Beyond ChatGPT: What Does GPT-4 Add To Healthcare?

Dust has yet to settle on the ChatGPT hype, but OpenAI, the developer behind the artificial intelligence (AI) language model underpinning ChatGPT, has launched the next iteration of its model: GPT-4.

OpenAI advertises its new AI model as being “more creative and collaborative than ever before” and “can solve difficult problems with greater accuracy”. The tool is already being adopted by organisations such as Be My Eyes to assist visually impaired people and Khan Academy in education; and we will surely encounter more medical use in the months to come. So let’s get acquainted with GPT-4 and what the technology means for healthcare.

What is different with GPT-4?

Like with the new iterations of practically every piece of technology, GPT-4 improves over its predecessors. More specifically, it performs 40% better than GPT-3.5 at producing factual responses based on OpenAI’s internal evaluations. For context, GPT-3.5 is the model on which the popular ChatGPT software currently runs.

Compared to the latter, OpenAI describes that GPT-4 is more capable by being able to handle much more nuanced instructions and deliver more reliable and creative output. The new software also performs better than the former version in benchmarks and simulated exams originally designed for humans such as the Medical Knowledge Self-Assessment Program (75% vs 53%).

While GPT-3 and GPT-3.5 accepted only text-based input, GPT-4 is multimodal as it accepts both text and image input to provide contextual text output. To illustrate this new ability, OpenAI provided the following example:

How to use GPT-4?

Based on what has been made available GPT-4 displays impressive results, but its access to the public is relatively limited; and even those with access have limited usage. Currently, it is available to ChatGPT Plus members, the $20 monthly ChatGPT subscription offered by OpenAI.

Even with a ChatGPT Plus subscription, users can ask only 100 questions within a 4-hour window and image prompts have not been made available. However, it could be rolled out to a wider audience in the coming months as there is a waiting list and other tools from other companies, such as Microsoft’s Bing chatbot, are leveraging the technology.

Illustrating GPT-4’s improvements for medical professionals

Even if accessibility to the tool is currently limited, GPT-4’s improvements also extend to medical queries; and this makes the software an interesting development in the digital health space. The differences between the responses of the GPT-3.5-powered ChatGPT and the new GPT-4 are stark. GPT-4’s output is more detailed, more nuanced and, admittedly, more creative than ChatGPT.

Who else can better illustrate this than GPT-4 itself? So we asked GPT-4 to provide prompts that would demonstrate the differences between Chat-GPT and GPT-4 for medical professionals. The following was the output provided:

Prompt

Diagnosing a patient with ambiguous symptoms

ChatGPT output

A patient presents with fatigue, weight loss, and occasional dizziness. What are some possible causes for these symptoms?

GPT-4 output

A 45-year-old male patient presents with a 3-month history of progressive fatigue, unintentional weight loss of 15 pounds, and episodes of dizziness. Please provide a differential diagnosis and suggest relevant diagnostic tests.

Prompt

Treatment recommendations

chatGPT output

What are some common treatments for type 2 diabetes?

GPT-4 output

A 55-year-old female with a recent diagnosis of type 2 diabetes has an HbA1c level of 8.5%. Outline a comprehensive treatment plan, including lifestyle modifications, pharmacological options, and follow-up monitoring.

Prompt

Patient education

chatGPT output

Explain high blood pressure in simple terms.

GPT-4 output

Create a patient-friendly educational handout on hypertension, including an overview of the condition, risk factors, symptoms, potential complications, and management strategies.

Prompt

Reviewing medical research

chatGPT output

Tell me about the benefits of exercise for mental health.

GPT-4 output

Summarize recent research findings on the relationship between physical activity and mental health outcomes, including potential mechanisms, types of exercise, and recommendations for various populations.

Prompt

Clinical case scenario

chatGPT output

Describe a patient with pneumonia.

GPT-4 output

Create a detailed clinical case scenario involving a 65-year-old patient presenting with community-acquired pneumonia, including history of present illness, relevant past medical history, physical examination findings, diagnostic test results, and treatment plan.

The potential medical purposes for GPT-4

Aside from such prompts, where could GPT-4 be integrated into medical practice? Its potential use is multifold, ranging from administrative tasks to assisting in medical consultations.

Given its ability to accept more than one input type, GPT-4 can provide context to images by analysing and classifying them. This feature could be useful in transcribing hand-written prescriptions or even assist in the transition of paper-based medical records to electronic ones.

As the new iteration is more reliable at providing factual responses, it can be used by patients to provide a general overview of medical conditions. It might not be as adept at the specifics but can serve as a primer for patient education.

In the doctor’s office, the tool could summarise a patient’s medical history as well as provide an overview of available treatment pathways. In another case, the AI could assist in summarising clinical notes for integration into electronic medical records. In fact, this is what the Microsoft-owned company Nuance announced its new GPT-4-powered tool will enable. This can save time for physicians, which they can dedicate to their patients and offer the human touch during consultations.

GPT-4 could also enable physicians to stay up-to-date with medical research by summarising lengthy academic papers and highlighting the key takeaways.

The potential uses of GPT-4 in healthcare are indeed manifold, but the actual implementation will depend on the availability of those tools and their ease to use by patients and physicians alike. However, we can expect to come across more digital health companies leveraging the technology soon that will materialise its potential. It will be a rapidly-evolving space, but one that will be interesting to observe; and we’ll be sure to keep you up-to-date with those exciting developments.