The technology has matured enough, and the algorithms have been honed in recent years, to the point where there are now two types of artificial intelligence: “old school AI and new school AI,” as medical journalist Paul Cerrato proposed during a recent HIMSS20 Digital Session.
For examples of the former, think DeepBlue – the chess-playing computer from IBM that beat Garry Kasparov in 1996.
For the latter, think AlphaGo, developed by Google’s DeepMind technologies, which in 2015 became the first computer to beat a human at the ancient strategic game of Go (and which now has since been succeeded by even more powerful iterations).
Deep Blue was trained with the rules of chess and then with simulations of millions and millions of actual games,” Cerrato explained. But AlphaGo took a “completely different approach: the program was only coded with the basic rules, and once it had those rules, it taught itself to be a champion.”
In healthcare, most providers still rely on old-school AI – “static, rule-based, pretty much encyclopedias,” as Cerrato explained it. “It does an adequate job, but we can do a lot more.”
Increasingly, new-school AI and machine learning are transforming clinical processes, with advances in image recognition, predictive analytics and neural networks having impacts on cardiology, dermatology, oncology, critical care and other specialties.
That’s certainly true at Mayo Clinic. Dr. John Halamka, who joined the health system as president of its Mayo Clinic Platform in January, also recently cowrote a HIMSS book with Cerrato on the ongoing AI-powered evolution of decision support.
During their HIMSS20 presentation, Reinventing Clinical Decision Support With Machine Learning, Halamka explained some of the most promising CDS advances there, and how they’re informing treatment decisions for an array of different specialties – and helping shape its understanding of COVID-19.
“New school” machine learning is impacting screening and diagnosis diabetic retinopathy and melanoma, helping with the management of sepsis – “and there’s even a whole field now known as digital pathology, analyzing slides with the help of AI,” said Halamka.
In a country where 18 million diagnostic errors lead to 74,000 deaths on average each year, those advances – helping avoid missed diagnosis, misdiagnosis, delayed diagnosis or overdiagnosis – have enormous potential to drive big improvements in quality, safety and outcomes, said Cerrato.
There are big challenges, to be sure, he said, with the “black box dilemma,” biased data sets and relative lack of empirical evidence so far just a few of them.
But the change from rules-based decision support to learning systems powered by AI holds immense promise as the technology continues to evolve and mature.
“Imagine the power to an AI algorithm if you could make available every pathology slide that has ever been created in the history of the Mayo Clinic. That’s something we’re certainly working on.”
Dr. John Halamka, Mayo Clinic
Success, of course, depends on curated and high-quality data. At Mayo Clinic, said Halamka, some of it is structured (Epic EHR data). Some is unstructured (CTs and MRIs). “Some has never been digitized – like 30 million pathology slides,” said Halamka.
“Imagine the power to an AI algorithm if you could make available every pathology slide that has ever been created in the history of the Mayo Clinic,” he said. “That’s something we’re certainly working on.”
Not for nothing, “I’m a big believer in these algorithms,” said Halamka. “I tested myself: Got a 12-lead ECG and ran them through all the algorithms.”
But for patients at Mayo Clinic and other health systems around the world, machine learning-powered CDS is helping provide more accurate answers to critical questions: “What are the likely complications you will experience in surgery based on patients like you? What are the likely drug targets that will be successful?”
For example, at Mayo there are newly-developed machine learning algorithms that can diagnose heart ailments such as atrial fibrillation, hypertrophic cardiomyopathy and low ejection fraction, Halamka pointed out.
“Typically these have been invasive procedures that have required hospital stays. And now from simple telemetry, we can, with very high sensitivity and specificity, determine what is likely to happen to you.
“And in a world of COVID-19, not surprisingly, we’ve used AI and machine learning approaches to look at the world’s literature, and the entire corpus of historical data, to identify what is likely going to identify a COVID-19 cure,” he added.
For just one example of that, Halamka explained (with the caveat that he is “apolitical” and was saying so without “connection to any person or policy”) that Mayo Clinic recent developed an AI algorithm that “looked at the effectiveness of antimalarial medications in patients with COVID-19.”
The soon-to-be published findings, he said, show that “administration of medications like hydroxychloroquine actually kill you twice as fast. So it’s really, really important, especially when you’re in the fog of war, which is the COVID-19 response, to be able to look at historical data and current data using these machine learning tools and discover patterns.”