How COVID-19 Is Accelerating the Adoption of AI in Healthcare
COVID-19 is exposing the need for adaptable, dynamic systems in healthcare. Fernando Schwartz, Head of AI at CitiusTech, shares trends in artificial intelligence and machine learning with the Health Transformer community.
The COVID-19 pandemic has brought about profound changes within both the healthcare system and the world at large. The ongoing and dynamic nature of this public health crisis is exposing the importance of having systems and solutions that are adaptable rather than fixed. Within the telemedicine and digital health industry, artificial intelligence and machine learning are increasingly utilized and integrated to analyze data and generate actionable insights and improve healthcare. In this Expert Office Hour with Fernando Schwartz, we explore how AI and ML are changing telemedicine and digital health solutions during COVID-19 and beyond. Schwartz is the VP of Data Science and Head of AI at CitiusTech, an IT service company that develops solutions for top healthcare companies. Here are a few key pieces of wisdom that we took away from our time together.
Overview of Artificial Intelligence
What is AI?
The field of AI and ML is extremely broad so there isn’t a sole definition of artificial intelligence. Generally speaking, when we are trying to make a decision based on a lot of data — for example, if there are 100,000 contributing factors that affect an outcome — it can be a daunting task for humans. As the number of factors grows, it becomes harder to determine the relationship between the inputs and the outputs. This is where machine learning and AI comes in. Though machines are not wholly smart, they are advantageous in making a rational decision based on the data and helping us scale the way we handle data.
What AI is not
Artificial intelligence is a shiny object that is often thrown around as a marketing term. It is important to remember that AI is not an end in itself but rather it is a process that is integrated into a workflow. Many companies are artificially going into the AI space just to make a claim. In order to avoid the pitfalls of AI, Schwartz recommends knowing the difference between data analysis and AI and deeply understanding the value chain and how AI fits into what your company is doing.
What are neural networks and how are they utilized?
A trained neural network is a technology that comes with pre-assembled knowledge and this is important because the process of training an algorithm is very expensive. Neural networks are utilized in digital health in areas such as imaging and image recognition. For example, if a technician is conducting an ultrasound exam of a patient’s heart and they freeze on a frame that they like to do a more detailed search, we can run measurements around that frame. Similar to how we have machines that can recognize people and cars on the street, you may want the machine to look at the ultrasound and determine the location and size of the right and left ventricle. In order to train a machine to segment and understand within context what an object is, you would need to show the machine a vast number of images. However, through a process called transfer learning, neural networks are able to learn from previous software and take knowledge from the old context into the new context.
How COVID-19 Has Changed The Perspective on AI
“Relearn the rules instead of rewriting the rules.”
With abject changes like COVID-19, companies are realizing that rules-based systems are not able to handle the rapid and dynamic changes in the world today. As a result, they are pivoting to solutions that are more adaptable and can be retrained. AI and ML allow systems to relearn the rules instead of rewrite the rules. The pandemic adds to the story about why AI is important within healthcare: when building AI and ML pipelines, some parts of the process revolve around monitoring. And with viruses, this monitoring is critical for faster public health response times. For example, CitiusTech is working with renal care clinics to analyze nation-wide transmission levels and demographic data to determine the rate of transmission for specific areas. Other companies are looking into the relationship between social determinants of health and hospitalizations. There are also claims about how machine learning can be used to diagnose COVID-19 from chest x-rays.
AI in the Future of Telemedicine
A Push Forward
When thinking about the interactions between physicians and patients, this space has traditionally been off-limits. But as telemedicine opens up more opportunities for physician-patient digital interactions, machine learning will continue to play a significant role in adding value to the robust digital pipelines that are being built. For example, machine learning can be integrated into the process of analyzing appointment videos and taking value and data from that.
Barriers to Implementation
While COVID-19 may accelerate the adoption of AI in healthcare, it is not easy to drive change in the health industry.
The first obstacle is the FDA. For example, I’m working with a pharma company that is working to help patients with epilepsy go into a second line of therapy. In order to go from the first to second line, and receive approval from the insurance company, the patient must meet the threshold of having three seizures. In order to help with the insurance approval process, we are trying to use AI and predictive modeling to analyze whether a patient is at risk for having three seizures. This product is still awaiting FDA approval, as they are still trying to understand how AI fits into the bigger picture of compliance. But the general trend is towards human factors and explainability.
The second obstacle is liability. Unlike the big tech industry, liability and AI often have an inverse relationship within healthcare. When talking to individuals that work in big technology companies like Amazon about a new algorithm, the response is always a positive “let’s do it!” But the mentality is different in healthcare. On one hand, companies that have a more direct impact on patient lives are more hesitant to integrate AI and ML into their workflows. For example, medical technology companies are hesitant to include AI and ML because they are concerned about intellectual property. On the other hand, companies in healthcare that are more removed from direct patient interactions — such as companies that work in the scheduling or maintenance space — are more receptive to the idea. Given the hesitation surrounding the use of AI in healthcare, Schwartz recommends doubling down on explaining the abilities of the model. There is a spectrum of algorithms: on one end we have a black box algorithm with high precision, and on the other end we have a low precision but simpler algorithm. It is possible for payers to choose the simple, less accurate model over the fancy, complex model because being able to explain how the data and recommendations came about from the algorithm is crucial when auditing decisions.
The third obstacle is bias. Bias in AI and ML is a huge topic. We could spend the whole year talking about it and still just be scratching the surface. The arguments for and against AI in healthcare fly in both directions. On one end of the spectrum there’s regulation that says that if you are an insurance company, you cannot price insurance based on genetic information. On the other end of the spectrum, if you are a patient and someone is able to identify that you have a mutation, you can now take this information, that you would not otherwise know, to make lifestyle changes. Bias is tied to both actions and knowledge. Every piece of data and knowledge can have both good and bad outcomes.
Digital health entrepreneur? Don’t make the journey alone. Learn more about the StartUp Health Community and how StartUp Health invests.
Sign up for StartUp Health Insider™ to get funding insights, news and special updates delivered to your inbox.
