“AI is a robot perched on our shoulder, not the creator at the keyboard.” –Ann Handley
Just five days after OpenAI released its chatbot ChatGPT on Nov. 30, 2022, it had signed up 1 million users. A month later it had more than 100 million, making it the fastest-growing consumer application in history.
The problem with ChatGPT is that while it can quickly generate words based on a given input, it doesn’t truly comprehend the meaning behind those words. As a result, the responses it generates typically lack depth or insight – what you might encounter in a high school term paper or a short Wikipedia entry rather than the ruminations of a thought leader. But that will change as artificial intelligence (AI) evolves.
Clever Parlor Trick or Valuable Business Tool?
The hype around ChatGPT has been exceptional. Multitudes of businesses from manufacturing to pharmaceuticals have accelerated plans to incorporate AI and machine learning into their business processes.
In the medical sphere, for instance, ChatGPT has enough “knowledge” to help medical students learn new concepts and study for exams, and help administrators streamline processes and manage massive quantities of data. That’s because ChatGPT “learned” from more than 570 gigabytes of textual data gleaned from books, articles, Wikipedia, and other internet content, including science and medicine-related texts.
Pharmatech companies have used ChatGPT and other medical-specific AI models to accelerate the long and challenging process of drug discovery. For instance, scientists at AstraZeneca are using it to search for new drug targets and then developing safe and effective medicines. By having AI crawl through massive amounts of biomedical data, AstraZeneca’s research teams have more resources and opportunity to identify drug targets– even find connections that human investigators may have missed. Using AI to shrink the time (and costs) associated with finding new drug candidates can potentially propel promising candidates to clinical trials more quickly than today’s norm.
In clinical practice, ChatGPT has the potential to refine the documentation process by quickly generating medical charts, progress notes and discharge instructions. And it can help health care workers examine a patient’s symptoms and recommend a course of action.
Yes, there’s concern that its responses may contain inaccuracies and bias, and even provide answers based on sketchy or discredited sources, so the message in the health care realm, where faulty information can ultimately lead to a patient’s death, is to tread carefully when it comes to AI.
ChatGPT in Medical Device Development
As I work for a software company that specializes in creating groundbreaking medical devices and applications, I wanted to explore the relevance of ChatGPT (and competing bots) in medical device development. Is there a role for ChatGPT?
I asked ICS’ Director of Medical Programs Milton Yarberry for his take. Here’s what he had to say:
ChatGPT is a tool that can potentially make us more productive at our jobs but it won’t replace humans. I don’t think it has any real immediate or even a medium-term future in medical device products. ChatGPT isn’t a knowledge engine. The underlying implementation doesn’t have a knowledge structure that it references or that can be verified. The ‘knowledge’ it has is merely a result of analyzing sentence structure. It is a chat engine designed to converse with people, so that’s its strong suit. Sure, there are a lot of other fringe benefits like understanding how to code, but that’s the result of the training set that was chosen.
Though I don’t see ChatGPT as essential to medical device development, AI (and machine learning) in general will grow in importance. The U.S. Food and Drug Administration (FDA) is already prepared. A few years ago, the FDA put out guidance for trainable machine learning, laying out expectations for what a company must do to update their machine learning without going through a full-scale submission process.
That guidance is expensive and requires systemic changes to an enterprise. The gist is that companies have to define ahead of time the areas in which they expect improvement and meticulously control/curate the data they use to train the system. What controls are there for machine learning in medical devices? The core of an organization’s medical device design includes extensive risk analysis, dFMEA and hazard analysis, and must include corresponding mitigations, traces to implementation, verification and validation, and software testing that ensure that safety and efficacy are maintained.
In this regard machine learning is no different than any other form of software. The FDA’s guidance makes the use of machine learning a risk-versus-reward proposition and holds it to the same standards that all medical devices were meeting before this recent boom in AI.
Chat GPT really doesn’t change anything on that front – it doesn’t help or hinder. It’s a language bot with generalized knowledge combined with some specialized knowledge. It will be helpful in chat-like scenarios or applications, but those weren’t really heavily regulated by the FDA to begin with.
ChatGPT has brought renewed attention to the discussion around AI and its potential to improve healthcare-related practices, processes and workflows – how medical students are trained, how clinicians treat patients, how administrative tasks are handled, and how new therapeutics are developed. Clinicians are already using it to forecast leukemia remission rates, help predict kidney disease, and accelerate drug discovery.
That’s why our medtech team is actively – but cautiously – exploring ways to safely (and securely as there are significant cybersecurity implications) leverage AI and machine learning where appropriate in medical device software development to benefit our customers and their users.
As for the takeaway, ICS’ CEO Peter Winston sums things up nicely: “The arrival of ChatGPT is a turning point. It is AI that we all can utilize as it is both accessible and useful to so many things. Whether that includes medical device development specifically remains to be seen. The next five years will be a race among manufacturers to add intelligence to everything. The whole landscape may change.”