The 56^th European Medical Writers Association (EMWA) Hybrid Conference Day took place on 10 November, with delegates able to attend sessions on Writing in Plain Language for Publications, and Artificial Intelligence (AI) in Medical Writing. The seminars covered these topics through presentations and panel discussions, with the aim of providing practical recommendations and advice to medical writers and communicators. A summary of the second session on AI in Medical Writing is provided below to benefit those who were unable to attend, and as a timely reminder of the key topics for those who did.

Our summary of the session on Writing in Plain Language for Publications can be found here.

---

KEY TAKEAWAY

• The advent of AI in medical communications will likely lead to the repurposing of medical writers; however, it should never replace them.

---

An introduction to AI – AI 101

Martin Delahunty (Inspiring STEM Consulting) and Sepanta Fazaeli (Stryker) co-presented the first presentation of the afternoon covering some basic concepts in AI and their relevance in the medical writing landscape.

Large language models (LLMs) have become synonymous with AI, with OpenAI’s ChatGPT perhaps the most famous example of an LLM. In response to a prompt, LLMs produce human-like text by recognising patterns between words and predicting what comes next. However, AI models do not ‘understand’ human language; they work by using a mathematical process to convert subcategories of a sentence (tokens) into a matrix of numbers. The more massive the matrix (ie, the larger the dataset on which the AI model is based), the more understanding and context it contains. Once converted, AI models run arithmetic functions on these numbers to process and produce an output.

Medical writing involves different types of tasks, eg, ranking and classification of data, information appraisal, information retrieval and extraction, and report generation. Some AI models are better suited to certain tasks than others. For example, AI models based on Bidirectional Encoder Representations from Transformers, can be better suited to extraction tasks compared with ChatGPT, which is a unidirectional generative AI tool. Nevertheless, LLMs can be adapted for specific tasks. In the example above, ChatGPT would need to be fine-tuned before being used for extraction tasks to avoid creating nonsensical or inaccurate outputs (known as hallucinations).

Some AI models are better suited to certain tasks than others. It is important to use the correct AI model for the task.

Delahunty and Fazaeli outlined some potential advantages of using AI tools in the medical writing landscape:

• AI tools can streamline processes, increasing efficiency.
• Using AI can increase the speed at which projects can be completed.
• AI allows the user to process vast data sets.
• Using AI can improve accuracy and consistency.

On the other hand, there are situations where using AI tools is either not appropriate or their use should be considered carefully. The International Committee of Medical Journal Editors (ICMJE) worked quickly to update their guidance on the role of authors and contributors with the advent of AI, emphasising the need for transparency and appropriate disclosure where AI tools are used in the production of biomedical publications. Yet there are still important and relevant issues to be addressed in the context of the biomedical publishing industry:

• Appropriate attribution of authorship/contributorship status needs to be considered when using AI.
• Using AI can have ethical implications.
• AI tools can produce incorrect or misleading outputs.
• There are limitations of current AI technologies.
• There are privacy considerations regarding the datasets used to train AI.
• Ownership of the outputs produced by AI can be a complex issue.
• Non-unique or duplicate outputs.
• Biased or offensive outputs.
• Detection of instances where AI use has not been disclosed.
• Amplification of AI papermills and fake image generation.

ICMJE have emphasised the need for transparency and appropriate disclosure where AI tools are used in the production of biomedical publications.

Harnessing the power of AI: an exploration of AI tools for the next generation of research

---

KEY TAKEAWAYS

• An understanding of the strengths and limitations of AI tools is fundamental to harnessing their potential whilst avoiding potential pitfalls.
• Medical writers need to be aware of the risks associated with AI to recognise where it can and cannot be used.

---

Avi Staiman’s (Academic Language Experts and SciWriter.AI) talk focussed on the potential uses and limitations of AI in medical communications. Current uses of AI include translating and editing of text, idea generation or brainstorming, summarisation, peer review, identifying potential funders, and data processing. Among these, AI-assisted summarisation can be particularly useful to medical writers due to the recent rapid growth of research literature. Staiman cautioned however that the body of research literature may be weakened if all research is written based on AI-generated summaries.

The body of research literature may be weakened if all research is written based on AI-generated summaries.

