What AI promised medical writers before generative models

Before generative models, AI was expected to take the routine work off medical writers: retrieving and transferring content between documents, finding and summarising the literature, compiling metrics, checking and formatting documents, and drafting reports and narratives straight from the data. Natural language processing and generation carried that promise, and sponsors piloted natural language generation on clinical study reports and patient safety narratives.

Computer-assisted documentation

In 1992 government and industry representatives from Europe and the USA discussed the development of computer-assisted marketing applications for new drugs (CANDA), whose advantages included better in-house documentation systems, better communication between companies and authorities, and less paper. ¹ In 2023 EMWA's president judged those advantages still valid. ¹

AI's expected reach into medical writing

By 2019 AI had started to affect the work of medical writers, and it was described as able to help them in regulatory writing, medical affairs, redaction and publishing. ² As the field advanced rapidly, writers were advised to be prepared for the time when AI-based innovations would knock on their door daily. ³

In 2019 AI was also redefining the clinical trial landscape, with repercussions for scientists, physicians, clinical staff, patients and medical writers. ⁴ Many pharmaceutical companies were investing in natural language processing tools for complex data analysis and interpretation, study design, and the documentation each step of clinical research generates. ⁴

Natural language processing

Natural language processing is the AI technology that helps computers understand the natural language of humans, and in 2019 it was already part of daily life in machine translation, email autocomplete, word processors and grammar checkers, and voice assistants. ⁵ Yahoo! was generating millions of automatic reports and match recaps for its fantasy football fans in a personalised manner, and writers and journalists had started to experiment with AI to help them generate original writing. ⁵

Text mining, which uses natural language processing, analyses language to the lexical level and can extract detailed information, reveal patterns across millions of documents and summarise large volumes of information automatically. ⁴ Because natural language processing tools can read unstructured data, they can scan metadata, perform tailored literature searches and accurately extract targeted information from vast unstructured databases. ⁶ AI algorithms could analyse networks as well as data, text and images. ⁷

Authoring structured regulatory content

In 2019 natural language processing algorithms might soon help regulatory writers replace one of their most time-consuming, tedious tasks: authoring structured content, which means retrieving information from various documents in order, for example, to transfer it with minor modifications to another place in another document. ⁵ A large portion of clinical protocols and clinical study reports could be created automatically by AI tools in a matter of hours instead of weeks, so that the writer could focus on work that needs a higher level of scientific interpretation. ^{5, 3} Such tasks could be done with the highest transparency and security of personal data, because no human eye would have seen the confidential information. ⁵ That level of productivity, confidentiality and consistent compliance with regulations was AI's big promise to regulatory writers. ⁵

The automation of standard documents decreases human errors and the need for clinical staff and writers to certify that information has been accurately entered in large patient databases. ⁶ That favours shorter early stages of clinical trials, results in a better selection of the patients taking part and promotes high quality control of the whole process up to market approval. ⁶

Searching the literature and clinical records

A growing number of articles are available in PubMed and hundreds of clinical studies in ClinicalTrials.gov, and it is basically impossible for a human being to keep up with the latest discoveries and relate the information on one study to hundreds of others already published. ^{6, 4} Writers could use natural language processing tools to scan metadata across different databases and identify key information that a classic keyword search cannot reach, with the potential to unravel new therapeutic options. ⁸ Natural language processing in clinical study design and pharmacovigilance was held to be revolutionary, because skilled writers would be able to gather and analyse a surplus of information quickly yet accurately. ⁴

AI tools could scan the medical literature and the scattered networks of physicians and scientists to create a list of key opinion leaders and the best strategy for contacting them, drawing on expert information, trend identification and sentiment analytics rather than simple web searches, and keeping the list current in real time. ⁷ They could also give writers and medical affairs staff essential information produced automatically by scanning databases, gathering scientific insights and producing summaries. ⁷

For freelancers, open-source software was appealing, and user-friendly natural language processing software offered a way to become familiar with what the future might look like. ⁹ SemEHR, Apache cTAKES, GATE and CLAMP are specialised in text analysis and in extracting target clinical information from electronic health records, unstructured clinical notes and narrative patient reports, and SemEHR also identifies contextualised mentions of biomedical concepts in clinical records and can pull out the cohort of relevant patients. ⁹

Metrics for medical writing managers

Medical writing managers need to collect metrics, create graphs, analyse data and make decisions to stay on top of their work or business. ^{10, 11} An AI-based system could perform these tasks automatically in real time and reduce the burden of doing them by hand. ¹¹ Metrics relevant to a writer's job include basic client information, departmental head count, skill sets, workload distribution, timelines and budgets. ¹¹ AI tools could collect and analyse every type of metric, even ones a human might at first have thought irrelevant, and then create graphs or narratives automatically, written differently for different readers: a chief executive may want to see different metrics from a line manager or a writer. ¹¹

