A natural language processing approach towards harmonisation of European medicinal product information

Abstract

Product information (PI) is a vital part of any medicinal product approved for use within the European Union and consists of a summary of products characteristics (SmPC) for healthcare professionals and package leaflet (PL) for patients, together with the product packaging. In this study, based on the English corpus of the EMA product information documents for all centrally approved medicinal products within the EU, a BERT sentence embedding model was used together with clustering and dimensional reduction techniques to identify sentence similarity clusters that could be candidates for standardization. A total of 1258 medicinal products were included in the study. From these, a total of 783 K sentences were extracted from SmPC and PL documents which were aggregated into a total of 284 and 129 semantic similarity clusters, respectively. The spread distribution among clusters shows separation into different cluster types. Examples of clusters with low spread include those with identical word embeddings due to current standardization, such as section headings and standard phrases. Others show minor linguistic variations, while the group with the largest variability contains variable wording but with significant semantic overlap. The sentence clusters identified could serve as candidates for further standardization of the PI. Moving from free text human wording to auto-generated text elements based on multiple-choice input for appropriate parts of the package leaflet and summary of product characteristics, could reduce both time and complexity for applicants as well as regulators, and ultimately provide patients and prescribers with documents that are easier to understand and better adapted for search availabilities.

Introduction

The product information (PI) is a vital part of any medicinal product approved for use within the European Union. It consists of the summary of products characteristics (SmPC) for healthcare professionals and package leaflet (PL) for patients, product packaging as well as an Annex which sets out conditions and restrictions for supply and the safe and effective use of the medicinal product. These documents, created from templates, explain how the product should be used and describe the expected benefits and risks associated with its use [1–3].

In 2020, following discussions with stakeholders, the European Medicines Agency presented key principles outlining a harmonized approach to develop and use an openly accessible and forward-compatible electronic format for product information (ePI) for human medicines across the EU. The ePI project mainly concerns the technical aspects of documents and aims to enhance accessibility of PI and promote best practices for creating product information for medicinal products [4]. Developing an electronic format with structured elements accommodates benefits brought by continuously evolving digital opportunities which will enable more efficient retrieval of information and facilitate the use of alternative e-platforms. As ePI can be read by machines, ePI information can flow to other systems such as electronic health records and e-prescribing systems. This will hopefully facilitate targeted delivery of the right information to the right user at the point of need.

In addition to the technical aspects, there will be challenges remaining after the expected output from the ePI project has been delivered. Several of these are related to the language itself contained in the documents, which are created by applicants and regulators in an iterative process to obtain a marketing authorization. Although there are current linguistic PI standards implemented manually for certain aspects of style, terminology and use of abbreviations, substantial variability is introduced between medicinal products in sentences where identical messages could be communicated. This process is time-consuming, may cause difficulties in search functions, creates uncertainty as to whether the content of different sets of PI is identical, and could potentially result in an increased risk of medication error [5, 6]. Whereas standardised messages could and should be used already today, a digitalization of the format will facilitate the use of exact standards, as well as to offer possibilities to streamline, simplify and speed up the regulatory processes.

The field of natural language processing (NLP) has been rapidly evolving since the introduction of transformer models [7]. These models are, without supervision or human imputation, trained on huge corpuses of text and contains high-dimensional semantic word embeddings based on textual context. In 2018, Devlin et al presented BERT, a Bi-directional Encoder Representation from Transformers model with 340 million parameters, which became state of the art in the field of NLP and has found many applications and undergone adaption to specific tasks [8]. Lately, BERT has been outperformed—particularly in creative NLP applications—by the much larger generative language models such as GPT-3 which have hundreds of billions of parameters [9]. However, the BERT architecture has remained a well-performing alternative for tasks such as classification and needs substantially less computing power.