Staiman next addressed several misconceptions regarding the use of LLMs in the context of medical writing:

Misconception	Explanation
LLMs are a personalised Wikipedia	Answers provided by LLMs are not always accurate and should not be relied upon
LLMs work consistently	There is a lot of variability; the same prompt on different days can lead to different answers
LLMs either work or do not work	Using LLMs should be an iterative process
The use of LLMs can be detected accurately	To date, there is no peer reviewed evidence that the use of AI in writing can be detected
Using LLMs is plagiarism	This depends on how plagiarism is defined; if it is defined as taking other’s work without attribution, using LLMs is not plagiarism
LLMs will have a negative impact on medical writing jobs	The subject area knowledge that medical writers have is critical for prompting the process. LLMs may just make medical writers more efficient

Staiman discussed how to get the best out of AI tools by providing effective prompts. The output quality from an AI tool is determined by the input quality. Iteration is key to getting the best results, said Staiman and he outlined six rules to follow for prompting AI:

1. Tell it your role.
2. Tell it your goal.
3. Break the prompt down into step-by-step instructions.
4. Few-shot prompting, ie, provide the AI with examples of good practice.
5. Personalisation, eg, which journal you are submitting to and the appropriate tone.
6. Tell the AI the constraints, eg, only include citations from the last 5 years.

The output quality from an AI tool is determined by the input quality.

Staiman had some cautionary advice regarding the use of generic AI tools. Tools including ChatGPT, Midjourney, Gamma, and Jasper are appealing due to their power, ease and (free) cost of access, and readily available training materials. However, these generic AI tools are also associated with hallucinations, misinformation, problematic citations, research integrity issues and authorship concerns. There is also a potential confidentiality issue with using generic AI. For example, in the free ChatGPT model, inputted information is collected and used to train the AI. Staiman advised not to put propriety information into this version. In the paid version, a private server is used and the input is not retained for training.

In closing, Staiman thought that funders must consider the implications of AI in undermining peer review, research integrity, confidentiality and bias. However, they must also recognise the opportunities afforded such as sorting through applications, supporting open infrastructure and levelling the playing field for researchers whose first language is not English.

The best-case scenario is the science writing community embrace responsible AI tools and create forward thinking policies on best practices for AI.

The more things change, the more things stay the same – the history of AI

---

KEY TAKEAWAY

• Interest in AI technologies – and associated fears – may be at a peak. As understanding of AI increases and its issues addressed, we can expect gradual mainstream adoption of practical applications.

---

Beginning by explaining that fear of new technologies is not a new concept, Phill Jones (MoreBrains Cooperative) provided a brief history of AI and outlined the current state of the art. Jones described how, at present, we have reactive machines that process complex information and provide a response. We have also started to use limited memory applications in self-driving cars for example. However, emotional intelligence and theory of mind remain purely theoretical.

Emotional intelligence and theory of mind remain purely theoretical.

Outlining the current development direction for AI, Jones referred to three important areas of research for LLMs and GPT:

• Machine learning – The science of how computers solve problems with no or minimal human guidance.
• Computer recognition – How computers recognise and characterise objects.
• Cognitive computing – How computers make decisions.

Jones went on to discuss the concept of natural language processing (NLP), describing human language as unstructured data whose understanding depends on context. NLP is the process of converting between unstructured human language and structured data that can be understood by a computer. Jones covered several approaches to NLP including:

• Simple language parsing: A sentence is broken down and classified into types of words, eg, nouns and adjectives.
• Semantic processing: The functions of words in a sentences are encoded.
• Named entity recognition: Recognition of proper nouns.
• Sentiment analysis: Recognition of emotional context.
• Dependency passing: Identification of word hierarchies in a sentence and the understanding of how important one word is to another.

Jones continued by describing how machine learning approaches are being developed to allow computers to account for missing information. LLMs are sometimes referred to as a stochastic parrot, referring to the ability of LLMs to generate convincing text without understanding its meaning. Neural networks learn by being given problems with known desired answers. Models are then modified iteratively until they can generate satisfactory outputs for each problem. In this way, supervised fine-tuning and reinforcing learning can be used to tailor LLMs to particular use cases.

