Solving the ‘Patient Did Not Attend’ mysteries
Did Not Attend (DNA) is a significant issue to the UK. The “free at the point of delivery” model that is key to the NHS means that someone not attending an outpatient appointment is an irrecoverable loss of the resources needed to run a clinic.
In RBH, the overall DNA rate is about 7% at a probable cost to the Trust of about £3m a year, not including the lost opportunity to see other patients.
The DNA project used the ready supply of data held in Bedrock to look for patterns in the properties of appointments that were not attended. With a large, high-quality data set there are many possibilities. Humans can use all manner of slicing and dicing to look over the data and come up with explanations for DNAs. Talking to NHS teams there are no end of explanations (urban myths?) that have arisen over the years. Age, sex, geography and time of appointment are known to affect DNA.
The stereotypes say that a 63-year-old woman will attend any appointment, but a 19-year-old man is less likely to attend anything but a Wednesday afternoon. And people also throw in explanations like public transport, demographics, working patterns and other issues. In our research at Trustmarque, we have also seen just plain wrong booking systems that don’t allow a patient to say they can’t attend so don’t discount the sample.
The AI alternative is to use Machine Learning to look at a large set of historical data and to look for patterns in the properties of DNAs and appointments. That was the approach at RBH.
The DNA issues at RBH used an AI approach that produce and probability score data from the Bedrock Platform that can were then split into two groups. First was “here are a few appointments that are very likely to be missed” and a second group of many more appointments that are less likely – but still quite likely to be missed.
The first was higher precision and the second was higher recall. Briefly explained higher precision only selected a subset of proposed appointments that were very likely to ‘DNA’. But it would exclude many that were quite likely to ‘DNA’. The Higher Recall result gave a larger volume of predicted DNAs, but was less discriminating.
Did it work? Yes – but the answer was also “It’s complicated”. In this case the AI could find a small number of more likely (>75%) to DNA (a few higher precision predictions) and larger number of quite probable predictions (higher recall). This was no magic solution but, a lot more information.