As the American healthcare system makes strides to develop financial incentives to drive patient health outcomes rather than volume, providers are making greater efforts to use new technologies that address inefficiencies and better harness the mountain of data stored in electronic health records.
Much of the current conversation around machine learning (ML) in healthcare remains theoretical because countless regulatory hurdles still stand in the way of many of its applications like robotic surgeries and personalized medicine. But machine learning is already transforming healthcare by optimizing provider schedules. The result: more patients are getting care in faster time.
Why does ML-powered scheduling matter?
Appointment scheduling is a clear first step in receiving health care services, but breakdowns in that communication often result in missed care visits, inconsistent follow-ups, last-minute cancellations and long wait times to see a doctor. To combat the risk of patients not showing up to their scheduled appointments, many healthcare organizations are doing what the airline industry already does: double-booking patients to avoid any gaps in the calendar and missed revenue opportunities. As a result, patient appointments are oftentimes delayed, and providers end up struggling with unpredictable schedules and cramming in too many appointments into too small windows. And for the patients that don’t show up, those missed appointments are lost care opportunities. It is also a warning signal that the patient is less engaged with their health, which puts them at a higher risk for worse health outcomes.
With machine learning, providers now have a way to predict, with consistently increasing accuracy, which patients will show up for their appointments and which will not. For those patients most likely to make the appointment, providers can prepare and offer longer appointment slots with the confidence that the time will be filled. For high-risk patients, more targeted and intensive pre-appointment outreach can help address the barriers that are keeping them from seeing their providers such as helping patients find a ride through ride sharing services like Lyft.
Lessons learned from 5 million patients
Armed with data from a population of 100,000 providers and 5 million patients, the Luma Health Machine Learning team sought to tackle the problem of predicting which patients were likely to miss appointments. From this analysis, the team made three key insights.
First, it is critical to build models from data that closely resembles the target population. Early in the process, it was discovered that building a model to predict cancellations using heterogeneous data simply does not work. Using country-wide data, for instance, to predict cancellations at a dermatology clinic in Seattle did not prove effective. Even two practices on the same street may serve completely different populations. An ideal predictive model uses only a target organization’s historical data.
Second, get comfortable with imperfect modeling. It is impossible to get a model that is 100 percent accurate, so it is critical that you first determine an acceptable level of accuracy that will provide sufficient insight to drive better outcomes and workflow efficiency. In this testing environment, models were made to prioritize accuracy in predicting cancellations rather than predicting which patients will show up, in order to meet the project’s goals.
Third, know when to quit. In technology markets, the approach is often that there’s no wall that can’t be broken. But for certain providers, there simply will never be enough data to build an accurate predictive model. Oftentimes data still lacks the insight into all the factors that influence cancellations, such as if a patient has a temporary caretaker, or if that patient is taking care of a sick loved one.
The use of machine learning in healthcare will continue to expand for use cases such as detecting medical anomalies, predicting outcomes and suggesting treatments. It has even been predicted that machine learning may replace some of the activities performed by physicians. What is certain is that machine learning is fundamentally changing the way patients and providers connect. People become doctors to treat patients, and with the use of machine learning physicians can now dramatically improve and maximize time spent with patients — starting today.
###
By Dr. Tashfeen Ekram, Chief Medical Officer and Co-founder Luma Health