Artificial Intelligence in Healthcare

Note: Robin Farmanfarmaian recently interviewed Michael Ferro about his ideas on artificial intelligence. Michael has been granted multiple patents and has founded or rebuilt many companies in AI including Merge Healthcare, which he sold to IBM for $1B in 2015.

Michael believes AI is a main tool that can be used to democratize healthcare. This series covers AI innovations we talked about that touch patient care directly, including remote patient monitoring, imaging analysis, digital therapeutics and more. Check back here weekly for the next installment — Robin Farmanfarmaian

Artificial Intelligence in Healthcare Series: Introduction & Series Overview

The healthcare industry is seeing an explosion in artificial intelligence enabled software. According to Markets and Markets, the global AI in healthcare market was $4.9B 2020, and is expected to grow to $45.2B by 2025. There are many uses for AI in healthcare — this article focuses on artificial intelligence that touches patient care directly through diagnostics, disease management or treatments. This usually means software that requires FDA oversight and/or HIPAA compliance.

What’s HIPAA and the FDA got to do with this?

Not every AI software program that can make a positive impact in healthcare needs to go through the FDA or be HIPAA compliant. AI software is superior at many different things, including pattern analysis, crunching vast amounts of data, and analyzing multiple data streams. AI software programs that work in supply chain management, demand forecasting, drug discovery and distribution modeling don’t require FDA clearance because they aren’t involved with “Practicing Medicine”.

“Practicing Medicine” is what a physician, clinician, or other licensed healthcare professional does when they treat patients. A general rule around if an AI software program requires FDA clearance is if the software enhances, improves, or replaces something that a licensed healthcare professional — such as a radiologist or neurosurgeon — would typically do or prescribe in their job as a healthcare practitioner treating patients. In terms of HIPAA compliance, HIPAA is a set of rules around privacy and security when dealing with patient data and communication in a healthcare setting, whether in-person or virtual. The general rule is if the software is dealing with patient data or communication, then it needs to be HIPAA compliant. In the world of AI software, if the software needs FDA clearance, then it usually also needs to be HIPAA compliant. The opposite is not necessarily true, many software programs need to be HIPAA compliant but don’t necessarily also need to go through FDA clearance, such as a messaging app between a doctor and a patient.

Data and Pattern Analysis

Though the term “artificial intelligence” was coined in the 1950’s, AI innovation, progress and adoption have sped up dramatically in recent years. On the TED main stage in 2018, Chris Anderson interviewed world-renowned Artificial Intelligence expert Dr. Ray Kurzweil. While talking about his groundbreaking work on AI at Google, he mentioned a key concept that explains the dramatic leaps we’re seeing with AI. He said that when working with artificial intelligence, “Life Begins at One Billion Data Points”. Anything less than that isn’t enough data to even begin to use AI. Especially when it comes to healthcare, it takes time to build up and aggregate the clinical grade databases needed to train AI in healthcare.

Back in 2008 when Merrick Ventures invested heavily into Merge Healthcare, artificial intelligence and machine learning (AI/ML) were just beginning to become focal points in the world of healthcare. IBM Watson was born only a few years earlier, and for many years they dominated the news in healthcare. The reason? IBM had massive databases of healthcare data in addition to medical textbooks to train Watson.

Strong Data = Value

The key to artificial intelligence is in using datasets that are giant, high quality and reliable. Too many AI software programs use small or incomplete datasets, which results in training the AI software incorrectly. Garbage in, garbage out.

Artificial Intelligence and the FDA

The FDA just released their latest guidance in January 2021 around AI software. Healthcare can expect to see the number of AI based software programs that clear the FDA to dramatically increase over the next 10 years. Bakul Patel, the founding director of the Digital Health Center of Excellence, is leading up these efforts. As a visionary in digital health, Bakul has made it a priority to ensure these innovations are safe, accurate, and improve patient outcomes — while getting them into the healthcare system to help patients as quickly as possible.

