Large Language Models (LLM) like ChatGPT have taken the world by storm, with everyone from researchers to tech enthusiasts excited about the almost limitless potential of these models. And yes, I’ll admit it that I too am jumping on the ChatGPT hype wagon—the opportunities are simply too big to ignore.
These Artificial Intelligence (AI) models can process and understand natural language at a previously unimagined level. This ability allows them to perform a wide range of tasks, from language translation to text summarization to conversation. Within the realm of clinical research, this opens up a host of possibilities to improve patient care and advance medical knowledge.
Health Benefit Consultants, Share Your Expert Insights in Our Survey
Derk Arts Derk Arts MD, PhD has over fifteen years of experience in medicine, research and technology. He founded Castor to solve the biggest issues in clinical research: a lack of inclusivity, patient focus and impact of data. Castor enables sponsors worldwide to run patient-centric trials on a unified platform, that helps them maximize the […]
I think I have spent about 15 hours being “productive” with ChatGPT by now, and I have to say, its capabilities do not cease to amaze me. At the same time, I have seen it make many mistakes. An important aspect of successfully implementing LLMs in any workflow is knowing where they are weak or strong and how to direct them for the best output.
ChatGPT’s potential use in engagement, screening, and more
As a conversational AI-powered bot, ChatGPT may be used in future clinical research to engage in conversation and understand participant responses. This ability could then could be used to improve engagement and screening for clinical trials. For example, an LLM could be combined with a digital avatar to have a recorded voice or video chat with a potential trial participant. Next, that recording could be fed into a validated digital biomarker for the desired indication. The output from this workflow could then be used to reduce the size of the recruitment funnel, thus removing a lot of pressure on sites running the screening process. This approach may work well in clinical studies for Alzheimer’s, depression, and Parkinson’s disease.
Even outside of the conversational domain, these models are impressive. I have already seen examples of LLMs supporting SDTM data transformations—a subset of tasks that I believe these models will become exceedingly good at. Mark my words, this technology is going to drastically change a number of data management activities in the next few years!
A Deep-dive Into Specialty Pharma
A specialty drug is a class of prescription medications used to treat complex, chronic or rare medical conditions. Although this classification was originally intended to define the treatment of rare, also termed “orphan” diseases, affecting fewer than 200,000 people in the US, more recently, specialty drugs have emerged as the cornerstone of treatment for chronic and complex diseases such as cancer, autoimmune conditions, diabetes, hepatitis C, and HIV/AIDS.
Combined with other systems that use more structured data, such as Wolfram Alpha, ChatGPT could even become a powerful tool for data analysis and decision-making in medicine. For example, a combination of these two systems could be used to answer participant questions about a protocol or informed consent form. It could even help them complete their questionnaires.
Potential risks of ChatGPT in clinical research
We need to be explicit about the challenges and limitations of technologies like ChatGPT as we bring them into the real world, particularly in healthcare and clinical research. LLMs are trained using massive amounts of text data, often sourced from the internet. While this approach has led to impressive capabilities in natural language processing, it also poses several risks.
One of the main risks is that the data used to train these models may contain biases based on the demographics, cultures, and perspectives of the individuals who generated the data. This approach can result in models that perpetuate and even amplify existing societal biases. For example, a model trained on data containing sexist language may generate text that also contains sexist language. Additionally, these models may not be able to handle diverse languages, dialects, and cultures effectively.
Another risk is that the models may not be able to distinguish between different types of text, such as news articles, fiction, and satire. This can lead to models generating text that is factually inaccurate or misleading. Therefore AI solutions must be paired with structured, scientific data in order to draw conclusions.
These limitations are particularly important in life sciences, where the stakes are high, and the consequences of errors or biases can be severe. It is crucial to have robust ethical and governance frameworks in place to ensure that these models are used responsibly and are held accountable for their outputs. Furthermore, it is important to continuously monitor and evaluate the models for biases and errors and take steps to mitigate them.
One option to make LLMs more fact-based around a certain task is to “fine-tune” these models by enhancing them with high-quality, structured data on a specific task. By fine-tuning an LLM we can greatly improve its ability to respond adequately within a certain domain. The big question that then remains is: What if we end up stumbling upon a question that the training set didn’t cover? How likely is it that we receive an answer that is once again “made up”? These subtleties will have to be understood completely before we take LLM Clinical Trial implementations out of beta.
Ethical implications of ChatGPT in clinical research
The main ethical concerns in the use of LLMs like ChatGPT are transparency, credibility, and accountability. These concerns are amplified if research participants are not made aware that they are interacting with AI.
First, there is the issue of informed consent. If patients or participants are unaware that they are interacting with AI, they may not fully understand the nature of the interaction and the potential risks and benefits. This could lead them to unknowingly provide sensitive information or participate in research that they would not have agreed to if they had known they were interacting with an AI. This undermines the very purpose of informed consent.
Second, there is the matter of trust and credibility. If patients or participants are not initially aware that they are interacting with AI and find out later, this could lead to mistrust. This may cause a lack of compliance or participant dropouts.
Third, there is the issue of accountability. Suppose patients or participants are not aware that they are interacting with AI. In that case, it may be difficult to determine who is responsible for any errors or inaccuracies that may occur. (Of note, there has already been much discussion about this in the world of radiology, as seen here and here.) This could lead to confusion and mistrust, making it more difficult to correct any errors that may occur.
While it’s important to recognize the limitations of Large Language Models (LLMs) like ChatGPT, there is no denying that its ability to act in a human-like way presents exciting opportunities within clinical research. Already, ChatGPT has been named as a co-author on at least four research papers. With proper management, AI generation could be an invaluable tool in clinical research—but only if we are mindful of the potential risks and biases that may arise from their usage and ensure transparency, fairness, and accountability in their use.
Photo: Vladyslav Bobuskyi, Getty Images
Derk Arts MD, PhD has over fifteen years of experience in medicine, research and technology. He founded Castor to solve the biggest issues in clinical research: a lack of inclusivity, patient focus and impact of data. Castor enables sponsors worldwide to run patient-centric trials on a unified platform, that helps them maximize the impact of research data on patient lives.
Dr. Derk Arts believes the key to achieving lasting change in the industry is through scalability and standardization. Technology to run better trials and maximize the impact of data should be available to all researchers.
Over the past 12 months, Castor provided pro-bono support to over 300 Covid-19 trials, and provided the entire infrastructure for the World Health Organizations’ Solidarity Trials.
