As artificial intelligence (AI) increasingly permeates healthcare and scientific communication, a solid grasp of AI terminology is beneficial for medical publication and communication professionals.
This article developed by the ISMPP AI Task Force explains the most important terms, offering an accessible overview of key AI terms you’re likely to encounter in discussions, research, and applications related to medical publications. While the world of AI is vast and complex, the aim of the AI Task Force is to demystify some of the most frequently used concepts, enabling you to navigate this new frontier with knowledge and increased confidence. You’ll notice certain terms in bold; these are also defined in the comprehensive AI Lexicon, a valuable resource available through the ISMPP AI web page.
What is the difference between Machine Learning, AI, and GenAI?
Artificial Intelligence (AI) is a general term for the ability of machines or computer systems to perform tasks that typically require human intelligence. This can include understanding language, recognizing images, solving problems, or making decisions. AI works by machine learning: computers learn from data, get better over time, and make decisions on their own. It involves training models with lots of data to, for example, create text that reads like a human wrote it, or interpret X-rays automatically, or both (multi-modal systems). This can greatly help with tasks like making medical documents and clinical reports and serve as an assistant to clinicians.
Generative AI (GenAI) is a specific form of AI, where an AI system creates new content, in contrast to, for instance, ‘just’ classifying input. Systems like GPT-4, Gemini, and Claude break text into tokens (small pieces of language) and use inference to guess what comes next. These kinds of systems have a particular design (transformer models) and are called Large Language Models (LLMs). But they don’t always get it right and sometimes fall into what’s called an hallucination, where they confidently generate completely wrong or made-up information. To reduce this, many systems now use retrieval-augmented generation (RAG), which allows them to pull in data from external sources to improve accuracy. This way there is less dependency on just the knowledge incorporated in the LLM, which, by the way, is outdated quickly, as training these systems is not done frequently because of the costs involved.
Do these GenAI systems just generate text and images?
Definitely not! A number of different input and output modalities are now commonplace: text, images, video, sound, and any combination of these, both as input and output. A few examples:
- Image with text to video: A system taking an image with a description of what happens next is used to generate a video of the scene described, or a PowerPoint slide deck converted to a podcast
- A chat companion: you talk to a system that talks back, or generates a document or an image based on what you say.
- Google’s NotebookLM allows various input sources that will be used to answer questions, and the latest OpenAI model supports talking with it, so speech input and output, combining it with uploading images to be discussed, for instance
The capabilities of these systems continue to expand rapidly, integrating more senses and modes of interaction beyond the traditional text-based interactions. As the technology progresses, we’re seeing GenAI move towards more natural, human-like communication and creative expression across a diverse range of mediums
Can GenAI systems do more than generate content?
GenAI systems can be trained to output text in any language, including computer programming languages. If that generated program is then executed, such systems do more than that: they also ‘do’ things. Such systems are called agents. There are frameworks designed to support the interaction of multiple intelligent agents working together to solve complex problems or perform tasks: multi-agent frameworks. There is no central control in such systems: each agent is independent and has its own role, knowledge, and capability.
Through these techniques, AI systems become more and more versatile and powerful. Here comes Artificial General Intelligence (AGI), a type of AI capable of performing any intellectual task a human can do, which might one day match or surpass human intelligence. This is, however, a topic of much controversy as some top experts disagree that this will happen (at least, within our lifetime).
LLMs are pretty interesting, can you tell me more?
LLMs are neural networks, currently mostly used for sophisticated text processing called Natural Language Processing (NLP). It is all about teaching computers to understand and generate human language. They are trained on vast amounts of text to handle tasks like translations or answering questions. These models rely on context windows (how much text they can consider at once) and embeddings (how they represent words in relation to each other). Transformer models are a particular neural network design focusing on relationships between words, helping systems understand the meaning behind what we say or write. They typically have two parts: one understanding the input text (encoder) and one generating the output text (decoder). Using just the encoder part allows a search using semantics rather than text similarity (semantic search). These networks are huge: the number of nodes runs into the hundreds of billions, and each vertex has a weight assigned to it (parameters). These huge LLMs are also often called foundation models, as they are so versatile and can perform a wide variety of tasks. Since they mimic human behavior, many users start to project human traits on them (anthropomorphism).
