What Clinicians Should Know About Healthcare AI, Part 1: Why Care About AI?
I’m glad you’re here! That’s because I believe artificial intelligence (AI) has incredible potential to transform healthcare for the better, but to do so, more people in healthcare need to understand what it’s all about.
My perspective is as a cardiologist finishing up a data science fellowship at Brigham and Women’s Hospital, which mostly involved creating AI models for ECGs and echocardiograms. The COVID-19 pandemic struck shortly after I started this work, and it laid bare the many dysfunctions of the US healthcare system. It also led me to realize how tech could help address many of these issues, from health inequity to clinician burnout.
Though I was heading toward becoming a physician-researcher up until the pandemic, now my passion is helping figure out how the healthcare system can more effectively use technology and data to improve patient care and advance our medical knowledge. I believe a key part of effectively incorporating tech into healthcare is having more clinicians involved with the process. There are challenges with using tools such as AI in healthcare, but these are challenges we can address proactively as we help build systems that work for us. So, welcome to this short series about healthcare AI.
What Is AI?
AI refers to using computers for tasks that simulate human abilities such as visual perception and speech recognition. Chances are, you’re already using AI. AI is working behind the scenes when you:
- check out recommendations from Netflix or Amazon
- use Siri or Alexa
- unlock an iPhone with Face ID
- get driving directions from Google Maps
- get a fraud alert from your bank
- filter e-mail with a spam filter
- search with Google
Think about how AI has made our lives easier. Think about what it can do in healthcare.
A subset of AI is machine learning (ML). When people talk about healthcare AI, they’re almost always referring to ML. Because of this, the terms AI and ML are often used interchangeably in healthcare. In ML, computer systems are programmed to “learn” from data. For example, you can set up an ML algorithm and provide it with 10,000 photos of cats and 10,000 photos of dogs. The system will start to find patterns in the data, and then you can use the model that is generated to differentiate between cats and dogs in new photos.
Similarly, you can set up an ML algorithm and provide it with spam e-mails and non-spam e-mails to create an ML model to detect spam. You can create a spam filter with this model and allow users to mark spam e-mails that make it past the filter. These missed spam e-mails can be continuously incorporated into the ML model to allow the filter to keep up with the latest words and patterns used in spam e-mails.
Why Use AI in Healthcare?
We should use AI in healthcare for the same reason search engine companies have shifted to using AI — it was a better approach to what they do. Like the use of AI in search engines, some of the ways AI can be added to healthcare are seamless and may not be very obvious to users. After all, as users we didn’t differentiate search engines based on whether they were AI-powered or not. We simply kept using search engines that were great at responding to our queries and stopped using search engines that weren’t as good.
It’s important to keep in mind that AI is a group of approaches rather than one single concept. Some simpler ML approaches overlap with the traditional biostatistics approaches we’re familiar with such as linear regression. There isn’t a clear line between biostatistics and machine learning. Rather, we have at our disposal a spectrum of tools to make use of data, and we can select from them depending on the need. Now that we’re incorporating complex ML techniques to our repertoire, not only can we improve some of what we were already doing, but we’re also able to start tackling a whole range of issues that we didn’t have the capability to properly address before. For example, AI can help diagnose conditions at early stages by detecting patterns in radiology studies that are not apparent to the human eye. Since there’s potential to improve patients’ lives by applying these newer approaches, I feel exploring their use is our responsibility.
Aside from improving patients’ lives, AI can also make our own lives easier. My first encounter with healthcare AI was during my cardiac MRI rotation in fellowship. For the first few studies, I painstakingly used the mouse to draw the borders of cardiac structures for one slice after another. This was initially interesting, but as you can imagine, it quickly became tedious. I started to wonder whether I was putting myself at risk for carpal tunnel syndrome. Then, someone pointed out a button in the software for an AI feature, and I tried it on the next study. I drew the borders on a couple of slices and then clicked the button. It was like magic — the software instantly filled in the borders for the remaining slices almost perfectly. I only had to make a few minor adjustments. It left me time to think about the study instead of mindlessly clicking through slices.
In addition to helping with tedious tasks like this, AI systems can potentially help us avoid mistakes or missed diagnoses. It can also help us improve our clinical decision making. As we know, there are many situations where the best management approach is unclear, especially for conditions that typically don’t get much research attention. Using AI can help set up a learning health system, where data from everyday practice is continually incorporated to gain insights and improve how future patients are managed.
What’s Next?
I’d love to say that healthcare AI will survive through the current hype and be a game-changer in the near future, but I don’t yet think we can take that for granted. The progress of AI in general has been hampered by periods of disillusionment called AI winters. Healthcare AI is not immune to this, and I believe the next couple of years will be a make-or-break period. If there are too many missteps and we continue to set unrealistic expectations, there’s a risk of declarations that “AI doesn’t work for healthcare.” Though some projects would survive, we may not see the transformative impact that is possible with healthcare AI.
How can we help this transformative impact come about? First of all, you’re already taking a step by deciding to read this post. You won’t be someone who says “AI doesn’t work for healthcare” because you know that AI encompasses a range of approaches that blends into traditional biostatistics approaches. I hope you’ll join for the rest of this series and explore other resources about healthcare AI. The better clinicians understand what AI is and its capabilities and limitations, the more we can get involved with helping incorporate it judiciously into healthcare. Healthcare AI holds great potential. If we get this right over the next few years, we could be at the verge of significant advances in how we provide patient care.
There will likely be a total of 5 posts in this series, and I’ll be posting on LinkedIn and Twitter when each comes out. If you prefer a more detailed video introduction to healthcare AI, I highly recommend AI in Healthcare Specialization on Coursera. If you’re interested in coding AI algorithms, see this list of resources that my colleague Pierre Elias and I recommend. See you soon for Part 2: What Does AI Model Creation Look Like?
