Back to AI in healthcare – we’ve looked at this from a number of angles, but what about some of the pros and cons of using AI/ML systems in a clinical context? And also, what about how to conquer disease with AI models?
There’s a broader theory that AI is going to allow for trail-blazing research on everything from cancer and heart disease to trauma and bone and muscle health – and everything in between. Now, we have more defined solutions coming to the table, and they’re well worth looking at!
In this IIA talk, cardiologist Collin Stultz talks about the treatment of disorders, and new tools, starting with a dramatic emphasis on heart disease.
Starting out, Stultz points out what the town criers of healthcare are telling us – that cardiology disease is endemic, and pervasive, and prevalent, and that many people have it.
When I teach cardiovascular physiology or pathophysiology, I typically ask the students or observers in the audience how many people here have cardiovascular disease? Everyone does. It is ubiquitous. It is a disease of the juvenile. Some of us will die from it, some of us won’t, and we just don’t know the answers to it.”
When he’s talking about rates of heart disease across the country and across the world, (pay attention to this part), he makes a particular point that we’re considering a spectrum of disorders, and making sophisticated diagnoses is a big part of the challenge.
Any aspect of this sort of process, he says, takes an information relay, from the sender to the receiver. In this context, he notes, the receiver is a healthcare provider.
Stultz talks about the two-fold goal of helping people to live longer and to feel better, and mentions some of the types of data at the clinician’s disposal, showing how these are ultimately important. But then, he makes a really interesting point about humans versus AI and their different approaches to systems – take the example of EKGs: as he points out, humans can’t take in all of the detail in one of the scans. They are obligated to ignore some of the detail and focus on other parts of the data.
AI doesn’t have this challenge – it can look at, as Stultz says, “every pixel,” every part of an image or data set, and make decisions.
Talking about a “healthy skepticism” in using deep neural networks, Stultz mentions a threefold goal: to integrate data, consider all sources and improve performance.
One of the challenges, though, he points out, is knowing when systems fail, and how.
“If the model is not accurate 100% of the time, then there is risk. These are high-stakes decisions. So if (there is) a patient … and (as a care provider or clinician) I have a medication that (the) patient needs to live longer, and a model predicts (that) this patient doesn’t need this medication, I’ve missed an opportunity to intervene and to make someone live longer, to avoid a disastrous event. So it is important in these models, not only to talk about their accuracy, and how they perform over a large cohort, but really to give some insight as to when they will fail, and deciphering when a given prediction is appropriate for a given patient … based on this.”
He also suggests that doctors have to ask themselves – even though the system may be so great in general, is it appropriate for my specific patient?
Quoting George Box, Stultz explains how models are inherently simplifications of complex real-world systems, and how, in some senses, “all models are wrong.”
Still, he says, new wearable technology gives us tools to improve diagnosis and care.
“We are encouraged by this,” he says, adding that the goal is to generalize to an outpatient setting what people get in hospitals.
Video: Collin Stultz has a lot of experience, and it’s shown him how AI can revolutionize cardiology
“We believe that we can do this by taking simple monitors with deep learning, to get the same information that one could get in the hospital,” he says.
That’s it, in a nutshell. Clinicians around the world are excited about using wearables and other tech to better customize medical care. It will save lives!
Read the full article here