How do these LLMs learn?
The size of the training set used to be a bottleneck, as it required human-labeled text (supervised learning). Using it without labels was easier but resulted in less accurate results (unsupervised learning). With the advent of self-supervised learning, the best of both worlds was used: an AI system creating its own labels (a form of synthetic data generation). What surprised researchers is that these large systems started to show behavior they were not explicitly trained for (emergent behavior). If privacy is an issue, federated learning lets models train without moving sensitive patient data around.
Once a model is trained on lots of (generic) text, it is called pretrained, and it can do a variety of tasks pretty well. To teach it a particular task, you can finetune it: further train it to do a particular task (transfer learning). That’s great, but since these foundation models are so large, when they are used for quite specific tasks, they are quite expensive to run. A way to remedy that is to create a smaller student model using a large finetuned model as a teacher (a process called distillation). Lately, this technique is becoming more and more popular also for more general tasks, as running inference using these huge models becomes prohibitively expensive. The performance of well-trained student models comes close to that of their teachers while using them is much cheaper and faster.
A model trained on a particular training dataset derived its knowledge from that dataset. If that training set contains, for instance, human prejudices, the model will have such bias as well. This can show up in how AI models make decisions, especially in sensitive areas like healthcare or hiring. That’s where Explainable AI (XAI) comes in—it’s about making AI decisions more understandable and accountable.
What new methods are shaping AI training and deployment?
Transfer learning and reinforcement learning are driving a lot of progress in AI right now. With transfer learning, models that have already been trained on one task can be adapted for something new, saving time and resources. Reinforcement learning helps AI make better decisions over time by rewarding successful actions. We’re also seeing neural networks with huge numbers of parameters—essentially the settings that help models make predictions—leading to surprising emergent behaviors (things they learn that weren’t explicitly taught). Lastly, synthetic data generation is becoming a popular way to train models without real-world data, especially in areas like healthcare, where privacy is key. Furthermore, models are so data hungry they run out of training data (the internet is too small). Generating new data avoids that limit.
I heard about Virtual Health Assistants, how is AI changing healthcare?
AI is starting to make waves in healthcare through Virtual Health Assistants (VHAs), which use natural language processing and multi-modal abilities (handling text, images, and more) to communicate with patients. VHAs enhance healthcare by improving convenience, patient engagement, and efficiency. They automate tasks like scheduling, data entry, and diagnostics, allowing healthcare providers to focus more on patient care. VHAs also help monitor chronic conditions, offering personalized alerts and insights, which can lead to timely interventions. Additionally, they play a role in mental health support through therapeutic interactions and assessments.
While VHAs complement rather than replace human workers, they transform job roles, creating opportunities in AI healthcare and data analysis. Sustainable integration requires ethical considerations, including workforce upskilling and data security. With continued advancements, VHAs are set to further revolutionize healthcare by offering personalized, efficient, and accessible care while addressing challenges like data protection and bias.
And how about the effect on Medical Communications?
This is mainly speculation although one can extrapolate from current trends. Medical writers will remain responsible for a manuscript but are expected to use GenAI in creating it. They are expected to deliver faster, and likely more diverse content (i.e., a plain language summary – PLS – as standard deliverable). That means writers will have to become expert users of GenAI, in all its facets, from quality control (think of completeness of references) to more sophisticated feedback (e.g., are the conclusions justified by the results?).
And Publishing?
Speculating again, since the review process is currently such a bottleneck, medical publishers will look to GenAI to make that process more efficient, e.g., adding an automated reviewer (like the grammar checking we’ve all used already for a long time), and provide additional tools for reviewers like “chatting with the manuscript” and reference checking and searching tools (“why did the authors not cite X?” or “why did they cite Y?”). We can expect that the time from submission to publishing will go down in the future.
On the other hand, there will be a flurry of low(er) quality papers submitted, and some will get through; the result being a larger number of publications with a decreasing average quality.
Acknowledgement: Article development was led by Bob Schijvenaars, Senior Vice President, Technology, Digital Science, in his role as a member of the ISMPP AI Task Force. The article was developed under the auspices of the ISMPP AI Task Force.