AI Model Can Help Predict, Prevent Pediatric Appointment No-Shows

A deep-learning model can accurately predict pediatric no-shows using data from patient EHRs and local weather information, which can then be used to implement no-show prevention measures, according to a new study published in npj Digital Medicine.

Appointment no-shows can negatively affect health outcomes and health system resource utilization because check-ups, preventative care, and treatment cannot be provided if the appointment is missed. According to the study, patients with a prior history of no-shows, public health insurance, and lower socioeconomic status are more likely to miss their appointments. Patients who miss appointments often cite traffic, scheduling issues, time conflicts, and environmental factors as the main reasons.

The researchers developed their deep learning model by retrospectively collecting EHR data for 19,450 patients between Jan. 10, 2015, and Sept. 9, 2016, at Boston Children's Hospital’s primary care pediatric clinic. These records included data from 161,822 medical appointments and information related to patient age, gender, previous no-show rate, and health insurance type. Of these appointments, 20.3 percent were no-shows, which means the patient neither showed up nor canceled.

The data was used to train the algorithm to predict the risk of a no-show when the appointment was scheduled. The model was trained using records with and without missing information, as 77 percent of the records gathered contained at least one missing data feature. The seven features that made up the missing data were “status of previous visits,” “mother’s education level,” “public or private insurance,” “insurance plan,” “payor,” “language,” “race,” and “clinical primary care provider’s name.”

The researchers hypothesized that local weather information was crucial for predicting no-shows. Thus, they included ambient temperature, wind speed, humidity, and atmospheric pressure on the day of the appointment in the city where the primary care clinic is located as predictive features.

After training their model, researchers compared its performance to those of the conventional persistence baseline approach, which predicts the no-show behavior of a patient based on whether or not they showed up to their last appointment, and a simple logistic regression approach as a reference.

Overall, the AI and logistic regression models outperformed the persistence baseline approach. Both also found that the history of patients’ no-show records and atmospheric pressure were the most important predictors of no-show behavior.

This research indicates that if accommodated properly, the inclusion of data about local weather patterns and patient records, even when they contain missing information, can significantly improve the accuracy of a prediction model.  

The researchers also suggested ways their model can potentially inform no-show reduction efforts. They suggest that their AI method can be used to identify patients who could benefit from a call, text, or email reminder for their appointments. Additionally, it could help calculate how frequently these reminders should be sent.

For clinics with fewer resources, the researchers posited that their model could help schedulers assess which days have favorable weather or pinpoint the days on which resources such as free transportation or language services are available, decreasing the likelihood of a no-show.