NLP techniques have previously been applied to extract information from FDA medicinal product information [10]. Also, there is previous work on model-based standardisation of clinical information and lexical simplification of technical terms [11–13]. In this study, based on the English corpus of the EMA PI documents for all centrally approved medicinal products within the EU, we use a BERT sentence embedding model together with clustering and dimensional reductional techniques to identify PI sentence similarities that could be standardized, for the benefit of patients, prescribers, and marketing authorization holders alike.

Materials and methods

Text acquisition and pre-processing

The text corpus was compiled on May 3, 2022, by scripted downloading of all available English language PI files for all centrally approved medicinal products within the EU, from the EMA website. PL and SmPC documents for each medicinal product, excluding multiplicate documents for medicinal products with more than one strength or pharmaceutical preparation, were used. The PDF files were scraped using the pdfplumber version 0.6.1 package in Python 3.8.10 to extract all text except page numbering, headers, and footers.

Line breaks and special characters (excluding punctuation characters) were removed, and punctuation was added to sentences where this was missing (such as headings) to avoid false aggregation. All paragraphs were tokenized on a sentence level using the NLTK version 3.7 tokenizer and filtered to exclude sentences shorter than three words. The complete data processing pipeline is shown in Fig 1.

[Figure omitted. See PDF.]

All sentences were embedded, i.e., transformed into a 768-dimensional output vector, using the pre-trained SBERT model all-mpnet-base-v2. Detailed information about the model is available in the model card [14].

Clustering and dimensionality reduction

DBSCAN from scikit-learn version 1.0.2 with ε = 0.45 and a minimum cluster size of 50 was used to aggregate adjacent sentence embeddings into similarity clusters within the full embedding space. For visualization purposes, projection algorithms were explored to reduce the high-dimensional space to a flat 2-d projection. Due to the high dimensionality of the data and expected shape of clusters, t-SNE from [15] scikit-learn version 1.0.2 was chosen over UMAP from umap-learn version 0.5.3 [16]. t-SNE was applied using principal component analysis initialization, was run for 500 iterations with a perplexity of 20, learning rate of 200 and random state 23. To allow further analysis of cluster shape and spread, cluster centroids were calculated using K-means from scikit-learn version 1.0.2.

Results

A total of 1258 medicinal products were initially included in the study, of which 5 were subsequently excluded due to document compatibility issues. From these, a total of 783 K sentences were extracted from PL and SmPC documents. The length and distribution of sentences among subsections is illustrated in S1 File.

From the representations in full embedding space, PL and SmPC sentences were analysed separately, generating a total of 129 and 284 similarity clusters, respectively (Figs 2 and 3).

[Figure omitted. See PDF.]

[Figure omitted. See PDF.]

Although the mean Euclidian distances used to estimate the spread of individual clusters in space should be analysed with caution, as interpretability can decrease with increasing dimensionality, the distribution among clusters (Fig 4) indicates separation into different cluster types depending on the level of linguistic variability. Examples with low spread—i.e., low variability—include those with identical embedding due to current standardization such as section headings and standard phrases. Others show minor linguistic variations, while the group with the largest variability contains variable wording but with significant semantic overlap at least on a thematic level. Examples from a range of categories with the corresponding cluster characteristics, are presented in Table 1.

[Figure omitted. See PDF.]

Cluster characteristics for PL (left) and SmPC documents (right). Distribution of total cluster size (top) and number of unique sentences (middle). Cluster spread histogram (bottom) shows mean Euclidean distance from cluster centroid, and suggests separation into different cluster types with regards to variability.

[Figure omitted. See PDF.]

Up to ten variants listed per cluster.

Discussion

Natural language processing holds great promise to generate value in the field of pharmaceutical regulatory science, including both drug development and formal regulatory processes. Here, we illustrate a relatively straight-forward approach to create meaningful sentence-level numerical representations of the nearly one million sentences from all EU centrally approved SmPC and PL documents, to facilitate both future standardization as well as NLP research. In addition to our similarity clusters, we provide a freely available database containing a total of 783 K sentence tokens that are indexed and mapped to the medicinal product, document type and document subsection from which they originated.

