Imagine a woman is working on a home improvement project indoors while her husband clears leaves from their house’s gutters. It’s the perfect idyllic Saturday — until she hears a sickening thud just outside.
Looking out the window, she notes with horror that he’s fallen from the ladder. Worse, he’s not moving. Frantic, the woman scoops up her cell phone and rushes outside. He’s breathing, but unresponsive. She calls 911.
“Ma’am,” the dispatcher tells her after a few fact-finding questions, “my computer indicates your husband may be suffering a cardiovascular issue. An ambulance is already on the way.”
How do you think the dispatcher came to this unlikely conclusion? Believe it or not, it has nothing to do with the questions she asked. Instead, a computer system that “listens” to all emergency calls made the determination. In the real-life version of this scenario, its diagnosis was just as accurate. The experimental system, which “taught itself” about symptoms of cardiovascular illness via medical data it was fed, detected the husband’s rattling breath in the background and alerted dispatchers to the possibility.
Situations like these highlight one of healthcare’s most exciting upcoming technologies: Machine learning, defined by one expert as computer systems that “improve [their] performance with experience.” As part of the larger big data movement, the technology holds immense potential in healthcare settings. Here’s what you need to know:
Self-Taught Systems, Prediction, and Diagnosis
Of course, we’ve seen no shortage of technologies promised to transform healthcare in the past decade, with a relatively small percentage have ultimately realized that potential. It will be a while before machine learning sees widespread adoption, but it already holds more promise than the average tech fad because it makes novel use of a skill computers are already very good at: sifting through large collections of data in search of patterns humans would struggle to detect.
The above-mentioned “roof” example provides a good example of this concept. The computer system, utilized by first response dispatchers in Copenhagen, was not programmed to look for specific symptoms. Instead, its creators fed it a collection of previous data related to cardiovascular illness — most notably calls to emergency dispatch — and told to make its own connections. A system “trained” in this way may notice certain words and phrases (“isn’t breathing right” or “clutching his chest,” for instance), background noises, or even tonal inflections build a pattern based on frequency and other trends, and use the connections it makes to inform dispatchers on future calls.
This isn’t to say machine learning technology’s diagnostic utility stops at cardiovascular problems. Researchers have demonstrated a long list of situations in which machine learning tools can keep pace with or even outperform trained medical professionals. In one instance, a tool studying pediatric brain MRI data was able to identify infants who might be at risk of an autism diagnosis later in their childhood.
In another, university researchers built a program that predicted acute myelogenous leukemia (AML) remission and relapse with 100 and 90 percent accuracy, respectively. Researchers “trained” the latter system by feeding it data on bone marrow from AML patients. And those are just two of dozens of examples in which machine learning tools come to highly accurate (sometimes astoundingly so) medical conclusions based on the data they’re fed.
To be clear, many examples of machine learning technology’s potential still occur at the academic level today, with researchers feeding their systems anonymized data and reporting their findings as a proof of concept. Even then, however, the outcomes reported in the news present an undeniably exciting view of the near future: one in which medical professionals bolster their diagnoses with the kind of data-backed insights entire teams of medical researchers may never be able to put together on their own.
Current and Near-Future Use Cases Underscore Machine Learning Technology’s Value
More, with all the promise machine learning tools hold, their value will eventually go beyond diagnosis. Larger health organizations have expressed a keen interest in machine learning as a larger analytical or fact-finding tool, with a large percentage in one survey (39 percent) claiming they’d invested in the technology for that purpose.
Population health presents one intriguing use case here. A system fed various demographic data, for instance, might conclude that residents of certain neighborhoods are at risk for a specific set of illnesses: a finding that could lead organizations to offer stronger preventative programs or targeted marketing campaigns in those areas. Similar tools could be used to optimize talent distribution for organizations impacted by the large and growing physician shortage, ensuring patients across a system get the best possible access to care.
By the same logic, short-term staffing services such as locum tenens organizations could soon benefit from machine learning solutions and their ability to identify patterns. As a purely hypothetical example, a system could determine that locum professionals with certain professional experiences tend to receive higher patient and client satisfaction scores in certain regional areas or clinical settings. While this pattern may not be immediately apparent to human eyes, a computer’s ability to draw direct lines between seemingly unrelated things could make for better talent matching and stronger overall talent pools.
Other examples extend into live medical settings, with health organizations achieving better health and economic outcomes thanks to machine learning initiatives. In one notable example, a commercial software package helped a Chicago-area hospital lower readmission rates by crunching data from past readmitted patients. Once it had determined factors and patterns related to readmissions, it used the data to apply a risk score to patients under the hospital’s care, helping the organization preemptively address problems with real potential to impact its bottom line. The same tool used similar data-mining techniques to find patients at risk of falling or contracting sepsis, further boosting economic outcomes.
If nothing else, these examples show how one high-level technique — namely, giving a system access to a large store of data and seeing what connections it comes up with — can provide utility in a multitude of settings. In a data-driven field like healthcare, that makes machine learning technology’s success all but a foregone conclusion: When data influences so many facets of an industry, from marketing and staffing to basic care decisions, any tool that elevates basic fact finding to superhuman levels is one too powerful to stay in an academic setting for long. However your organization ultimately deploys it, expect machine learning to play a serious role in business and care decisions soon.
