Predicting scheduled hospital attendance with artificial intelligence - Amy Nelson et al., April 2019

Original article: https://www.nature.com/articles/s41746-019-0103-3#MOESM1

One-sentence summary

In this paper, a research group from UCL trained an AdaBoost artificial intelligence algorithm which used 81 variables to predict who would not attend their scheduled hospital appointments with an AUC of 0.852.

What did they do?

They used data from 22,318 MRI scan appointments over a three-year period at two north London hospitals to develop several different classification algorithms to predict non-attendance to appointments.

In total, 81 variables were included in the best performing model; these included appointment day and month, scan type, previously missed appointments, geographical location and waiting time until appointment.

Additionally, they developed a formula for calculating the cost-benefit of using such predictions to guide implementation aimed to boost adherence (such as a phone call or a text message).

What does this mean?

This has the potential to save hospital trusts money, by helping them to target their interventions for boosting attendance, such as targeted phone calls.

Given that the typical cost of an MRI is around £150, a phone call to remind patients to attend costs an estimated £6 and it has reported efficacy of improving attendance in around 33% of calls, their is a clear potential financial benefit demonstrated by such a predictive model.

The authors noted that roll-out of such an algorithm would have an upfront cost, but that this could be financially justified as it would pay for itself within 3-83 days (depending on how widely it was rolled out). On top of this, there would be an assumed medical benefit of more people attending their appointments.


Here is a visual summary of the paper - credit to Andy Callow.