We chose to analyse SmPC and PL documents separately, in line with the regulatory practise to compile these documents in separate processes tailored for their respective group of professional and non-professional end users. However, given that many messages are shared across the two types of documents, the linguistic relationship and potential for a more unified workflow should be explored in the future.

Although not specifically trained on a corpus of medical or pharmaceutical language, the BERT sentence embeddings allowed for semantic analysis of medicinal product information documents. This could to some extent be a result of regulatory language standards, making these documents contain far less medical specialist terms compared to health records. However, a language model trained on a corpus including medical and regulatory text would likely perform even better.

With the parameters for clustering that were used in this project, a total of 15% of sentences in the SmPCs and 23% of sentences in the PL documents were assigned to a semantic cluster. The absolute level of clustering should be interpreted with caution as it is highly dependent on what level of in-cluster similarity and minimum cluster size is sought, but the relative difference in clustering rate between SmPCs and PLs indicates that the PLs have a lower degree of linguistic variability.

By mapping the complete PI language space for European centrally approved medicinal products, we can show that there is a high level of semantic similarity in substantial parts of the documents where there currently is no fixed standard, or where standards exist but supporting systems should be developed or vocabularies used more frequently.

Semantic similarities may also point out the need for a certain structure for expanded electronical use, where existing standard sentences could be automatically populated or linked via reference data vocabularies. Such format would facilitate the use of existing established standard references, already translated into all EU languages. Where information such as “No dose adjustment is needed/No dose adjustment in children is needed” has been found, it may not need harmonisation as such, but rather electronic structure connected to both standard and subheading, i.e., both wording and place in the document will be important when forming structured standards.

Fields where standardized sentences are of great value, such as storage precautions and certain warning texts have already been harmonised by presenting standard sentences in EMA guidelines. The approach developed in this study could be used for further standardization, potentially valuable for regulatory processes as well as for down-stream use of the product information. Nevertheless, although the sentence clusters identified could serve as candidates for future standardization, the PI will always contain unique product-specific statements.

In the current approach only centrally approved medicinal products were included, which are all fairly new and where a single version of product information is used as a base for translation in all EU member states. Expanding the analysis to products approved in decentralized or national procedures, where documents by procedural nature vary between products of even the same active substance or therapeutic class, could provide further insight and further the potential for standardisation.

Conclusion

In this study, we have shown that currently available data science tools can identify semantically similar statements in medicinal product information. This could serve as a basis in a future process of standardisation, which would allow automation for parts of the PI compilation process. Moving from free text human wording to auto-generated text based on multiple-choice input for appropriate parts of the summary of product characteristics and package leaflet would reduce both time and complexity for applicants and regulators, and ultimately provide patients and prescribers with documents that are easier to understand and better adapted for search availabilities. For the foreseeable future, it will remain essential that the final documents are assessed by domain experts at the competent authorities involved, keeping the human in the loop throughout the process.

Supporting information

S1 File. Product information document characteristics.

https://doi.org/10.1371/journal.pone.0275386.s001

Acknowledgments

Disclaimer: The views expressed in this article may not be understood or quoted as being made on behalf of or reflecting the position of the European Medicines Agency.

Citation: Bergman E, Sherwood K, Forslund M, Arlett P, Westman G (2022) A natural language processing approach towards harmonisation of European medicinal product information. PLoS ONE 17(10): e0275386. https://doi.org/10.1371/journal.pone.0275386

About the Authors:

Erik Bergman

Roles: Data curation, Formal analysis, Investigation, Validation, Visualization, Writing – review & editing

Affiliation: Swedish Medical Products Agency, Uppsala, Sweden

Kim Sherwood

Roles: Validation, Writing – review & editing

Affiliation: Swedish Medical Products Agency, Uppsala, Sweden

Markus Forslund

Roles: Validation, Writing – review & editing

Affiliation: Swedish Medical Products Agency, Uppsala, Sweden

Peter Arlett

Roles: Writing – review & editing

Affiliation: European Medicines Agency, Amsterdam, Netherlands

Gabriel Westman

Roles: Conceptualization, Formal analysis, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft

