Machine Learning in Healthcare: How it Supports Clinician Decisions—and Why Clinicians are Still in Charge

Many clinicians view the practice of medicine as an art rather than a science. There’s a craft to extrapolating diagnoses from symptoms, determining which pieces of information are important and which are distractions, designing and refining treatments, and delivering compassionate care. So, when machine learning is introduced as a clinical decision support tool, some clinicians respond, “we don’t need a machine to tell us what to do; we understand our patient base better than any technology can.”

Machine learning in healthcare is frequently compared to self-driving cars, in which robots control the vehicle and humans are mere passengers along for the ride. So, similar perceptions may come to mind when machine learning is mentioned in healthcare, many of which are misguided. Unlike potential applications in transportation, the application of machine learning to healthcare is intended as a powerful tool to support human decisions—not as a fully automated process to supplant it.

Everyone in healthcare—clinicians, executives, and administrators—should be enthusiastic about the opportunities machine learning presents for providing clinical teams with more precise, faster, data-driven insight to help them achieve better outcomes for their patients. When made accessible to the right person (clinician) in the right place (point of care) at the right time (real- or near-real time), machine learning-enabled models can give clinicians more pertinent information to support their patient care decisions.

This article explores machine learning and its applications, benefits, limitations, and potential impacts on healthcare. It not only describes how machine learning supports clinical decision making, but also clarifies that clinicians—not predictive models—will continue to oversee and be responsible for their patients’ care.

How Machine Learning in Healthcare Works

Machine learning in healthcare most often refers to predictive analytics models that are refined using machine learning algorithms. It is the process of analyzing a dataset for patterns and applying those patterns to new data that allows a model to offer a prediction of the likelihood of a defined clinical outcome of interest, whether that is a positive outcome that the clinician wants to achieve for the patient, or a negative outcome the clinician wants to avoid for that patient. After the initial deployment of a machine learning model, it should be repeatedly retrained as the data set evolves and grows, and, whenever it is brought to a new setting, should be retrained on that specific dataset to ensure applicability to local patterns.

Health Catalyst® applications apply predictive analytics algorithms using data in the enterprise data warehouse (EDW) to offer insight into each patient’s risk of a positive or negative outcome of interest, as well as the factors contributing to that patient’s level of risk. This is most often delivered as a worklist. For example, machine learning can predict which patients to monitor more closely for central line-associated blood stream infection (CLABSI), because they are exhibiting similar patterns or characteristics to past patients who had higher incidences of CLABSI. These machine learning insights can be presented in an actionable dashboard for clinicians at the point of care.

The presentation of the results from predictive analytics algorithms can separate those factors that are malleable to intervention from those that are not. For example, there’s no changing a patient’s gender, age, or existing diagnosis, but the prescribed interventions can change. In the case of CLABSI prevention, the predictive algorithm can analyze the contribution to risk of CLABSI of the duration, number, and type of central lines, and present this insight to the clinical team. It is up to the clinician to decide whether and how to act on that information.

The Clinician Is Still in Charge

The previous CLABSI example illustrates how the clinician is still in charge. Just because someone is at risk for CLABSI, doesn’t call for removing the line; the patient may still require central line access for their care, and there may also be reasons why it may be inadvisable or impractical to attempt to remove or replace it. The model just surfaces the information that the patient is at risk so the clinician can take that information into account. The clinician can consider a variety of approaches:

  • Should I monitor the patient more closely?
  • Do I need to order more tests?
  • Is a different type of access less prone to infection?
  • Was the central line inserted during an emergency, which makes it more likely to have been contaminated?

The predictive algorithm can only present the likelihood based on similarity to patients who did develop CLABSI, and offer that information to the clinical team who can consider the entire clinical picture and determine the most appropriate plan for the patient to manage that risk.

Understanding the Differences between Predictive and Prescriptive Analytics

Predictive analytics only surface information and risk; however, machine learning can also be used to suggest orders for easing clinician workflow. For instance, it can offer the insight that the predicted risk would be significantly reduced if a specific medication were added, and give the clinician a shortcut to place such an order if they agree with that suggestion, after considering the whole clinical picture. This is the important distinction between predictive and prescriptive analytics.

A predictive analytics solution may predict that a patient is at higher risk for central line infection, and present the contributing risk factors. While it can be extended into the prescriptive realm by also offering an additional step of suggesting the type and dosing of antibiotics that could reduce the risk, ultimately, it’s up to clinicians to decide whether to place the order. This is as close as the model gets to prescriptive. It can offer a specific suggestion, informed by applying a machine learning predictive model to the health system’s data, and comparing a specific patient’s characteristics with those in the predictive model.

