Oxford PharmaGenesis

Have you ever wondered what can cause a publication to ‘go viral’? Natural language processing (NLP) enabled us to analyse the growth of a very popular scientific journal article (a recent clinical trial of dostarlimab) by summarizing within minutes the thousand news articles that mentioned it. We found that these articles used terms of exaggeration (e.g. ‘revolutionary’, ‘unprecedented’, ‘scientific miracle’), which suggests that over-inflating results in news reporting may be a method to achieve popularity (a strategy we would not recommend!). The timely analysis of the vast quantity of these articles was made possible by NLP, a technique that is well suited to processing large volumes of text.

Several million scientific publications are now added to PubMed each year, compared with about 300 000 in 1980. The volume of health-related text (including publications, but also electronic health records) has grown exponentially with time: the resulting volume can only be analysed fully and efficiently with the help of automation. NLP helps pharmaceutical companies and other researchers to derive insights from this flood of health-related information, ultimately to benefit patients.

Considerable progress has been achieved in NLP in the last thirty years. Early approaches to NLP, which focused on the logical application of rules, failed to capture linguistic complexity and have proven limited in their usefulness. Using artificial neural networks, modern NLP performed well at ‘understanding’ language, by creating abstract representations of words and their contexts. This has opened up applications for healthcare, and medical and pharmaceutical research.

Exciting applications of NLP for health

NLP can summarize documents, focusing on the topic (topic modelling) or the opinion and emotional tone (sentiment analysis). Topic modelling can quickly condense themes in the published research on a therapy area. This application can inform the publication strategies of research institutions and pharmaceutical companies by enabling them to choose whether to contribute to common themes or to address potential literature gaps. Sentiment analysis can determine whether a publication was cited by others to confirm or to refute it, or for information only; such information augments a simple citation count to enable better judgement on publication reliability.

"Aside from publications and electronic health records, NLP tools can also analyse non-traditional text sources such as social media posts to monitor the experiences of patients with disease and treatment"

NLP can also be used to extract data from publications, such as individual entities (via a method called named-entity recognition) or relationships between entities (through relationship extraction). For example, named-entity recognition could extract the clinical outcome of a trial treatment arm, and relationship extraction could link this outcome to the drug and dosage used in that arm. This may accelerate the data extraction phase of systematic literature reviews, which are the cornerstone of evidence-based medical decision-making.

Avoiding the pitfalls of biomedical NLP

The successful biomedical application of NLP requires expertise in both NLP and medicine to ensure that we ask the right questions and interpret the answers correctly; however, such expertise is scarce. Successful implementation of NLP in the biomedical sphere may require the services of consultants who have the interdisciplinary knowledge to understand, to implement and to communicate biomedical NLP findings.

In particular, the interpretation of NLP-derived outputs needs special care. Modern NLP approaches require training data from which correlations are first derived, and then extrapolated onto new data to make predictions. Consider the example of a tool trained to extract drug dosage from clinical trial publications. If drug dosages are always presented using one unit (e.g. mg) in training data, NLP might be unable to extract dosage presented with another unit (e.g. µg/mm²). This reliance on good training data limits the power of NLP; if the training data contain mistakes, these will be reproduced in NLP-derived outputs. Awareness of these limits by interdisciplinary experts enables appropriate interpretation and communication of results, building trust in the true potential of NLP.

A promising future for biomedical NLP

Aside from publications and electronic health records, NLP tools can also analyse non-traditional text sources, such as social media posts, to monitor the experiences of patients with disease and treatment. For example, social media listening can be used to inform drug development by evaluating the impact of chronic disease on the quality of life of patients. A 2019 study used NLP to identify and to evaluate social media accounts of life with dry eye disease, which yielded insights into the needs of patients that have yet to be met by the pharmaceutical industry.

Pharmaceutical companies may also find new opportunities for NLP application. Information derived from NLP can be augmented with existing biological and clinical databases (e.g. UniProt for protein information, ClinicalTrials.gov for clinical trial data) to find new relationships between data. Such relationships could link an existing drug with a novel therapy area, which may constitute a repurposing opportunity.

In the last few years, considerable progress has been made in the field of NLP and we are now seeing powerful medical applications emerging. The biomedical opportunities offered by NLP are vast; to stay ahead, pharmaceutical companies should embrace NLP and benefit from the support of interdisciplinary experts to help avoid potential pitfalls.