Normalizing Dietary Supplement Product Names Using the RxNorm Model

Rui Zhang, PhD, Facmi, Famia, Fiahsi, Jake Vasilakes, Yadan Fan, Rubina F. Rizvi, Anusha Bompelli, Olivier Bodenreider, Rui ZhangMedinfo 2019 · DOI

TL;DR

This paper tackles the challenge of inconsistent naming for dietary supplements (DSs) by adapting the RxNorm model, a standard for normalizing drug names. They developed a rule-based natural language processing system that significantly improves the coverage of DS product name normalization compared to existing methods, paving the way for better data interoperability and research.

What Problem Does This Paper Address?

The use of dietary supplements (DSs) is on the rise, making it crucial for consumers, clinicians, and researchers to easily find information about these products. However, DS product names are highly variable due to loose labeling regulations, making it difficult to search and aggregate information. Unlike prescription drugs, which have stricter naming conventions and are normalized by systems like RxNorm, DSs lack a standardized nomenclature. This inconsistency hinders critical tasks such as cross-platform communication, knowledge discovery, and pharmacovigilance. The paper aims to bridge this gap by developing a normalization system specifically for DS product names, inspired by the successful RxNorm model.

Methods

The authors developed a rule-based natural language processing (NLP) system to normalize DS product names. The system leverages three existing terminologies: the Therapeutic Goods Administration (TGA), RxNorm, and iDISK (an integrated knowledge base for DSs). The process involved:

  1. Data Extraction and Preprocessing: Extracting 12,383 product names from the Dietary Supplement Label Database (DSLD), focusing on single-ingredient products to reduce initial variability. This data was split into development (80%) and evaluation (20%) sets.
  2. NLP Pattern Development: Iteratively developing term types (e.g., Brand Name, Ingredient, Dose Form) and regular expression patterns to match these components within product names. They used keyword lists from TGA and RxNorm, augmented with manual additions for abbreviations and plural forms. Brand names were matched using a rule-based method and a curated list. Common stop words were also identified and marked.
  3. Template Generation: For each product name, the system outputs an ordered list of term type codes, forming a "template." These templates are designed to correspond to higher-level RxNorm term types.
  4. Evaluation: The system's performance was evaluated based on coverage (percentage of product names fully or partially matched) and accuracy (correctness of term type assignments).

Datasets

  • Dietary Supplement Label Database (DSLD): Used for extracting 12,383 single-ingredient product names. This dataset served as the basis for both training (development set) and evaluation.
  • Existing Terminologies: TGA, RxNorm, and iDISK were used to inform the development of NLP patterns.

Key Insights and Ideas

  • RxNorm's Limitations for DSs: The study highlights that existing RxNorm templates do not adequately cover the nuances of dietary supplement product names, necessitating a DS-specific approach.
  • Brand Names are Dominant: Brand names (BN) play a significantly larger role in DS product labeling than in drug labeling, with 91% of the generated patterns containing BN. This is attributed to the competitive DS market emphasizing branding.
  • Normalization is Feasible but Challenging: The research demonstrates that normalizing DS product names is achievable, but the inherent variability and complexity of these names require ongoing refinement.
  • Trade-off Between Coverage and Specificity: The system achieved good coverage but also encountered challenges with ambiguous terms and phrases that could be interpreted as ingredients, dose forms, or claims.

Key Results

  • Coverage Improvement: The developed system achieved a 72% coverage on DS product names, representing a 32% increase over the existing RxNorm model's coverage (39.89%).
  • Accuracy: The system achieved a normalization accuracy of 0.86 on the evaluation set, improving to 0.95 for fully matched product names.
  • Template Generation: The system generated 136 unique templates across the development and evaluation sets.
  • Common Templates: The most frequent templates involved Brand Name (BN) and Ingredient (IN), often combined with Strength (STR) or Dose Form (DF). For example, "BN IN STR" was the most frequent template, accounting for 33% of fully matched product names.
  • Term Type Accuracy: Many term types, including BN and IN, achieved high accuracy (above 0.90), while others like Animal Source (ANM) had lower accuracy (0.61).

Limitations and Future Work

  • Single-Ingredient Focus: The study was limited to single-ingredient products, which represent only 22% of all DS products in DSLD. Generalizing to multi-ingredient products is a key area for future work.
  • Assumption of Single Ingredient: The system assumes at most one ingredient per product name, which is a simplification. Modifying the system to handle multiple ingredient mentions is suggested.
  • Contextual Disambiguation: The system has limited ability to disambiguate context, leading to potential misinterpretations of polysemous keywords. Future work should explore advanced NLP and machine learning methods for better contextual understanding.
  • Unseen Brand Names: Many unmatched words were brand names not present in the development set. Machine learning approaches could improve coverage for these.
  • Ingredient vs. Preparation/Plant Source Errors: Errors in ingredient (IN) term type assignment were sometimes due to phrases belonging to Plant Source (PLNT) or Preparation (PREP) being incorrectly identified as ingredients. This suggests a need for refined pattern searching order or logic.

Summary updated

BibTeX
@inproceedings{fiahsi2019normalizing,
  author = {Rui Zhang, PhD, FACMI, FAMIA, FIAHSI and Jake Vasilakes and Yadan Fan and Rubina F. Rizvi and Anusha Bompelli and Olivier Bodenreider and Rui Zhang},
  title = {{Normalizing Dietary Supplement Product Names Using the RxNorm Model}},
  booktitle = {Medinfo},
  year = {2019},
  doi = {10.3233/shti190253},
}

Loading related content...

Figure