Self-driving cars, on the other hand, are truly prescriptive; they not only calculate the route, but also take action: they turn the steering wheel, accelerate, decelerate, and stop. This is not the role machine learning plays in healthcare; rather, machine learning supports clinical teams in providing high quality care.

Machine Learning Tackles Data Variability and Complexity

Machine learning can quickly parse the many patient attributes and complexities that help make medicine more of an art than a science. Clinical practice guidelines deliberately remove complexity and isolate patients to a single problem at a time, but clinicians know that real patients are more often a complex mix of multiple diagnoses, characteristics, and other factors that should be considered in determining a plan for treatment.

For example, consider the variability and complexity of data assigned to a patient on blood pressure medication. A single patient may be prescribed many different blood pressure medications over the years by various providers in various settings, from the attending physician, to the primary care physician, to the cardiologist. New formularies become available, insurance coverage changes, the patient moves from one healthcare setting to another and develops new co-morbid conditions, and provider preferences change. The patient goes through these medications with varying degrees of success and side effects. She may stop taking certain medications because of those side effects or changes in the out-of-pocket cost.

The Joint National Committee (JNC) guidelines on treating blood pressure can factor in a limited list of diagnoses and labs. They cannot take into account a patient’s entire medication history, which, as described previously, can be extensive. Machine learning can handle this level of complexity, especially with access to a vast repository of data from multiple sources.

As with any technology, machine learning has its limitations, which healthcare organizations must thoroughly understand and communicate to their clinical teams.

Machine Learning’s Limitations

There are limitations to what predictive analytics models can take into account, and clinicians need to be aware of these potential pitfalls and information gaps. As with any industry, it’s important to ensure that new technology is implemented with realistic expectations, in the appropriate context of existing clinical processes. Regardless of where a recommendation originates, it’s still the clinician’s responsibility to own the decision, and much can be done with designing models to ensure that this bright line remains very clear. Machine learning can be used in the same way papers and textbooks have always been referenced for decision support. But, ultimately, clinicians make the decisions about every patient’s treatment.

Machine Learning, Algorithm, and Clinical Decision Support Governance

The FDA recently announced that it is likely to publish guidance on digital health devices in the fall of 2017. If, as expected, this is interpreted to include clinical decision support and/or the use of algorithms to deliver information into the workflow, this new arena of oversight may include some of the applications Health Catalyst is developing. The new FDA Commissioner, Dr. Scott Gottlieb, published a blog post titled Fostering Digital Innovation: A Plan for Digital Health Devices, in which he indicates the FDA will take a risk-based approach, making sure to continue supporting innovation, but also paying attention to safety. The plan describes a program, like Meaningful Use, where the FDA will look at “employing a unique pre-certification program for software as a medical device (SaMD).”

Rather than requiring post-market surveillance or a detailed pre-market review, the FDA would focus regulatory efforts on those features which warrant oversight due to higher potential risk to patients, and plans to continue to clarify which types of technologies will generally fall outside FDA regulatory scope due to the low risk they pose to patients. Gottlieb goes on to suggest the FDA is looking into a third-party certification program, which is firm-based rather than product-based, and focuses on assessing whether the company “consistently and reliably engages in high quality software design and testing (validation) and ongoing maintenance of its software products” rather than requiring rigorous pre- and/or post-market testing and surveillance, product by product.

To ensure patient safety, Health Catalyst, which has started deploying algorithms that, for the first time, can deal with the real-life complexity of individual patients and populations, will remain mindful of its accountability for the quality of data and the processes that put this data in front of clinicians. It will also continue to participate in and follow any regulatory development process for machine learning, and will take steps to comply with emerging requirements.

The Next Steps in Machine Learning’s Game-Changing Trajectory

Machine learning is a game-changing tool in how it can surface insights for clinicians at the point of care, but it is not a self-driving car; it allows healthcare organizations to achieve more complex analysis, more rapidly, to provide a powerful new input into the care decisions that clinicians will continue to make. Health Catalyst already sees the positive results among health system partners that have applied machine learning to various complex data patterns, and is excited to continue working with healthcare organizations to explore new applications of machine learning that will improve patient care and outcomes.

Additional Reading

Would you like to learn more about this topic? Here are some articles we suggest:

  1. There Is A 90% Probability That Your Son Is Pregnant: Predicting the Future of Predictive Analytics in Healthcare
  2. Machine Learning 101: 5 Easy Steps for Using it in Healthcare
  3. How Machine Learning in Healthcare Saves Lives
  4. An Inside Look at Building Machine Learning for Healthcare
  5. Prescriptive Analytics Beats Simple Prediction for Improving Healthcare
Loading next article...