E-mail: [email protected]

Affiliations Swedish Medical Products Agency, Uppsala, Sweden, Department of Medical Sciences, Uppsala University, Uppsala, Sweden

https://orcid.org/0000-0001-9402-772X

References

1. EMA. How to prepare and review a summary of product characteristics. In: European Medicines Agency [Internet]. 17 Sep 2018 [cited 16 Aug 2022]. https://www.ema.europa.eu/en/human-regulatory/marketing-authorisation/product-information/how-prepare-review-summary-product-characteristics

2. EMA. Product information: Reference documents and guidelines. In: European Medicines Agency [Internet]. 17 Sep 2018 [cited 16 Aug 2022]. https://www.ema.europa.eu/en/human-regulatory/marketing-authorisation/product-information/product-information-reference-documents-guidelines

3. EMA. Product-information templates—Human. In: European Medicines Agency [Internet]. 17 Sep 2018 [cited 16 Aug 2022]. https://www.ema.europa.eu/en/human-regulatory/marketing-authorisation/product-information/product-information-templates-human

4. EMA. Electronic product information for human medicines in the European Union—key principles. In: European Medicines Agency [Internet]. 30 Jan 2019 [cited 16 Aug 2022]. https://www.ema.europa.eu/en/electronic-product-information-human-medicines-european-union-key-principles

5. Fuchs J, Hippius M, Schaefer M. Analysis of German package inserts. CP. 2006;44: 8–13. pmid:16425965

6. Goedecke T, Ord K, Newbould V, Brosch S, Arlett P. Medication Errors: New EU Good Practice Guide on Risk Minimisation and Error Prevention. Drug Saf. 2016;39: 491–500. pmid:26940903

7. Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, et al. Attention is All you Need. Advances in Neural Information Processing Systems. Curran Associates, Inc.; 2017. https://papers.nips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html

8. Devlin J, Chang M-W, Lee K, Toutanova K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). Minneapolis, Minnesota: Association for Computational Linguistics; 2019. pp. 4171–4186.

9. Brown TB, Mann B, Ryder N, Subbiah M, Kaplan J, Dhariwal P, et al. Language Models are Few-Shot Learners. arXiv; 2020. http://arxiv.org/abs/2005.14165

10. Shi Y, Ren P, Zhang Y, Gong X, Hu M, Liang H. Information Extraction From FDA Drug Labeling to Enhance Product-Specific Guidance Assessment Using Natural Language Processing. Front Res Metr Anal. 2021;6: 670006. pmid:34179681

11. Kreimeyer K, Foster M, Pandey A, Arya N, Halford G, Jones SF, et al. Natural language processing systems for capturing and standardizing unstructured clinical information: A systematic review. Journal of Biomedical Informatics. 2017;73: 14–29. pmid:28729030

12. Pathak J, Bailey KR, Beebe CE, Bethard S, Carrell DS, Chen PJ, et al. Normalization and standardization of electronic health records for high-throughput phenotyping: the SHARPn consortium. J Am Med Inform Assoc. 2013;20: e341–e348. pmid:24190931

13. Buysschaert, Joost. The development of a MeSH-based biomedical termbase at Hogeschool Gent. In: Zweigenbaum, Pierre and Schulz, Stefan and Ruch, Patrick, editor. LREC 2006 workshop on acquiring and representing multilingual, specialized lexicons: the case of biomedicine. ELDA; 2006. pp. 39–43.

14. sentence-transformers/all-mpnet-base-v2 · Hugging Face. [cited 16 Aug 2022]. https://huggingface.co/sentence-transformers/all-mpnet-base-v2

15. van der Maaten L, Hinton G. Visualizing Data using t-SNE. Journal of Machine Learning Research. 2008;9: 2579–2605.

16. McInnes L, Healy J, Saul N, Großberger L. UMAP: Uniform Manifold Approximation and Projection. JOSS. 2018;3: 861.