Machine learning approaches are being developed to allow computers to make a best guess based on incomplete information.

The ability of generative AI to write sensical but useless text may contribute to the amplification of papermills. Papermills target journals with weak review processes and LLMs will likely make it harder to detect papermill articles. Accordingly, Jones thought it important for journals to have better editorial oversight and more quality control processes. However, Jones also explained how AI can be used positively to tackle misinformation, citing the example of ChatBawa, an AI-powered chat bot that can monitor misinformation and check the veracity of public and social media claims.

Jones ended by showing the Gartner Hype Cycle for 2023 – a graphical depiction of the maturity of emerging technologies. According to the chart, AI has gone through the early phase of innovation and has reached the so called ‘peak of inflated expectations’. A natural waning of the hype surrounding AI might be expected in the short-term before the technology is more widely understood and adopted and we see a rise in people’s faith in AI.

Medical writers as guardians and leaders of medical communication

---

KEY TAKEAWAY

• Medical writers have a responsibility to define and implement ethical and accountable approaches to the use of AI in scientific communications.

---

In her presentation on ethical use of AI in scientific communications, Uma Swaminathan (GlaxoSmithKline) focussed on the key themes of ethics, accountability, and trust. Swaminathan detailed how scientific communication comprises a wide spectrum of activities, all of which require an ethical, accurate, and factual approach. She emphasised the importance of proof of ethics at every level and thought that, when used in the correct way, AI can support all these activities.

Ethicists are not here to stop the advancement of AI, but to look at how AI can be liberated to catalyse innovation and support scientific communication.

Swaminathan explained further that the data we work with is provided with trust by research participants for a broader good, and that employing AI in scientific communication raises legitimate concerns about how this data might be used. In this respect, scientific communicators are uniquely placed to act as guardians of an ethical approach to using AI.

Ultimately, Swaminathan thought that the responsibility to incorporate an ethical approach cannot rest with AI; human control and accountability are key. She proposed the following approaches to integrating AI into scientific communications ethically:

• Building capabilities to ensure AI is used correctly, taking a 70/20/10 learning approach: 70% learning on the job, 20% learning from the peer network and 10% formal learning.
• Adapting policies, systems, processes, and controls to integrate AI.
• Proactive risk management and robust governance.
• Defining and sharing best practices.

There is a lot of discussion, energy and focus around ensuring the ethical, appropriate use of AI. Ultimately, it is about maintaining trust, human accountability and explainability.

Is the hype real? Real-life user experience of an AI tool for clinical study report production

---

KEY TAKEAWAY

• TriloDocs is an AI-enhanced medical writing tool that can quickly collate a draft clinical study report from appropriate source documents.

---

In the next presentation, Lisa Chamberlain James (Trilogy Writing & Consulting) provided a medical writer’s perspective on using AI to produce a clinical study report (CSR). Chamberlain James introduced TriloDocs, an AI tool that can consolidate existing text and draft de novo content to generate an initial draft CSR.

TriloDocs uses source documents to populate the CSR sections in a TransCelerate template. The tool can write results text and populate tables for key study endpoints as well as write text to highlight clinically relevant data. The tool does not use machine learning or generative AI. Importantly, this means the tool cannot hallucinate – anything not included in the source documents needs to be added manually.

Chamberlain James highlighted key advantages and disadvantages of using an AI-enhanced medical writing tool such as TriloDocs:

Advantages	Disadvantages
Collates information quickly and accurately Reduces time taken to produce first drafts and key messages memos Negates the need for quality control of numerical data Frees up time for crafting text, applying critical thinking, and contextualising the data Reduces risk of overlooking important signals Guides the writer to think about study context by asking questions	Cannot populate information that is not in standard source material Cannot write the conclusions or discussion Cannot process unusual analyses

 Chamberlain James noted AI tools are dramatically changing how medical writers work; they allow writers to focus on the clinically meaningful messages and have discussions with clinical teams earlier. Furthermore, because AI cannot contextualise data in terms of what it means for patients, nor interpret new data, medical writers are still an essential part of the writing process for CSRs.

Because AI cannot contextualise data in terms of what it means for patients or interpret new data, medical writers are still an essential part of the writing process for CSRs.

