SUPPLEMENT – COMBINATION/OTHER	16531	36.4%
DIETARY SUPPLEMENT – HERBAL OR BOTANICAL	8681	19.1%
DIETARY SUPPLEMENT – NON-NUTRIENT/NON-BOTANICAL SUPPLEMENT	5448	12.0%
BOTANICAL SUPPLEMENT WITH VITAMIN/MINERAL	5099	11.2%
DIETARY SUPPLEMENT – VITAMIN	2922	6.4%
DIETARY SUPPLEMENT – MINERAL	1727	3.8%
DIETARY SUPPLEMENT – AMINO ACID OR PROTEIN	1558	3.4%
FATTY ACID OR FAT/OIL SUPPLEMENT	1444	3.2%
MULTI-VITAMIN AND MULTI-MINERAL SUPPLEMENT	1266	2.8%
SINGLE VITAMIN AND SINGLE MINERAL SUPPLEMENT	509	1.1%
DIETARY SUPPLEMENT OTHER NUTRITIVE SUPPLEMENT	270	0.6%

About 36% of supplements were classified as “dietary supplement – combination/other”. According to the LanguaL™ code, “dietary supplement – combination/other” is a subcategory of “dietary supplement – combination” where 16,531 dietary supplements are included under this category. Among these, 16,522 dietary supplements were classified as “combination/other” subcategory and the remaining 9 products were classified in the “combination” category. We combined these two product types and listed them as one type in the table. However, this is not the only overlap found in this classification system. For herbal and botanical supplements with vitamin or mineral, they could either appear in the “dietary supplement – herbal or botanical” or “botanical supplement with vitamin/mineral” listed in Table 1 since the second category does not specifically exclude the supplements that contain vitamin or mineral. A graphic presentation of the distribution of each type is shown in Figure 1.

Figure 1.

Distribution of the number of ingredients present in the supplements listed in DSLD. Due to the long tail of the data, 1,091 supplement (2.4%) containing more than 50 ingredients were not shown in Figure 1.

Normalizing and analyzing ingredient list

Before normalizing the ingredient information, the distribution of the number of ingredients in each supplement is shown in Figure 1. It can be seen that the histogram peaks at 13,004 for one ingredient, implying about 30% of the supplements have one dietary ingredient. The distribution is also very long-tailed. The 25^th percentile, the median, the 75^th percentile and the mean for the number of ingredients are 1, 5, 12, and 9.6, respectively. Among 45,455 dietary supplements in the DSLD database, 1,091 items (2.4%) included more than 50 ingredients on their supplement labels, which were not shown in Figure 1.

These ingredients are classified into 20 categories. The only irrelevant category listed in the database is the “header” category, which only serves the purpose of organizing the dietary ingredient facts, and thus will not be discussed further. The normalized results for the other 19 categories were shown in Table 2. The “Before Normalization” column lists the number of items extracted from the dietary ingredient fact sheets of all the supplement products in the DSLD, while the “After Normalization” column lists the number of unique items remaining after normalization. Each category was slimmed down by 81.6% to 99.5%, which facilitates faster cross-referencing among different databases.

Table 2:

Comparison of the scale of ingredient data in DSLD before and after normalization.

Ingredient Category	Before Normalization	After Normalization
Blend	842	147
Bacteria	6535	506
Protein	5929	145
Fiber	5395	76
Element	13160	109
Chemical	23920	1359
Vitamin	92604	504
Fatty Acid	10632	541
Hormone	512	21
Animal Part or Source	1896	348
Fat	20148	139
Carbohydrate	15919	94
Amino Acid	25387	739
Polysaccharide	207	28
Botanical	80614	6046
Mineral	54721	484
Enzyme	9320	341
Other	34884	4572
No Label	60005	9213

Evaluation of supplements term coverage

The mapping coverage across different terminologies is described in Table 3. The UMLS column presents the number of CUIs mapped by MetaMap. As the mapping process returns explicit semantic types, the quantity of matching CUIs is affected by relevant semantic types as those are not related to dietary supplements were filtered out. But for RxNorm and NDF-RT, the mapping was performed by RxMix that does not output explicit semantic types. RxNorm assigns a unique concept id (RxCUI) for each individual product, making it possible that RxNorm generate more matches than the UMLS.

Table 3:

The number of unique concepts mapped into each terminology by the respective ingredient category.

Ingredient Category	UMLS (CUI)	RxNorm (RxCUI)	NDF-RT (NUI)
Amino Acid	149	192	188
Animal Part or Source	59	78	64
Botanical	1004	1097	515
Chemical	398	463	454
Enzyme	75	78	70
Hormone	4	5	7
Vitamin	128	137	154

The mapping coverage across different resources (e.g., online resources, and EHR medication lists) was also compared. As shown in Figure 2, the DSLD database has more unique terms than online resources and the medication list. However, the mapping percentages of UMLS and RxNorm terminologies to the unique terms of DSLD are much lower than the MedList and online database. The NDF-RT terminology produced similar mapping percentage. It is noteworthy that these ingredient categories are not orthogonal. The unique concepts in one category might overlap with the ones in another category; therefore, the final mapping percentage is slightly lower after removing the duplicates.

