Ten Questions to Ask AI Companies in Medical Imaging (and Healthcare At Large!)

Falgun H. Chokshi,
Founder, FalgunChokshi

The following is part of a Medium.com article I wrote before about vetting AI start-up companies. If you like it, please "clap" for article on Medium and share on LinkedIn and Twitter. And check out more at falgunchokshi.com...

So, what do you do to tell the AI snake oil salesforce from seemingly visionary companies that may legitimately help radiologists improve their contribution to patient health? Here’s a list of questions I ask vendors at meetings and conferences.

This list is by no means comprehensive, so use it as a launching pad to think for yourself:

“What Does Your Software Really Do?; explain to me the algorithm(s) being used and whether they are to look at a specific imaging finding for many diseases, or a specific disease process that can have many imaging findings?”

This is an important start to the discussion. Beware of representatives (reps) marketing their wares as some revolutionary algorithm(s) that can cure cancer. The best uses of AI are for targeted, medically relevant clinical questions, not for platitude-ridden aspirations. Most of the companies will use supervised machine learning models that rely on a “ground-truth” data set of imaging studies to train their algorithms. Ask them to describe the training set, how they validated (at different clinical sites), and what was the balance of positive (disease present) to negative (“normal”) studies. If their training data-sets are skewed towards too many positive or negative samples, their algorithms’ accuracy will sway significantly. Plus, if they ask you to “collaborate” with them, ask them if they want clinical and imaging data. If that’s the gist of it, walk away. Once you give the data, you won’t hear from them.

“Tell me about your founders and leadership team and your advisers, especially the medical subject matter experts you worked with to develop, train, validate, and test your software”

There are many companies out there that have no clue about the medical side of imaging. Rather, they just see everything as images and want to find open source data-sets to develop and train machine learning models. Are they actually working with practicing radiologists? What are their areas of expertise? If the reps have blank stares as you question them about this, politely pretend you receive a phone call and walk away. They don’t have a clue about what they are doing.

“What kind of Machine Learning methods are you using in your software?”

If you hear anything other than “neural networks”, “deep learning”, “recurrent neural networks”, “convolutional neural networks”, then they are using outdated technology. If they mention terms like “support vector machines”, “decision trees”, or other terms not including “neural” or “deep”, then they are merely using high end statistical modeling techniques that do not actually “learn” anything. There’s no “intelligence” in artificial intelligence when using these latter techniques.

As a followup to the last question, ask “tell me about the hardware requirements to use your software? Is it a cloud-based computing model with a subscription requirement? Or does my group need to invest in servers (including GPUs) and trained IT professionals to maintain them? Tell me about your company’s support of the software, in detail.”

Machine learning applications, especially the state-of-the-art neural networks, require a lot of processing power… A LOT. So, make sure you know exactly what you are buying from them versus the infrastructure needs for which you will have to budget.

“Does your software have an accountability or auditing feature that allows me to see the rationalization of its decisions?”

This is important, as most of these programs are “black-boxes” and no one, except MAYBE the company’s data scientists, actually know what the machine is really doing. Asking about this will help separate out vendors with whom you should keep talking. Examples of such features would be heat maps that show areas of an image where the machine “saw” or highlighting of specific text in a report or medical record using natural language processing (NLP).

“Where in the Radiology Information System (RIS)-Picture Archiving and Communication System (PACS)-Electronic Health Record (EHR) system does your software operate? Is it a fully automated end-to-end pipeline system requiring minimal radiologist input? Is their an easy to use interface that allows the radiologist to give feedback to the algorithm about its misses, thereby improving its accuracy?”

These questions will help you understand the level of integration their software has with your existing IT infrastructure. The best software, whether, AI or not, works in the background, without the user really noticing its presence.

“Tell me where you are in the FDA approval process and who are your investors, including how much capital you have raised and if you have passed Series A funding?”

This will tell you two things: Are these folks serious about their company and how likely are they to succeed given they need capital to grow the company and make it successful. Plus, you will find out the level of FDA clearance they aim to obtain. Check out this page for more on that: https://www.fda.gov/AboutFDA/Transparency/Basics/ucm194438.htm

“What are your sales to date and who are your current customers?”

This will give you an idea of their current success and if they will fizzle soon. You really don’t want to buy their AI software and then they are out of business the following year.

“What other products/algorithms do you have in the pipeline? How many different AI features are in your software package?”

You don’t want to buy software that only finds lung nodules, rather finds multiple abnormalities with a high accuracy on chest x-rays, but better yet, on both x-rays and chest CT scans. Ideally, you also want those abnormal findings to be integrated with EHR data about the patient’s medical encounter, including vitals, lab data, and physician assessment (one can only dream!). Remember, you want to maximize the value as the buyer. So the vendor should really have a comprehensive package of software to make it worth your while.

“Why should I buy your AI software compared to your competitors’, who claim to do what you do, but better? What is the value proposition?”

See how good they are at selling you their product and vision. There are a lot of green data scientists looking to make it big and have their companies bought out by the Googles and Amazons of the world. But, you are the potential customer. They have to maximize the value to you as the buyer, otherwise they will have only short term success in sales, but no sustainability, and hence, no buy-out by the big tech giants.

I hope you get some useful and interesting answers to these questions and many more. Despite the pessimism of these questions, I remain a cautious optimist about the power and value of AI to revolutionize healthcare! At the end, AI is an amazingly powerful technology, but it is amoral and increasingly available for use and manipulation by the masses. But, like all revolutionary technology, its REAL adoption and dissemination will come after a big hype cycle… and we are definitely in one now!