AI has been appearing as a hot topic for the past couple years – nearly omnipresent in newly launched apps and being integrated into currently existing tools. Many of us, however, have already used AI, Machine Learning (ML), and Computer Vision indirectly via Google, face recognition, social media, and a host of other now-ubiquitous technologies.  

We asked our employee experts to share how they use AI at work in medical device development.  They explored where it shines, where it’s taking over, and where it falters (and they take over). Most employees were fans, but a few remain unconvinced.  

Read on for 9 ways to use AI in medical device development and a couple of reasons to think twice before engaging with proprietary information. 

Connected Medical Devices 

 AI data analysis of data generated by medical devices can do computations many orders of magnitude better than a human can. It Is a great ally in helping look for trends and other things in the data. Then I take over for further analysis and interpretation. 

QA/RA  

We’re incorporating AI into one or more medical devices and lean heavily on our regulatory team to support our efforts because AI is to a degree, kind of a black box. There are some misgivings about giving it too much power or authority for diagnosing or recommending. Often it seems that the FDA and regulatory bodies want AI to be sort of like an AI doctor team. It is regularly being incorporated into medical devices.  

In QA/RA, AI can automate documentation processes, improve the accuracy and speed of data analysis, and reduce risk by predicting potential compliance issues, which in turn can lead to more efficient quality assurance and regulatory adherence. But AI should be used for assistance only. I always check and refine or fine tune work, otherwise it can lead to over-reliance on automated systems.  

By using AI to perform data analysis for trends and inconsistencies or outliers, potentially AI could gather data during regulatory submittal research. For example, finding similar devices in the marketplace, and assessing their worthiness as predicates for a desired device submittal. 

We are developing an AI tool to help with medical device regulatory searches. AI is great at summarizing and compiling, provided the information it pulls from is relevant. We provide the tool with information chosen manually for relevance. We’ve restricted some of the garbage in, garbage out. We are still working on refining input data so it will get better over time.  

Developing (real time) Simulations 

We need Python for hardcore programming in computational modeling software. I use ChatGPT instead of going through programming books. AI provides a lot of information in a very short amount of time.  

ML is revolutionizing Simulation. To develop simulations, we prepare meshes and complex models. We run them using a lot of computational power. If we train a physics-informed neural network on the situation instead, we can get results in real time. This is the next front for the field of simulation. There’s a lot of work going on in the simulation side. The basic purpose of that work is to make simulation more viable for clinical use. You need to run the simulation in the clinic to get the results ASAP instead of waiting for days or sometimes longer. That’s where simulation and AI are headed.  

I plan to use AI more and more because it is a requirement of the field itself to build a capability for patient specific modeling. With AI we can prepare a device and test it in different conditions. Using AI for patient specific testing will make a big impact on testing and reliability of the device. 

Expedite Design Process 

AI is a tool to expedite my design process. Vizcom.ai released a new generative AI feature to turn a simple sketch into a realistic image (which I use for rapid visualization of industrial design concepts), or 3D models. Although AI generated models are not as accurate as traditional methods, I can 3D print the AI generated models to get a sense of the overall form. This would take a few hours to do manually in SolidWorks due to the complexity of the surfacing involved. I’m hopeful AI software will improve its model generation accuracy so rapid prototyping complex geometries can truly be ‘rapid’ and more industrial designers integrate this AI tool into their workflow. 

Project Reviews

AI can more accurately identify potential roadblocks, gaps and opportunities for improvement. With AI, I streamline my project activities including planning and identifying potential risks. 

Using AI in project reviews makes it easier to identify themes from team feedback. AI tools help quickly sift through input and pinpoint critical pain points and areas for improvement. This facilitates concise summaries faster, enabling more efficient knowledge sharing and better-informed project decisions. 

Supplier Evaluation 

AI for supplier evaluation and selection is super cool and useful to understand a supplier’s capabilities with feedback from the market. I use AI to generate Ideas for contracts with suppliers and quality agreements. I would love a plug-in for Syspro so our engineers could punch in their requirements, “I have a phase one project and need sheet metal. My tradeoffs are quality and cost”, and then Syspro AI provides the history of suitable suppliers in our database.  

