Translational Learning with Orange Data Mining

Abstract

Abstract for the e-NATCONPH 2021 (International Conference) TRANSLATIONAL LEARNING WITH ORANGE DATA MINING Raqib F 1, Dunne MCM 1*, Gurney JC2, Harle DE 3, Sivapalan T 2, Sabokbar N 2, Bhogal-Bhamra GK 1 1 Ophthalmic Research Group, Optometry School, Aston University, Birmingham, UK 2 Acute Primary Care Ophthalmology Service, West Kent CCG, Aylesford, UK 3 Acute Primary Care Ophthalmology Service, West Kent CCG, Tonbridge, UK *Corresponding author’s email ID: m.c.m.dunne@aston.ac.uk BACKGROUND. Health Education England’s Topol Review has recommended preparation of clinicians for a digital future. Orange Data Mining software enables hands-on exposure of machine learning to practitioners that traditionally lack this training. PURPOSE. This case study presents a translational learning approach, used for teaching undergraduate optometrists, that includes (a) gathering clinical evidence (b) learning from the clinical evidence and (c) translation to evidence-based teaching and practice. METHODOLOGY In this approach, students are taught about research ethics before creating an Orange Data Mining canvas containing widgets to upload clinical data (File), remove missing data (Impute), assign variables (Select columns), carry out machine learning (Naïve Bayes and Logistic Regression), master cross validation and hyperparameter tuning (Test and score) before gaining new knowledge and clinical decision support (Nomogram). This is demonstrated with 1351 real clinical cases for determining the relative importance of clinical data, recommended by the College of Optometrists’ Clinical Management Guidelines, for investigating an anterior eye disease (uveitis). RESULTS. Students discover that Naive Bayes has higher informedness (96%) than tuned Logistic Regression (90%). The Naïve Bayes nomogram reveals the relative importance of the clinical symptoms and signs while the Logistic regression nomogram indicates possible redundancy. A presentation of acute unilateral discomfort and visual disturbance with mild red eye and anterior chamber inflammation results in 90% and 68% probabilities of uveitis according, respectively, to Naïve Bayes’ and Logistic Regression nomograms. CONCLUSION. Our students enjoy this translational learning approach and we ask if it might also be useful for training other health scientists. Key Words: Education, Health Sciences, Translational learning