Peer review and editing

More and more authors of scientific studies were frustrated with traditional peer review, among other things for its serious problems in fighting bias, its slow speed and its lack of transparency. ^{12, 3} AI tools could help with most of its laborious processes. ³ For example, AI-based programs could identify and control parallel communication strategies with reviewers until the required number of reviews had been reached. ³ Journals were testing tools to detect plagiarism, bad reporting and manipulated images, which could improve the publishing process if used with caution. ³ Editing could be automated with AI-based tools that control and amend documents to comply with required styles and formats. ³

Natural language generation

Natural language generation is programming that can turn data into language: a software process that collects raw data and transforms them into written natural human language. ^{13, 14} It is widely used by businesses and organisations in many sectors. ¹⁴ The generation of narratives from structured data had been in the life sciences for more than a decade by 2023, previously as text hardcoded into programs and by then through natural language generation models that generate the text up front for the structured data being processed. ¹⁵ Depending on the models used, such generation can involve machine learning (deep learning), preset parameters or both, and the output is a narrative as a medical writer would have written it. ¹⁵

Advances in AI and natural language generation were held in 2020 to have the potential to transform medical writing by automating routine content generation. ¹³ Using AI, natural language generation tools convert complex data sets into written narratives quickly and accurately, so they can be useful for text that needs no analysis or scientific interpretation. ¹⁶ Collating data from a source document such as a clinical study report into a manuscript's results section, as with adverse event tables, may introduce quality issues if not done carefully, and natural language generation provides tools to create narratives that summarise tabular data accurately. ¹⁷

Yseop's sales vice-president gave the benefits of integrating AI and natural language generation as a shorter time to market for new drugs, less stress for writers, a better ability to analyse complex data, a single voice, fewer quality reviews and less time needed to react to new requirements. ^{13, 18} He compared the technology to the spreadsheet: just as the spreadsheet freed data analysts from the repetitive work of collecting, analysing and reporting data, AI and natural language generation will free writers from tedious copying, pasting and formatting. ^{13, 18} AI and natural language generation solutions are customised to each client's report content standards, data sources, data analysis methods, text style and preferred interactive dialogue options. ¹³ Yseop's AI-powered application automated regulatory report writing, with natural language generation for clinical study reports, patient safety narratives and pharmacovigilance, and freed writing teams to focus on writing that needs scientific knowledge, interpretation and strategy. ^{19, 20}

Sponsor pilots on clinical study reports and narratives

Sanofi evaluated natural language generation as part of a strategy to shorten the time from last patient visit to a completed clinical study report, and implementing it reduced that time from a baseline of 6.8 months. ²¹ Its analysis of report content found that 55% was reused from other documents, 40% could be provided by natural language generation and 5% needed to be authored. ²¹ In a demonstration of Sanofi's tool, Business-Oriented Bot, its head of medical writing created the safety section of a report in a few minutes with a custom Word plug-in: one click of a table insertion tool populated the whole report with properly formatted tables, data and cross-referencing links, and a dialogue tool then generated short paragraphs summarising the safety data in seconds. ^{22, 18} Human intelligence, she emphasised, will always be required for meaningful clinical interpretation. ²²

Orion built a customised Yseop tool to produce the clinical summary of a client's patient safety narratives; the client's existing automated content tool had drawbacks including long lead times, an inability to perform calculations and an inability to use existing data outputs without modification. ²³ Through short development cycles of daily calls and weekly meetings, a pilot tool was in place after three months that calculated dates, aggregated continuous adverse events and flagged missing data. ²³ The writing was accelerated and the need for quality control reduced. ²³

In 2020 automatic content generation at Roche was part of an initiative aimed at reducing the critical path from database lock to filing from 20 weeks to 14. ²⁴ The areas to optimise included the output generation process, the regulatory publishing process, and data-dependent writing, review and approval. ²⁴ For narratives, Roche implemented SAS JMP Clinical's autonarratives software. ²⁴ Optimising the output the data keepers give through existing statistical programming software can minimise reformatting of the data and so save effort and time. ²⁴ In the view of a Roche regulatory documentation scientist, the role of a regulatory writer includes much more than writing: regulatory writers manage the entire writing process as strategy, content and process experts, and natural language generation tools that reduce the mundane tasks of producing documents can let them focus more on those other parts of the role. ^{24, 18}

According to staff of the natural language generation company Narrativa in 2021, automating clinical study reports, patient safety narratives and electronic common technical documents had let writers work faster and more efficiently, and pharmaceutical companies had substantially reduced costs and accelerated regulatory submissions, bringing new medical developments to market more quickly. ^{14, 25, 26}

