AI Model Improves Outpatient MRI No-Show Rate

September 11, 2020

A decision tree-based machine learning algorithm can help departments identify and contact patients at highest risk for skipping appointments.

Artificial intelligence (AI) can be used to decrease the number of patients who do not show up for their MRI appointments, saving practices resources and money.

In a study published this week in the American Journal of Roentgenology, investigators from Changi General Hospital in Singapore shared their experience with an AI algorithm that identified patients who were most likely to skip their appointments. By using this tool, they decreased their no-show rate by more than 17 percent from baseline in only six months.

When patients do not show up for scheduled appointments, said the team led by Le Roy Chong, deputy chief of the diagnostic radiology department at Changi General, healthcare resources are wasted. Successful interventions could decrease wait times, increase scanner utilization rates, and help control costs.

“For radiology departments, accurately predicting the individual patient risk of an outpatient scan appointment no-show may enable more intelligent and reliable intervention strategies, such as selective appointment overlooking, telephone, emails, and short message service text reminders,” they said.

Chong’s team called the rate of outpatient MRI no-shows a “pressing problem” because the demand for these scans is increasing globally, leading to longer wait times. In fact, they said, between 2016 and 2018, their institution maintained an upward-trending average no-show rate of 17.4 percent -- their baseline rate. To turn this tide, they opted to research whether using a decision tree-based ensemble machine learning algorithm (XGBoost) could help them pinpoint which patients were most likely to skip appointments to potentially change their behavior.

Using 32,957 outpatient MRI appointments from 25,461 individuals between January 2016 and December 2018, as well as a holdout test set of 1,080 patients from January 2019, the team trained their model. They identified the 25 percent of patients who were at highest risk for missing an appointment and determined that patient age and appointment time were the two factors that contributed most to skipped appointments. Most radiology departments and hospitals can gather such data from their existing information technology systems, the team explained.

Related Content: How to Minimize the Impact of Patient No-Shows

For six months, the team had MRI technologists call these patients with reminders about the time and location of their scan. Technologists were already routinely contacting roughly the same number of patients, so this initiative did not increase their workload.

But, at the end of the intervention, the team determined this targeted approach was 4.2 times more productive than calling random patients with reminders. In fact, among patients who were successfully contacted via telephone, the hospital’s outpatient MRI no-show rate fell to 15.9 percent compared to the 19.3 percent from the immediate preceding 12 months.

The no-show rate among patients the technologists were unable to reach was 40.3 percent, indicating that the model has a strong prospective predictive ability. It also highlights the possibility that the algorithm could be even more effective, Chong’s team said.

“We anticipate that a larger decrease in no-show rates could be achieved if the predictive model was set at a lower threshold, albeit at the cost of having to make an increased number of reminder telephone calls,” they said.

The overall improvement rate from baseline, they said, was 17.2 percent – a change they estimated equates to a $180,000 annual efficiency gain for the department.

Ultimately, they said, their goal was to create an AI tool that could be easily and rapidly implemented to improve patient compliance with appointments.

“The aim of our study was not to produce a highly complex model,” they explain, “but, rather, to produce one that could be developed relatively quickly, would require minimal data processing, and would be readily deployable in workflow practice for quality improvement.”