Figure 2.

Distribution of the number of matched concepts among different resources (DSLD, medication list, and online databases) in UMLS, RxNorm, and NDF-RT terminologies. The percentage of unique terms that have been mapped to each different resource was shown above the corresponding bar.

Although most of the ingredients can be exactly or partially mapped to the UMLS, there is still a portion of ingredients that cannot be mapped to any terminology. Some examples are given in Table 4. Most of these non-mapping results are due to misspelling or errors which occurred during the transfer from physical labels to digital data; instances where the ingredients are simply names created by manufacturer can also lead to a non-mapping result; for botanical ingredients, the name of the ingredients might be coming from another language (e.g., Chinese) that is not in the terminologies; and for chemical ingredients, some nomenclature of organic compounds are not included in the terminologies.

Table 4:

Selected examples not covered by terminology.

Non-mapping Terms	Problem and Possible Suggestions
Artichock	Spelling error - Artichoke
Sheng Jiang	Names in foreign language – Chinese Pinyin for “raw ginger root”
Dicyclopentanone	Organic compound nomenclature
Aminogen	Manufacturer names

Discussion

Due to the known and potential adverse effects on patient safety, DSIs and supplement side effects have attracted a lot of attention due to potential patient risk. Many researchers have found DSIs through pharmacological experiments; however, they usually focus on a small set of supplements or drugs. To largely explore such DSIs from a variety of resources, standardized terminology of supplements is necessary. We have evaluated the terminology coverage among online resources and medication lists in EHR data. Supplement labels contain rich information, including product information and especially complete lists of their ingredients, which are the main components interacting with drugs. Thus astudy on the coverage of terminology concerning the supplement ingredients by standard terminologies such as UMLS, RxNorm, and NDF-RT needs to be investigated.

A relatively unique problem for supplements as opposed to comparable over-the-counter and prescription medications is the large numbers of ingredients noted in Figure 2. The use of multiple active agents is fairly infrequent in prescription medications. It shows up more frequently with over-the-counter medications, but it is rare to include more than 3 or 4 components in a product. The supplement data indicate a very large number of ingredients, which may be challenging to practical management of supplement content in secondary document sources such as EHRs. Standardization of ingredient terms is vital for information extraction and information retrieval of dietary supplement in many resources, including both structured data and free texts. Standardization can also promote the accurate use of higher level generic or trade names to represent known ingredient combinations whenever it is feasible. The example of multiple vitamins is a typical case where in the medical record only the multivitamin is noted, since it is known to contain a typical set of usual vitamins. Effective terminology and mappings to higher level names such as the trade name of supplement products can make clinical documentation and research much easier to manage.

The low sensitivity of mapping may also due to the fact that the normalization is unable to cover all possible variations of the ingredients. The rule-based regular expression filters cannot eliminate all possible typos, affixes and duplicates for the ingredients. For example, “copper” and “Cu” were treated as two distinct concepts, but they actually refers to the same concept, which is the copper element. On the other hand, only exact matches were taken into account when mapping the UMLS concept, thus yielding a relatively lower number of mapped concepts. All of these factors contribute to lower mapping sensitivity. Another observation is that RxNorm mapped more ingredient terms in DSLD and medication lists than the UMLS Metathesaurus, which is an anomaly as RxNorm contains more detailed information than UMLS Metathesaurus. For example, there are two distinct RxCUIs that correspond to glycine (1311532 (glycine hydrochloride) and 4919 (glycine)). While MetaMap specifies the semantic type for each CUI, RxMix does not reveal the semantic type difference between the different CUIs, thus making the mapping percentage higher for RxNorm. This could cause further complications as these RxCUIs would be consequently applied in RxMix to find corresponding NUIs. More detailed information is required to eliminate all the one-to-many mappings in order to derive more accurate term coverage for dietary supplements.

This study has certain limitations. We only evaluated the term coverage in a limited number of terminologies, and only considered the exact matches, which may ignore some related concepts from UMLS. This study lays the foundation for additional examination of current drug terminologies and the improved understanding of current dietary supplements. Without terminology examination, further steps such as information extraction would be very challenging. Our future study will investigate how to better represent dietary supplements using existing resources and terminologies to enhance dietary supplement research.

Conclusion

In conclusion, we extracted the product information and the dietary ingredients of all the supplements listed in DSLD. While generic names for the supplements in this database were not available, the dietary ingredient facts became the source of data for analysis of term coverage. The distribution of dietary supplements according to the product types and the ingredient types was derived. The ingredient terms have been mapped to concepts in three different terminologies including UMLS, RxNorm and NDF-RT, to evaluate the ingredient term coverage with the aid of MetaMap and RxMix. It has been observed that the RxNorm provides the best coverage (19.65%).