Machine Learning 101: The Healthcare Opportunities Are Endless

Both supervised and unsupervised machine learning can help executives better enhance care delivery, Stanford algorithms expert says.

To understand how machine learning can aid healthcare organizations, healthcare executives first must have a basic grasp of what machine learning is and what it can do. “Machine learning is about discovering new knowledge,” said Zeeshan Syed, director of the clinical inference and algorithms program at Stanford Health Care and clinical associate professor, anesthesiology, perioperative and pain medicine, at the Stanford University School of Medicine. “At a high level, artificial intelligence is getting an agent, software, to behave like it’s smart. One example might be a thermostat. If it’s cold, the thermostat turns the heat on. That’s a system behaving in a smart way, a very crude form of artificial intelligence. Knowledge you are using is pre-derived and embedded into the device. Machine learning goes a step further: How do we derive this knowledge that we are using? It’s knowledge derived from the data itself.”

So in a nutshell, machine learning is all about new knowledge that leads to providing intelligence. And there are two different kinds of machine learning – supervised and unsupervised.  “With supervised, the goal is you have some data and you have an outcome of interest, and what you are interested in learning is how is the data related to the outcome,” Syed said. “So in the context of patients, you might have a lot of information on the patients, their lab values, medication histories, and so on, and you want to predict an outcome, whether they will experience a cardiac event. The knowledge is the relationship between the attributes of the patients and the outcomes you are trying to discover. That is supervised learning.”

With unsupervised machine learning, there is no target or label or outcome in mind. “You have a bunch of data and you are trying to find structure in that data, something that is interesting in its own right,” Syed said. “Say you have all of that same patient information: Can you identify the kinds of patients that exist, say maybe five different classes of patients? It’s trying to find statistically interesting things about the data itself, and not in that context of relating it to something else.” For example, what patients in a database are outliers or anomalies not relevant to a predefined endpoint? “Just which patients look anomalous because they have unusual combinations of labs and comorbidities,” Syed said. “You are looking for interesting structures within the data, not to classify something but related to the properties of the data itself. That is unsupervised machine learning.”...