The Application of a Bayesian Machine Learning Platform to the Clinical Activity of Independent Prescribing Optometrists

Abstract

Background: The NHS Long Term Plan has laid the foundations for healthcare transformation. Artificial intelligence, digitally enabled care, and decision support are mentioned in the NHS Long Term Plan and the Topol Review as enablers for transformation. Independent Prescribing (IP) optometrists practising in isolation lack “live” peer support, seeking guidance from clinical guidelines and literature reviews instead. This sacrifices the currency of evidence for its quality and contributes to an evidence-to-practice gap. Aims of study: - to apply machine learning (ML, a subset of artificial intelligence) to the clinical activity of IP optometrists - to develop the concept of an interactive and evolving “live” evidence-based support system for IP optometrists and those in training Methods: Over a year, 1351 first patient consultations were collected by the Acute Primary Care Ophthalmology Service in West Kent (APCOS), a service delivered by IP optometrists. A digital learning platform was developed (MyDLP) to apply supervised machine learning (naïve Bayes’) to the data. A combined “intelligent” electronic patient record and virtual patient tool (iEPR/iVPT) within MyDLP provides decision support and automated grading. MyDLP also evaluates the performance of ML (accuracy, informedness and markedness) using cross validation and learning efficiency curves. The data in MyDLP can be manipulated to promote an understanding of ML concepts amongst clinicians. Results: A ‘proof-of-concept’ was demonstrated using the diagnoses and prescribing decisions for keratoconjunctivitis sicca (KCS) and uveitis. Maximum learning efficiency was reached, meaning more data would not have improved model performance. The study findings indicate that Bayes’ ML results in good replication for diagnoses and prescribing decisions. Conclusion: ML can be used to power “live”, “white box” decision support tools, useful to both qualified IP optometrists and those in training. As far as the author is aware this was the first time ML was applied to the clinical activity of IP optometrists.