The Breakdown

Machine learning in healthcare

By Dr. Sanjay Basu

“The Breakdown” is a series examining the latest developments in health and wellness, sharing what the research says about some of today’s most pressing issues, what it doesn’t, and where we still have more work to do. At Collective Health, we take a rigorous approach to our research to ensure we’re making informed decisions that improve the health of our members. In “The Breakdown,” we’ll share the insights we’ve learned through that process. You can read more about our approach in STAT.

Machine learning. From email filtering to smart speakers and fraud detection, we’re constantly hearing about the latest app or product to employ its use. You’re probably being sold solutions underpinned by “sophisticated machine learning” capabilities, but vision and promise are often detached from reality when it comes to this type of technology. How should we turn a critical eye to it to ensure your people are getting appropriate care? Here’s what you should be paying attention to.

What is machine learning?

In a recent peer-reviewed tutorial on the subject, we defined machine learning as an exercise in which a machine “is given experience (say, a set of input–output pairs) and learns to perform a task (say, predicting hospital readmission).” The machine learning applications most common in healthcare are predicting outcomes (which patients will do better or worse when given this medication?) and classifying data (which patients with a spot on their CT scan have a brain aneurysm versus a benign anatomical variant?).

Increasingly, machine learning tools have been used to help employers, insurers, and healthcare providers make predictions about healthcare cost, utilization, and quality. One example: helping to identify which diabetes patients may need more intensive care support and which patients would benefit from more intensive treatment of their blood pressure. These use cases seem like an excellent application of technology in theory, but—as is often the case with emerging technologies—we need to take a step back and look at the bigger picture to ensure we’re using the technology in meaningful, useful ways that are safe and reliable.

Takeaway for HR and benefits leaders:
At its most basic, machine learning is simply giving a computer information then letting it predict an outcome or classify data based on that information. There are several applications in healthcare, but we need to take a step back before going all in on them.

Make sure your vendor has expertise on healthcare processes

While it can seem like an abundance of predictive healthcare information is helpful, that isn’t always the case. In healthcare, algorithms often exclude content expertise. For example, one recent machine learning algorithm claimed to identify which patients who went to the hospital with pneumonia ultimately needed more or less care support. The algorithm said, rather counterintuitively, that those with asthma needed less support—something a physician would’ve known to be false—because it didn’t know the hospitals were automatically triaging patients with asthma to the intensive care unit. They were, in fact, receiving additional care support, which the algorithm had confused with having lower risk.

The best predictive machine learning models will often combine machine learning methods with detailed content expertise, rather than replacing one with the other. In one study, a machine learning algorithm went through electronic health record data to flag patients likely to benefit from palliative care services to a care team. Rather than dictating the course of action this team would take, it allowed the team to use their healthcare expertise to then proactively make recommendations.

Beyond medical treatment and diagnosis, machine learning also holds promise in the processes holding back health systems the most: more menial tasks like billing, scheduling, and processing claims. Contrary to what popular TV shows suggest, it’s rarely medical diagnosis or treatment selection that’s the problem with healthcare delivery; it’s usually executing the treatment plan well in the context of the mess of insurance, forms, confusion, and paperwork. By applying machine learning to these processes, we can reduce costly healthcare errors.

Takeaway for HR and benefits leaders:
Ask your vendors how they incorporate healthcare expertise and context into their evaluations, and consider working with vendors who apply themselves to some of the most human-error-prone tasks like paperwork and scheduling.

Look inside the black box

If you’re thinking about adding machine learning-powered tools to your benefits program, make sure to ask vendors about strategies to look inside and understand how the machine learning actually works. In particular, it is useful to see explanatory models that help identify how different people are treated by the model and whether that follow common sense and medical expertise, and ask your vendors to show performance metrics on more than one dataset (“external validation”).

Takeaway for HR and benefits leaders:
It should serve as a red flag if machine learning vendors aren’t willing to share metrics or discuss their model’s ability to predict/classify data in datasets specific to you.

Protecting privacy and against bias with machine learning

However, one of the criticisms holding machine learning back in healthcare is its tendency to reinforce human bias and invade privacy. While there are valid reasons for those critiques, we can, in fact, design machine learning algorithms to help detect bias and improve privacy by training machine learners to help identify our vulnerabilities. At Collective Health, we’ve been using machine learning as a team of bloodhounds to examine the different players in the healthcare system—hospitals, doctor’s offices, etc.—and identify where these systems are prone to threats and need to strengthen security and privacy. We’ve similarly applied peer-reviewed strategies to develop risk scores that avoid bias and increase accuracy when we identify people who may benefit from care interventions.

Takeaway for HR and benefits leaders:
Ask your vendors whether they use processes to detect bias in their algorithms and what strategies they have to protect privacy beyond simple “de-identification” measures that are easily overcome by hackers.

Ultimately, machine learning shows promise in both the business and diagnostic sides of healthcare—provided we take context into consideration and ensure our expertise on complex healthcare processes are woven into the development and validation of machine learning tools. As we continue to apply machine learning to common healthcare tasks, we’ll be able to refine algorithms to ensure they’re useful tools to augment our human abilities. With the right guardrails, machine learning can help us usher in a safer, more effective future for healthcare.

In future articles, I’ll share the latest updates on health and benefits-related topics from some of the industry’s leading journals. Sign up for our email newsletter to get more information like this in your inbox.

Our customers include:
Red Bull Activision Blizzard
Zendesk Counsyl
Red Bull Trace3 DataSafe
Red Bull
Activision Blizzard
Zendesk Counsyl
Trace3 DataSafe
Red Bull Activision Blizzard Zendesk Counsyl Trace3 DataSafe