Detecting Clinical Intent in Electronic Healthcare Records in a UK National Healthcare Hospital

Following a patient consultation, the clinicians and supporting staff are responsible for determining and arranging the next clinical steps for the patient. A clinician may, for example, request a follow up appointment, order an imaging procedure or even discharge them. Each of these follow up-steps need to be placed in the hospital's system as an ‘order’. In some cases however, due to administrative error, these orders are not created which can subsequently place patients at risk of being ‘lost-to-follow up'. In most cases however the clinicians will record their ‘intention’ to create these orders in the patient's free-text notes. In this paper we investigate the use of machine learning to detect ‘clinical intent’ in the free text notes. We train and evaluate a machine learning model using real hospital data from a UK based National Healthcare Services Hospital and report macro-precision of 0.91, macro-recall of 0.90 and a macro-f1 of 0.90. Additionally we discuss the pilot deployment of the model at the hospital as part of and alerting workflow to warn clinical staff of patients at risk of being lost-to-follow up. In particular we discuss a prototype dashboard interface, currently being trialled within the hospital, that surfaces patient's who are at risk of being lost-to-follow up to clinical support staff.