Written: September 2022
Radiology is the area of healthcare which stands out as a candidate to take advantage of AI. However, it is not the only area of healthcare which will benefit. In Robot Attitude: How Robots and Artificial Intelligence Will Make Our Lives Better, I described how AI can revolutionize the stethoscope, diagnose eye disease, predict falls in the hospital, enhance accuracy of emergency calls, forecast fatal conditions, optimize prescription of medicines, and diagnose disease from x-rays. AI systems are learning to diagnose disease across a wide range of medical conditions, and gradually they are becoming as accurate as human doctors. In time, the AI diagnoses will become more accurate than human diagnoses and it will be applied to most aspects of healthcare.
At the heart of AI is a branch of the science called machine learning (ML). As humans, we learn from experience. An AI lives on a server or mobile device and, like us, it learns from experience. The difference between us and an AI is the experience we learn from is not always written down, we just remember it. For an AI, the experience is in the form of data, a lot of it. An AI follows instructions which are based on algorithms which contain the instructions. An AI learns, using data and algorithms, how to identify patterns and make decisions with minimal human intervention.
An AI becomes intelligent after it has learned a lot. For example, let’s consider the weather. If a meteorologist considers a number of factors including the size and shape of clouds, temperature, dew point, wind direction and speed, and barometric pressure, he or she can forecast the weather for the next hour will be X. With a different set of conditions, the forecasted weather would be Y. If you applied machine learning software to thousands of conditions and resulting forecasts, an AI could be trained to forecast the weather more accurately. As more and more data with sets of conditions and actual weather are accumulated and submitted to machine learning, the accuracy of the forecasts would get better. With enough data and machine learning, the weather AI may produce better forecasts than meteorologists. The concept can be applied to numerous areas, including business, government, social sciences, and healthcare. In healthcare, machine learning can be applied to medical data and enable an AI to learn how to diagnose a medical condition. It might not be long before algorithms routinely save lives. A more recent area of AI in healthcare is in the field of cardiology.
Researchers at Stanford University, led by Andrew Ng, a prominent AI researcher and an adjunct professor there, have shown machine learning can identify heart arrhythmias from an electrocardiogram (ECG) better than an expert cardiologist. Heartbeat irregularities can be deadly and more accuracy in diagnoses can be important in some cases. Automated diagnoses can also be important in geographic areas where availability of cariologists may be limited.
One of the reasons AI is expanding its impact in healthcare is researchers are looking beyond imaging data and looking at forms of data such ECGs. An ECG contains a lot of data but some of it is subtle. The data may contain an irregularity which could lead to serious health complications but it can be difficult for a doctor to distinguish between benign irregularities and those which can require treatment.
The researchers at Stanford partnered with portable ECG device maker iRhythm. Deep learning involves feeding large quantities of data into a powerful computer and fine-tuning the parameters of the AI until it accurately recognizes an ECG which represented a serious health problem. They applied the AI to 30,000 30-second clips from iRhythm-wearing patients with varying forms of arrhythmia. To evaluate the accuracy of their algorithm, the researchers compared performance of the AI to five cardiologists on 300 undiagnosed ECGs. They had a panel of three experts to provide diagnoses they were certain of.
The diagnoses by the AI applied to the 300 cases turned out to be more accurate than the five cardiologists. Doctors amass an amazing amount of anecdotal data in their minds. They compare what they see to what they have seen in the past. Looking to the future, it will be hard for any cardiologist to compete with what an AI can do after applying machine learning to tens of thousands and eventually millions of data sets.
Eric Horvitz, managing director of Microsoft Research and both a medical doctor and an expert on machine learning, said others, including from MIT and the University of Michigan, are applying machine learning to the detection of heart arrhythmias. The research is promising but will doctors trust it? As I have written before, trust is a potential barrier. I am optimistic because researchers are beginning to open up the “black boxes” which contain the AI and show doctors what logic led to the diagnoses the AI made. Ten years from now, I believe the use of AI will be commonplace in healthcare settings and be a routine part of the workflow for diagnosing and treating patients.
See more at “The Machines are Getting Ready to Play Doctor“.