The machine learning method worked even though some patient data was missing. The group used a data imputation strategy, based on a collection of missing type flags, to address frequently missing information in patient records. Including patient records with missing information significantly improved predictive accuracy compared to a gold standard approach that can only be trained and assessed on patients for whom complete information is available.
The exit model could be used by healthcare providers to identify patients who would benefit from a text, call, or email reminder about their appointment. No-show predictions could also be used to optimize scheduling systems and minimize patient wait times.
“Our results suggest that choosing the day of the week and the time of day that would be easier for patients and their parents to come to their medical appointments, and further using a language service and choosing a day with a likely nicer weather would help reduce no-shows,” wrote first author Dianbo Liu, of Boston Children’s Hospital and Harvard Medical School’s Department of Pediatrics, and colleagues. “These effects turned out to be the most related to parents’ working hours, city traffic and potentially patients’ school schedules.”