One thing about AI/ML (ML is machine learning, an application of AI) is that the software actually does learn — it adapts itself based on new data. Just like a human being, machine learning software changes and improves the more it is used. Previously, once a product had cleared the FDA, the software would need to be fixed and can’t be altered without going through the FDA again. Obviously this is in direct opposition to one of the main benefits of AI. So the FDA has focused on making sure AI software, known as SaMD (Software as a Medical Device), has enough freedom to improve after it has been cleared, while still keeping patients safe and secure. Check out the FDA’s latest guidance in this article or in the FDA’s press release.

The US’ first FDA cleared AI based software program was back in 2012. While the next few years only saw a handful of other AI based software clear the FDA, that has accelerated dramatically in the past 3 years. We’ve now seen over 70 AI software programs clear the FDA, and this is just the beginning.

Most of these FDA cleared programs are centered around specialties like radiology, cardiology, ophthalmology, and neurology — all specialties that require imaging analysis. In general, AI is the best at pattern analysis, which makes it a great fit for anything centered around image analysis. An AI software program can spot abnormalities human eyes aren’t always able to see, no matter how great the physician or specialist.

Some AI based software programs help shift where the patient receives their care. Patients with diabetes should have their eyes checked once a year for diabetic retinopathy, a leading cause of blindness in diabetes. Typically that means having to make an appointment with an ophthalmologist. Not only is this another appointment to think about and schedule, but most of us don’t do our annual exams perfectly every year. FDA cleared IDx-DR software from Digital Diagnostics helps shift that one exam from the specialist’s clinic to primary care. The doctor takes an image of a patient’s eye with a fundus (retinal) camera, and IDx-DR analyzes that image in 60 seconds, with a yes or no binary answer: yes, the patient may have diabetic retinopathy, or no, the patient is good to go, come back next year to get checked again. This is a gamechanger for areas without easy access to ophthalmologists, as there are now fundus camera smartphone attachments that are considerably less expensive than a full size fundus camera in a clinic. That means it is easier to outfit primary care clinics, actually get patients tested, and hopefully prevent more diabetes patients from losing their sight.

While IDx-DR is the first software cleared by the FDA that doesn’t require a healthcare professional to interpret the results, the software is used to only do this one very narrow task, for this one test. Even if there is something else noticeably wrong with the patient’s eyes, IDx-DR won’t pick it up — the software is only trained to find retinopathy. Currently only Artificial Narrow Intelligence (ANI) actually exists, where the software program is superior when doing one narrow task, like checking for retinopathy in an image. Ophthalmologists look for diabetic retinopathy, in addition to any other abnormality or problem with the eyes. Because of this, AI won’t replace a physician completely anytime soon, but what will happen in the near future is that physicians who use AI will replace physicians who don’t.

BrainCheck is another interesting company that is able to shift life-altering exams out of expensive, scarce specialist clinics and into the easier to access, more prevalent primary care clinics. Around 30% of dementia cases are preventable, yet most of us don’t have a baseline on our brain health the same way we do blood pressure or cholesterol, so it can be difficult to notice subtle differences day-to-day. In addition, many patients don’t go see a neurologist until their memory loss starts affecting their daily life. That typically means they are so far advanced into a form of dementia that lifestyle behavior modification and even pharmaceutical intervention won’t help.

With dementia, prevention or early diagnosis is the main goal, as those early stages are treatable. To solve this problem and help prevent up to a third of dementia cases, BrainCheck has taken the standard neurological exams found in a neurologist’s clinic, and turned the tests into a gamified smart tablet app that takes patients 5–10 minutes to complete. Suddenly, those standard neurological exams don’t require a healthcare professional to administer. The primary care office just needs to hand the patient a smart tablet, or send the app for the patient to use at home on their own device. BrainCheck is FDA cleared and falls under two reimbursement codes at CMS (Medicare) for the cognitive test and cognitive care platform, so doctors can get paid when patients use the software under their clinical guidance.

Check back next week for the second installment where we cover Digital Therapeutics

Healthcare Tech Entrepreneur Founder & CEO of Merrick Ventures

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store