In the pipeline series: outpatient cancellations predictor

If you are interested in aquiring this tool, then please contact us at analytics-support@draperanddash.com.

One of the team will be more than happy to assist with any enquiries.

Alfonso Portabales – Data Scientist and Gary Hutson – Head of Solutions and AI

Cancellations affect the capacity of the outpatient system – leading to wasted slots, allied health professional time and general wastage in terms of capacity. We were inspired by a number of articles suggesting that robots and ML can cut down on this form of wastage: https://www.esneft.nhs.uk/robots-save-time-and-cut-down-on-wasted-appointments-in-outpatients/.

The team at D&D have been working to tackle this problem. This has been achieved by the application of a custom ML algorithm, designed to solve the problem of predicting cancellations.

Additionally, this custom algorithm is also used alongside our command centre tool, which is currently being deployed in a number of NHS trusts now. For information regarding this, please read our previous blog post: https://www.draperanddash.com/machinelearning/2019/10/command-centre-amplification-with-predictive-analytics-and-machine-learning/.

To make this prediction we use the behavioural and situational information, at a patient level, to reach a prediction. This prediction uses a probabilistic model to make a prediction of the likelihood of a patient cancelling an appointment unexpectedly.

We include generic features that are available to all trusts:

• Arrival hour bandings
• Patient Type
• Patient Age
• Source of referral
• Assessment wait
• Attendance category
• Where the patient lives in approximation to their local hospital

We are constantly improving our algorithms and we will work on building more features into the model, however with the minimal set the model performs well.

The algorithm we use is a popular ensemble method – random forests. This model builds individual decision trees, using different data samples, each of the tree then makes a prediction of whether they think the patient will cancel the appointment and then they have a mini general election to determine what classification they think the prediction should belong to.

In order to find an effective and reliable solution, we bucket the different variables into sensible groups, and train our system using a Random Forest algorithm with 3-fold cross validation and kappa boosting.

For more extensive information on decision trees – see: https://en.wikipedia.org/wiki/Decision_tree#Decision_tree_elements.