An integrated platform for search, summarisation and generation

Authoring products that let several people collaborate on one document had been in use for more than a decade, alongside referencing and formatting to journal requirements, but these capabilities could come from various tools that would have to be integrated into a single system. ¹⁷ A 2023 design from a team that had built a cognitive search and medical writing platform at Infosys set out such a system, taking the scientific manuscript as its example. ^{27, 28}

The case for the platform

Medical writing is complex and time-consuming because the writer must refer to multiple sources, sift through many documents, maintain data integrity, review the literature, interpret results and summarise. ²⁷ The design holds that cognitive search, natural language processing and natural language generation can address these challenges and minimise them substantially. ²⁷ Given the recent advances in language models for generation, it holds that the time is ripe for a product that integrates and automates search, provides cognitive processing and generates content. ²⁷ In its view, AI's applications are nowhere more relevant than in life sciences, pharmaceutical, regulatory and medical writing, for documents such as scientific manuscripts, fact sheets, literature reviews, disease awareness material and oral posters. ²⁹ The literature review is a canonical use case, and natural language generation models were being used to automate its summarising, though manual intervention is required before machine-generated text can be submitted for publication. ²⁹

Entity recognition and recommendation

AI or machine learning can help automate the literature search, content extraction, content enrichment (which includes named entity recognition) and intent detection, enabling unified searching across several data sources and databases. ³⁰ Named entity recognition classifies text into predefined categories such as drug name, disease name or location, and can tell "apple" the fruit from "Apple" the corporation by context. ³¹ A recommendation model uses machine learning to predict the content relevant to a user in a given context, often combined with a ranking that shows the most relevant first. ³¹ Once trained, recommendation models identify the relevant sentences or paragraphs in the articles, which can make the most time-consuming part of the literature review more efficient. ³⁰ The system can also manage citations. ³⁰

How the platform works

The platform draws documents from external sources such as PubMed, Ovid and ClinicalTrials.gov, from regulatory documents such as the FDA's, and from the deploying organisation's own sources. ^{32, 33} During ingestion, document structure is extracted and tagged with metadata for indexing and classification, including classifying sentences, paragraphs and sections by the drugs, diseases and other biomedical terms they deal with. ³³ A user's question in plain English is translated into a machine-readable query and searched against the indexed documents, and the results are ranked by the rules and boosting criteria in use and returned as recommendations. ³³ Once the user selects documents, natural language generation produces extractive or abstractive summaries of the selection. ³³

The envisaged platform has these key features: ³⁴

• unified search across internal and external databases
• built-in deep learning models for article recommendation in pharmaceutical clinical trials, regulatory intelligence and medical research
• language models fine-tuned for life sciences for summarisation and generation
• real-time ingestion of structured or unstructured documents from varied sources, such as PubMed articles, the FDA, the European Medicines Agency, clinical study reports and protocols
• automatic document structure extraction, content enrichment and sentiment analysis
• dynamic document editing with scientific lexicons and ontologies
• workflows for collaborative authoring
• citations from a machine-generated summary back to its original sources
• templating of the final document

³⁴

Onboarding a manuscript template

An authoring template is a composite of individual sections, and onboarding each one entails data ingestion, article recommendation and content summarisation. ³⁴ A manuscript's introduction draws mainly on PubMed articles for disease description, epidemiology, burden of disease and a drug's mechanism of action; for it, ingestion indexes the relevant article sections such as the abstract and introduction, and pipelines classify sentences into those categories and run named entity recognition for diseases, drugs and molecules. ^{34, 32} For article recommendation, natural language query understanding pipelines are defined and search ranking rules and boosting criteria refined as required. ³² For summarisation, the platform evaluates the available language models against each section's needs, samples are curated for fine-tuning and pipelines are defined for later active learning. ³²

Expected gains and challenges

In the design, with pretrained language models such as GPT-3, GPT-J, BART and BERT fine-tuned for the biomedical domain, summarisation and final document generation could take a few days rather than a few weeks, even after the final manual review and approval. ³⁴ With all workflows automated, the scope for error would be minimised. ³⁴

The design names these key challenges in applying AI to medical writing: ³⁵

• ingesting documents of many formats and structures from internal and external sources
• understanding document semantics, extracting key entities unambiguously, capturing synonyms from scientific ontologies and identifying contexts and intents, which needs compositional semantic analysis including word sense disambiguation and relationship extraction
• re-ranking search results by learning to rank
• combining natural language processing, generation, deep learning, language models, lexicons and ontologies in one platform
• detecting and removing biases that can enter at various steps, from the curation of training data to feature engineering, model choice and implementation

^{35, 36}

The design concludes that AI and machine learning combined with natural language processing and generation promise to reduce the manual work of medical writing by automating many steps and to improve quality and reliability, and that the time and effort saved can be substantial for large organisations, which often spend a considerable amount of both during a drug's lifecycle. ²⁸