AI also helps search for suppliers and vendors by validating and performing a gap analysis to plug holes in manual bottom-up searches. We recently conducted a vendor search for pressure sensor suppliers without using AI. From the bottom up, we looked at the whole industry, all the different suppliers and broke down the different products that would work for us because we felt the search didn’t suit itself to something like ChatGPT large language model. At the end of the search, we made a few AI queries to see if there were any vendors missing, and we got 1 or 2 that we hadn’t found through a traditional search engine.  

Analysis and Research 

When conducting research facilitating the engineering behind medical science, AI often finds different publications, the latest data, etc, for us to digest. Some AI tools are good at scanning through the literature, summarizing it and identifying key points. There is a bit of risk because you need to understand the subject. AI sometimes makes mistakes. It’s getting better and better as it goes but it still it takes some bandwidth to review AI results. AI usually gives back more bandwidth for us to doing actual engineering as opposed to scanning what may or may not be good information coming from the literature.  

AI is a very important add for a literature review. Recently we received a couple of projects and had to do a quick literature survey to get an idea of what we were getting ourselves into. After that level of information, I don’t trust AI fully. I only trust it to provide me with guidelines, the broader points that I can search for.  

AI helps to speed things up if you get the queries right when looking for papers you might have missed in a landscape search for R&D, or even within the landscape search, or where you’re trying to break down complex research with huge reams of information and you need certain points. You can pull AI derived information out, and then break it down for yourself. Ultimately, remember that you are responsible for accuracy.  

I used Perplexity AI while writing a speech about conflict. I wanted to find all the research I could find about conflict. It gave me a synopsis and included links to all the papers and websites it referenced. That was helpful and added a couple of authoritative tidbits to my talk.  

Brainstorming 

AI can be very useful in the brainstorming stage. It is nice to input a vague idea of what you’re thinking, see what AI returns and determine if you can continue brainstorming off of the AI content. AI is a good sounding board to brainstorm ideas back and forth. 

I ask ChatGPT questions to stimulate ideas. I wouldn’t rely on its results to give me any concrete numerical answers or anything like that. But, for creative ideas, or a solution to a problem, it’s worth asking ChatGPT.  

Administrative Help (Faster than Google) 

Use AI to take meeting notes. I summarize the AI notes and then regurgitate them back. It’s super beneficial for project management. Before AI, I would sit and take notes using my grade ten keyboarding skills or have a team member dedicated to taking the meeting notes.  Having AI capture and regurgitate the notes and then review and send them to participants is super helpful.  

When I am struggling with the correct formula in Excel, I will tell AI the data I have in each cell, and ask for a formula that will provide the information I’m seeking in another cell. It is often faster than googling the same thing.  

I leverage ChatGPT to perform menial tasks like generating citations in American Psychological Association (APA) format, very preliminary data extraction and compilation from multiple sources, and for alternative or differentiated explanation of an unfamiliar concept.  

AI helps summarize large files by providing key points and takeaways. Sometimes it is better to start with a summary draft generated by AI than to start from scratch. It helps save time and enhance overall productivity. 

Edit or Enhance Writing and Communications 

AI helps refine or change the tone of content and messages I write. For example, I tell the system to use the voice of an expert who is explaining a complicated idea to an audience with minimal subject knowledge. It provides adjusted text that I then augment. It is much easier to edit text than start from a blank page.  

AI helps review documents before sending. It’s almost like a thesaurus but for full sentences rather than just words. 

Personal Budgeting 

AI helps improve my personal budgeting practice and saving opportunities. Running an itemized receipt through AI helps me categorize each line item into broader categories and set a budget that can be controlled. I more accurately predict how much and where I will spend in a month. 

Developing AI as Tools 

In the field of medical ML development, we are keenly aware of the careful balance we need to reach between the potential benefit of new technologies, and hidden risk to the quality of our medical devices and the health of patients. Language models open many new opportunities for automating low-risk but laborious tasks in text processing and classification pipelines. However, they are difficult and expensive to set up at scale for security-conscious organizations that can’t risk using third party tools.

Small, custom language models could see use as lightweight sentiment analysis or text classification in specialized contexts or environments. Finally, the transformer architecture has been shown to be excellent for general time series analysis tasks, and could be particularly useful in medical signal analysis.