Reimagine the future of medical and regulatory writing with generative AI and LLMs

---

KEY TAKEAWAY

• Copilot is a generative AI tool that can be adapted to support medical writers with a variety of repetitive tasks in the production of regulatory documents such as CSRs, patient narratives, and clinical trial summaries.

---

Enrica Cavedo (Yseop) and Dominique Mariko (Yseop) presented the final talk focussing on how generative AI could support scientific communications and showcased Yseop’s work in this area. Setting the scene, they noted that global demand for medical writers has increased in recent years due to unprecedented levels of drug development. This demand is only set to increase and will significantly outpace the availability of trained professionals in the field. Generative AI could increase the efficiency of medical writers producing regulatory documents, supporting writers with repetitive tasks, and allowing them to focus on using their expertise elsewhere in the process.

Generative AI could increase the efficiency of medical writers producing regulatory documents, supporting writers with repetitive tasks, and allowing them to focus on using their expertise elsewhere in the process.

Being able to navigate the different AI technologies suited for different purposes can help reduce writing time. Cavedo and Mariko described Yseop’s Copilot tool (which is built on pre-trained LLMs) that can generate reports for regulatory documents automatically in a few seconds. They explained how Yseop first evaluated which parts of clinical content generation could be automated, starting with CSR and patient narrative generation, with plans to move towards AI-assisted generation of clinical summaries and clinical overviews.

They outlined some of the ways the technology they have developed meets the needs of regulated industries including:

• Data security and confidentiality being embedded in the model. Unless specifically agreed, a user’s data is not kept for training and an in-house model is used so data does not travel all over the world.
• Self-quality metrics to provide human evaluation and ensure text is factual, accurate and informative.
• Traceability to remain compliant by ensuring everything needed to prove auditability is available to the end user.
• Control – configurations and narrative options allow users to get exact results.

Cavedo and Mariko concluded with Yseop’s stated aim to gain 80% productivity across the whole workflow, and by doing so, impact the productivity of the whole clinical content generation chain from data collection to submission.

Panel discussion

The seminar closed with a panel session chaired by Martin Delahunty, bringing together the speakers from the afternoon session. Key discussion points included:

• The balance between human input and automation: panellists concluded that the level of acceptable AI use depends on the level of risk of generating false outputs. Concerns were raised relating to the use of generative AI in regulatory documents.
• The lack of transparency surrounding AI tools: there is a general lack of understanding regarding why AI tools work. This fundamental opacity makes transparency difficult which in turn makes it difficult to trust the validity of outputs. Transparency will be dictated by the data used to train the models. The quality, traceability and transparency of the data will be reflected in the quality of the output. Medical writers need to be aware of the risks associated with AI to understand where it can and cannot be used.
• The use of AI tools when working with patient data: when using AI with sensitive data, the approach should depend on the robustness of the controls which need to be built into the tools. Some institutions use private servers only accessible to certain individuals. When dealing with patient data these tools should be General Data Protection Regulation-compliant.
• AI in informed consent forms: there is an obligation to inform research participants about the future use of their data, including that AI will be relied upon more and more. However, it is not possible to inform participants of all possible future uses of AI because this is unknown. Ultimately, it is important that the use of AI happens with governance and controls.
• Current use cases: there are several ways we can use AI already, including risk assessments, monitoring real world evidence in real time, adverse event reporting and complaint handling.
• Advice for new medical writers: the advice remains the same as always; gain experience, find a company that offers training and learn to write.
• Sustainability of AI tools: AI tools consume a lot of energy and this is increasing as they become more powerful. The benefits conferred needs to be assessed in relation to the power consumption. The panellists agreed that to remain sustainable, AI tools should only be used when they are effective and the right size tool should be selected.
• Spreading of misinformation due to hallucinations: the panel agreed that humans should always be involved in the final decision making and approval processes; this should not be left to AI. Regulatory authorities have a responsibility to prevent the spread of misinformation.

Why not also read our summary of the morning session on Writing in Plain Language for Publications.

——————————————————–

Written as part of a Media Partnership between EMWA and The Publication Plan, by Aspire Scientific, a proudly independent medical writing and communications agency that believes in putting people first.

——————————————————–